Daily Papers

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

Key-value (KV) cache growth is a major bottleneck in autoregressive decoding, as memory and bandwidth scale linearly with context length. Existing KV eviction methods often rely on static heuristics or proxy scores, which poorly track future token utility and cause brittle eviction as relevance shifts. To address this, we introduce KVpop, which learns a fixed-budget KV eviction policy by directly supervising the keep-or-drop decision. The scorer is trained against a novel future-attention target, computed efficiently without materializing dense attention maps. We further introduce a delayed memory-based scorer that, uniquely among learned eviction methods, defers scoring for a fixed number of steps to exploit near-future context. On AIME and HMMT mathematical reasoning, KVpop retains 98% of full-attention performance on Qwen3-4B at 75% KV cache compression and 97% at 88% compression, consistently outperforming established eviction baselines. Qwen3-8B shows even stronger results, reaching near-full teacher performance. These results show that supervising eviction with future-attention signals cuts memory costs while maintaining quality.

The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token's future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of lightweight per-head RL agents trained on pre-computed generation traces using only key and value vectors. Each agent learns a specialized eviction policy guided by a holistic reward, derived from future utility, that evaluates the quality of the ranking across all cache budgets, requiring no modifications to the underlying LLM or additional inference. Evaluated across two model families on the long-context benchmark RULER (up to 128K tokens) and the multi-turn dialogue benchmark OASST2-4k, KVP significantly outperforms strong baselines. Zero-shot tests on standard downstream tasks (BoolQ, LongBench passage retrieval, GovReport) further show that KVP generalizes beyond its training distribution and to considerably longer sequence lengths. These results demonstrate that learning to predict future token utility is a powerful and scalable paradigm for adaptive KV cache management.

Recent advances in large language models (LLMs) have significantly boosted long-context processing. However, the increasing key-value (KV) cache size poses critical challenges to memory and execution efficiency. Most KV cache compression methods rely on heuristic token eviction using all attention heads in Grouped Query Attention (GQA)-based LLMs. This method ignores the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrades the performance of LLMs. To address the issue above, instead of using all the attention heads in GQA-based LLMs to determine important tokens as in the previous work, we first identify the attention heads in each layer that are not only capable of retrieving the initial and final tokens of a prompt, but also capable of retrieving important tokens within the text and attending to their surrounding semantic context. Afterwards, we exploit such heads to determine the important tokens and retain their corresponding KV cache pairs. Furthermore, we analyze the cache eviction error of each layer individually and introduce a layer-adaptive KV cache allocation strategy. Experimental results demonstrate the proposed CompressKV consistently outperforms state-of-the-art approaches under various memory budgets on LongBench and Needle-in-a-Haystack benchmarks. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV.git.

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-p selection more suitable than fixed top-k sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36times prefill speedup at 1M context and about a 2.01times decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value vector information on top of attention-based eviction scores. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process.

Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBench and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes will be released on Github.

Large Language Models (LLMs) are increasingly expected to operate over long contexts, yet standard softmax attention incurs a KV cache that grows linearly with sequence length, quickly becoming the bottleneck for long context inference. A practical remedy is to evict less important KV entries; however, existing eviction policies are largely heuristic and struggle to capture the rich, input-dependent distribution of token importance. In this work, we introduce a learnable indexer that predicts KV importance, enabling more accurate retention of critical tokens. Meanwhile, naively evicting tokens permanently discards their information, leading to irreversible forgetting and degraded retrieval over long ranges. To address this, we propose a lightweight latent memory module that compresses evicted tokens into a compact, online-updated state and provides residual readouts to compensate for the attention contributions lost through KV eviction. Collectively, our method enables accurate long-context inference under a bounded KV budget, delivering consistent improvements on RULER (4K/16K) across Qwen, Mistral, and Llama models (up to 25 points under aggressive eviction), markedly more stable Needle-in-a-Haystack retrieval, and superior LongBench scores and compression curves compared to existing eviction policies.

Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware. Existing KV cache eviction methods typically apply heuristic token scoring over all heads in GQA-based LLMs. These methods ignore the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrading the performance of LLMs. To address this issue, we propose CompressKV, a resource-efficient KV-cache compression framework for GQA-based LLMs. Instead of aggregating attention scores from all heads, CompressKV identifies Semantic Retrieval Heads (SRHs) that capture both the initial and final tokens of a prompt and semantically important mid-context evidence, and uses them to select tokens whose KV pairs should be retained. Furthermore, CompressKV allocates cache budgets across layers according to offline estimates of layer-wise eviction error. Experiments on LongBench and Needle-in-a-Haystack show that CompressKV consistently outperforms existing KV-cache eviction methods across memory budgets. Notably, it preserves over 97\% of full-cache performance using only 3\% of the KV cache on LongBench question-answering tasks and achieves 90\% accuracy with just 0.7\% KV storage on Needle-in-a-Haystack. These results demonstrate an improved resource--performance trade-off for long-context LLM inference. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV

Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.

Large language models (LLMs) support long-context inference but suffer from substantial memory and runtime overhead due to Key-Value (KV) Cache growth. Existing KV Cache eviction methods primarily rely on local attention weights, neglecting the influence of value representations, output projection, and inter-head interactions. In this work, we reformulate KV Cache eviction from a conventional head-wise, weight-averaging approach into an output-aware, layer-wise matrix multiplication approximation problem. We introduce LaProx, a novel eviction strategy that explicitly models the multiplicative interaction between attention maps and projected value states to accurately quantify token contributions while accounting for inter-head dependencies. Building on this metric, we propose the first unified eviction strategy that assigns globally comparable importance scores to tokens, enabling model-wide selection instead of local, head-wise decisions. Experimental results across 19 datasets on long-context benchmarks LongBench and Needle-In-A-Haystack demonstrate that our approach maintains model performance with only 5\% of the KV cache and consistently outperforms prior works across all configurations. Notably, our method achieves up to 2times accuracy loss reduction under extreme compression scenarios compared to existing state-of-the-art baselines with minimal overhead.

Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL.

Large Language Models (LLMs) exhibit enhanced reasoning capabilities by employing Chain-of-Thought (CoT). However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache size, particularly in tasks requiring long reasoning sequences, such as mathematics and programming. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a Token Importance Recurrence phenomenon: a large proportion of tokens receive renewed attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose LazyEviction, a lagged KV eviction framework designed to maintain reasoning performance while reducing KV memory. LazyEviction is an Observation Window-based Lagged Eviction Mechanism retaining latent recurring tokens by performing lagged evictions across decoding steps, which contains two key components: (1) Recurrence Interval Tracking for capturing temporal variations in token importance, and (2) an Maximum Recurrence Interval-Centric Eviction Policy that prioritizes eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache size by 50% while maintaining comparable accuracy on mathematics reasoning datasets, outperforming state-of-the-art methods. Our findings highlight the importance of preserving recurring tokens, which are critical for maintaining knowledge continuity in multi-step reasoning tasks.

As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present TokenPilot, a dual-granularity context management framework. Globally, Ingestion-Aware Compaction acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, Lifecycle-Aware Eviction monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on PinchBench and Claw-Eval under both isolated and continuous modes demonstrate that TokenPilot reduces costs by 61% and 56% in isolated mode, and 61% and 87% in continuous mode, while maintaining competitive performance compared to prior systems. TokenPilot has been integrated into LightMem2 at https://github.com/zjunlp/LightMem2.

Despite the recent success associated with Large Language Models (LLMs), they are notably cost-prohibitive to deploy in resource-constrained environments due to their excessive memory and computational demands. In addition to model parameters, the key-value cache is also stored in GPU memory, growing linearly with batch size and sequence length. As a remedy, recent works have proposed various eviction policies for maintaining the overhead of key-value cache under a given budget. This paper embarks on the efficacy of existing eviction policies in terms of importance score calculation and eviction scope construction. We identify the deficiency of prior policies in these two aspects and introduce RoCo, a robust cache omission policy based on temporal attention scores and robustness measures. Extensive experimentation spanning prefilling and auto-regressive decoding stages validates the superiority of RoCo. Finally, we release EasyKV, a versatile software package dedicated to user-friendly key-value constrained generative inference. Code available at https://github.com/DRSY/EasyKV.

Large language model (LLM) agents accumulate heterogeneous context, including system instructions, plans, user turns, retrieved documents, tool outputs, and intermediate reasoning, whose key-value (KV) cache can become a major memory bottleneck. Existing eviction policies generally apply the same attention- or recency-based rule to every token, ignoring semantic structure already available to the agent orchestrator. We introduce MemDecay, a training-free, region-aware KV-cache eviction policy. MemDecay assigns tokens region-specific base priorities and decay rates, refreshes retention scores when tokens receive attention, and evicts the lowest-scoring pages under a fixed cache budget while allowing critical regions to be pinned. We also provide a procedure for calibrating decay rates from measured attention lifetimes. We evaluate MemDecay at approximately 450 and 1,700 token contexts using Qwen2.5-1.5B and 3B. Across all settings, attention lifetimes differ by an order of magnitude across regions: system-token half-lives range from 148 to 189 decoding steps, compared with 14 to 16 for scratchpad tokens. Pinning preserves system-region facts at full-cache accuracy in every setting, while no baseline preserves more than 13 of 24. Region-aware retention remains effective as context grows, whereas recency-based retention collapses. Accumulated-attention retention performs better on unpinned content, however, and ablations identify attention-score normalization as the main limitation of the current formulation. These results establish semantic prompt structure as a robust signal for KV-cache management while clarifying how it should be combined with attention-based importance.

Linear-attention models that compress the entire input sequence into a fixed-size recurrent state offer an efficient alternative to Transformers, but their finite memory induces forgetfulness that harms retrieval-intensive tasks. To mitigate the issue, we explore a series of hybrid models that restore direct access to past tokens. We interleave token mixers with intermediate time and space complexity between linear and full attention, including sparse attention with token eviction, and the query-aware native sparse attention. Particularly, we propose a novel learnable token eviction approach. Combined with sliding-window attention, an end-to-end trainable lightweight CNN aggregates information from both past and future adjacent tokens to adaptively retain a limited set of critical KV-pairs per head, maintaining linear attention's constant time and space complexity. Efficient Triton kernels for the sparse attention mechanisms are provided. Empirical evaluations on retrieval-intensive benchmarks support the effectiveness of our approaches.

Key-value (KV) cache management through compression and eviction strategies has emerged as an important research direction in recent years. Computational demands of large language models (LLMs) and their multi-modal variants during output generation can be partially alleviated by caching previous key and value calculations needed by subsequent scaled dot-product attention operations. However, this leads to another problem: the size of the resulting KV cache grows linearly with context length and quickly consumes all available GPU memory when either the prompt or the generated output are long. KV cache management periodically prunes entries from the cache thereby reducing its memory footprint while attempting to retain sufficient information for accurate generation. A by-product is faster inference speed. We propose a simple yet effective KV eviction scheme motivated by the insight that past tokens which can be well-predicted from more recent tokens are redundant and their associated keys and values can be removed from the cache. To score entries for eviction we run the model on the tokens in their original order, reusing the key and value representations already stored in the KV cache, and applying a counter-causal attention mask so that each position attends only to its future context. This is in-distribution, tied directly to the actual cache contents, and requires no additional training. To further reduce cost, we additionally propose a fast single-layer approximation that restricts the counter-causal pass to the last transformer layer, achieving a significant speedup per refresh cycle at marginal accuracy cost. We evaluate our strategy on various open-source LLMs and benchmark datasets showing competitive or improved performance over other state-of-the-art methods. Reference code is available at https://github.com/metacognitionai/counter_causal.

Long-horizon research agents solve open-ended tasks through iterative retrieval, aggregation, and synthesis, but context grows rapidly while the marginal value of additional evidence often declines. This leads to unnecessary token cost, higher latency, and noisier inputs for final report generation. We study marginal value estimation for context management in deep research agents and present the first systematic stage-aware comparison of pruning strategies across the pipeline. We evaluate lightweight heuristic criteria and a learned value model at pre-retrieval, post-retrieval, and pre-synthesis stages. Our results show that pruning effectiveness depends more on where pruning is applied than on the specific scoring rule: early pruning yields the largest end-to-end savings, while later pruning mainly refines the final synthesis context. Lightweight heuristics reduce token usage by up to 73% with little quality degradation, learned pruning remains competitive on selected trade-offs, and no single method dominates across quality, efficiency, and faithfulness. These findings provide practical guidance for designing efficient long-horizon agentic systems.

The key-value (KV) cache is a major bottleneck in long-context inference, where memory and computation grow with sequence length. Existing KV eviction methods reduce this cost but typically degrade performance relative to full-cache inference. Our key insight is that full-cache attention is not always optimal: in long contexts, irrelevant tokens can dilute attention away from useful evidence, so selective, learnable eviction can improve generation rather than merely approximate the full cache. We introduce a global retention-based KV eviction method that learns each token's future utility under a unified memory budget. Lightweight retention gates assign utility scores to cached KV entries, and a shared final scoring projection calibrates these scores across all layers and heads. This enables a single global eviction policy in which tokens from different layers, heads, and modalities compete directly for cache capacity. We further provide theoretical analysis showing that preferentially retaining useful tokens reduces attention dilution, and we justify geometric retention as a query-agnostic proxy for future utility. Across diverse long-context language and vision-language reasoning, and multi-turn dialogue benchmarks, our method substantially reduces KV memory while matching or surpassing full-cache inference. These results suggest that learned, globally calibrated KV eviction is not only a compression technique, but also a mechanism for improving long-context reasoning.

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a "cake-slicing problem." CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.

The KV-Cache technique has become the standard for the inference of large language models (LLMs). It caches states of self-attention to avoid recomputation. Yet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system, especially when confronted with ultra-large models and long-context queries. A natural remedy is to discard the KV-Cache for less important tokens, with StreamingLLM as an example, but the used static eviction strategies cannot flexibly adapt to varying contexts. Remedies like H2O leverage accumulative attention scores to perform dynamic eviction but suffer from the attention bias issue in capturing contextual information. This paper bridges this gap by devising a parameterized KV-Cache eviction mechanism, dubbed as Attention-Gate, which accepts the whole context as input and yields eviction flags for each token to realize in-context eviction. The subsequent self-attention module proceeds according to the flags and only the KV states for the remaining tokens need to be cached. The Attention-Gates can vary among different heads and layers and be trivially plugged into pre-trained LLMs, tuned by cost-effective continual pre-training or supervised fine-tuning objectives to acquire what to discard. The computational and memory overhead introduced by Attention-Gates is minimal. Our method is validated across multiple tasks, demonstrating both efficiency and adaptability. After a highly efficient continual pre-training, it achieves higher average accuracy and evicts more tokens compared to traditional training-free methods. In supervised fine-tuning, it not only evicts many tokens but also outperforms LoRA-finetuned LLMs on some datasets, such as RTE, where it improves accuracy by 13.9% while evicting 62.8% of tokens, showing that effective eviction of redundant tokens can even enhance performance.

Most KV-cache eviction methods ask, in effect, which memory appeared important while reading the prompt? We instead ask, which memory will matter while answering? Since decoding queries are unavailable at eviction time, prior future-aware methods rely on pseudo-responses or synthetic future-query estimates. We cast fixed-budget future-aware eviction as distributional estimation over plausible model-conditional query trajectories and introduce LORE-KV (Lookahead Output-perturbation with Reliability-weighted Ensembles for Key-Value caches), a training-free method that samples short autoregressive continuations from the frozen target model and uses their response-side query states to estimate prompt-token utility. Tokens are scored by projected leave-one-out attention-output deletion cost and aggregated across sampled futures with optional trajectory weighting. The temporary continuations are discarded before final decoding, requiring no auxiliary model or training. Ablations isolate the mechanism: at B=128, a single response-side continuation recovers about 89% of the gain over the prompt-window control, while additional futures provide smaller improvements. At B=128, LORE-KV raises the LongBench average on Qwen2.5-14B from 45.49 to 48.24 (+2.75) and the 16K RULER average on Mistral-7B from 45.20 to 51.05 (+5.85). Gains diminish at larger cache budgets and coexist with task-level regressions. LORE-KV incurs 1.46-2.77x AnDPro's per-sample wall-clock time as a one-time compression overhead across six dense and hybrid-attention backbones.

Transformer-based large language models (LLMs) rely on key-value (KV) caching to avoid redundant computation during autoregressive inference. While this mechanism greatly improves efficiency, the cache size grows linearly with the input sequence length, quickly becoming a bottleneck for long-context tasks. Existing solutions mitigate this problem by evicting prompt KV that are deemed unimportant, guided by estimated importance scores. Notably, a recent line of work proposes to improve eviction quality by "glimpsing into the future", in which a draft generator produces a surrogate future response approximating the target model's true response, and this surrogate is subsequently used to estimate the importance of cached KV more accurately. However, these approaches rely on computationally expensive draft generation, which introduces substantial prefilling overhead and limits their practicality in real-world deployment. To address this challenge, we propose LookaheadKV, a lightweight eviction framework that leverages the strength of surrogate future response without requiring explicit draft generation. LookaheadKV augments transformer layers with parameter-efficient modules trained to predict true importance scores with high accuracy. Our design ensures negligible runtime overhead comparable to existing inexpensive heuristics, while achieving accuracy superior to more costly approximation methods. Extensive experiments on long-context understanding benchmarks, across a wide range of models, demonstrate that our method not only outperforms recent competitive baselines in various long-context understanding tasks, but also reduces the eviction cost by up to 14.5x, leading to significantly faster time-to-first-token. Our code is available at https://github.com/SamsungLabs/LookaheadKV.

Long-output reasoning shifts the KV-cache bottleneck from the fixed prompt to the generated trace. Existing reasoning-cache eviction methods largely treat cached entries as routing objects, estimating whether an old key will still be read, will recur, or can be replaced. This routing-only view overlooks two effects: low-attention entries can carry large value payloads whose removal changes future predictions, and newly generated states can appear stale before later queries have had a chance to read them. We introduce AvoKV-E, a training-free eviction policy that first delays eligibility for recent states and then ranks eligible entries using candidate-normalized read pressure, key redundancy, and value-payload potential. According to empirical evaluation across different models and datasets, AvoKV-E matches or exceeds redundancy-aware, recurrence-based, and thought-adaptive eviction baselines at matched active-KV budgets, with its largest gains in the tightest-cache regime. Component and counterfactual analyses further connect these gains to delayed observation, payload-aware scoring, redundancy, and scale-robust normalization. Together, the results show that long-reasoning KV eviction should preserve not only keys that are likely to be read, but also the value payloads that sustain the reasoning trajectory.

The KV cache used in large language models has linearly growing time complexity, so LLMs face memory blow-up and reduced decoding efficiency when they process long contexts. Current KV Cache eviction has become an important research direction; however, existing methods based on fixed Soft Tokens (e.g., Judge Q) rely on a static parameter set as the query to evaluate the importance of KV pairs, so they cannot adapt dynamically to different input prompts, and they cannot precisely capture complex and changing task relevance. Also, evicted KV pairs are discarded permanently, so this causes irreversible information loss and context breaks. To address this problem, we propose Meta-Soft, a dynamic compression framework based on probe-driven context integration. Specifically, we build a meta-library with a learnable orthogonal basis matrix L, and we use a selector network with Gumbel-Softmax to produce differentiable sparse combination weights, so we dynamically synthesize the most targeted k Soft Tokens from the input prompt features. We append these Soft Tokens to the end of the input sequence to probe key information. We also introduce an attention-flow based integration mechanism, which redistributes the semantic information of removed tokens into retained tokens, and this keeps the dropped context information effectively. Experiments on multiple datasets show that our method outperforms existing state-of-the-art eviction methods and provides a new solution for KV Cache compression.

Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.

Heuristic functions are essential to the performance of tree search algorithms such as A*, where their accuracy and efficiency directly impact search outcomes. Traditionally, such heuristics are handcrafted, requiring significant expertise. Recent advances in large language models (LLMs) and evolutionary frameworks have opened the door to automating heuristic design. In this paper, we extend the Evolution of Heuristics (EoH) framework to investigate the automated generation of guiding heuristics for A* search. We introduce a novel domain-agnostic prompt augmentation strategy that includes the A* code into the prompt to leverage in-context learning, named Algorithmic - Contextual EoH (A-CEoH). To evaluate the effectiveness of A-CeoH, we study two problem domains: the Unit-Load Pre-Marshalling Problem (UPMP), a niche problem from warehouse logistics, and the classical sliding puzzle problem (SPP). Our computational experiments show that A-CEoH can significantly improve the quality of the generated heuristics and even outperform expert-designed heuristics.

Heuristics are widely used for dealing with complex search and optimization problems. However, manual design of heuristics can be often very labour extensive and requires rich working experience and knowledge. This paper proposes Evolution of Heuristic (EoH), a novel evolutionary paradigm that leverages both Large Language Models (LLMs) and Evolutionary Computation (EC) methods for Automatic Heuristic Design (AHD). EoH represents the ideas of heuristics in natural language, termed thoughts. They are then translated into executable codes by LLMs. The evolution of both thoughts and codes in an evolutionary search framework makes it very effective and efficient for generating high-performance heuristics. Experiments on three widely studied combinatorial optimization benchmark problems demonstrate that EoH outperforms commonly used handcrafted heuristics and other recent AHD methods including FunSearch. Particularly, the heuristic produced by EoH with a low computational budget (in terms of the number of queries to LLMs) significantly outperforms widely-used human hand-crafted baseline algorithms for the online bin packing problem.

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at https://github.com/SalesforceAIResearch/Random-Attention.

Chain-of-thought reasoning has driven striking advances in language model capability, yet every reasoning step grows the KV cache, creating a bottleneck to scaling this paradigm further. Current approaches manage these constraints on the model's behalf using hand-designed criteria. A more scalable approach would let end-to-end learning subsume this design choice entirely, following a broader pattern in deep learning. After all, if a model can learn to reason, why can't it learn to forget? We introduce Neural Garbage Collection (NGC), in which a language model learns to forget while learning to reason, trained end-to-end from outcome-based task reward alone. As the model reasons, it periodically pauses, decides which KV cache entries to evict, and continues to reason conditioned on the remaining cache. By treating tokens in a chain-of-thought and cache-eviction decisions as discrete actions sampled from the language model, we can use reinforcement learning to jointly optimize how the model reasons and how it manages its own memory: what the model evicts shapes what it remembers, what it remembers shapes its reasoning, and the correctness of that reasoning determines its reward. Crucially, the model learns this behavior entirely from a single learning signal - the outcome-based task reward - without supervised fine-tuning or proxy objectives. On Countdown, AMC, and AIME tasks, NGC maintains strong accuracy relative to the full-cache upper bound at 2-3x peak KV cache size compression and substantially outperforms eviction baselines. Our results are a first step towards a broader vision where end-to-end optimization drives both capability and efficiency in language models.

We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection scheme based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.

Large language models have revolutionized natural language processing, yet their deployment remains hampered by the substantial memory and runtime overhead of the transformer's Key-Value cache. To mitigate this, recent methods employ a scoring-aggregation framework to evict unimportant cache entries, based on the stability assumption-that a fixed subset of entries remains consistently important during generation. However, prior work has largely focused on refining importance indicators for scoring, while defaulting to mean aggregation due to a faithful trust in the stability assumption. In this work, we argue that this underlying assumption is inherently fragile, making mean aggregation highly vulnerable in extreme cases. To counter this, we propose a simple yet elegant defensive aggregation strategy: a two-step, linear-time approach that controls worst-case risk, thereby defending against extreme cases with negligible computational overhead. Embodying this strategy, we propose a novel cache eviction method, DefensiveKV and its extension, Layer-DefensiveKV, which incorporates layer-wise budget allocation. Across seven task domains (18 datasets), our methods reduce generation quality loss by 2.3x and 4.3x respectively, versus the strongest baseline under a 20% cache size. These results set new performance benchmarks and pioneer a promising direction for optimizing cache eviction against underlying fragility through worst-case risk management. Our code is available at https://github.com/FFY0/DefensiveKV.

Streaming visual transformers like StreamVGGT achieve strong 3D perception but suffer from unbounded growth of key value (KV) memory, which limits scalability. We propose a training-free, inference-time token eviction policy that bounds memory by discarding redundant tokens while keeping the most informative ones. Our method uses significantly less memory with little to no drop in accuracy: on 7-Scenes with long sequences it reduces peak memory from 18.63 GB to 9.39 GB while accuracy and completeness drop by only 0.003. Under strict memory budgets, eviction enables denser frame sampling, which improves reconstruction accuracy compared to the baseline. Experiments across video depth estimation (Sintel, KITTI), 3D reconstruction (7-Scenes, NRGBD), and camera pose estimation (Sintel, TUM-dynamics) show that our approach closely matches StreamVGGT at a fraction of the memory and makes long-horizon streaming inference more practical.

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39\% compared to the Full KV Cache baseline.

Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.

Despite its simplicity and efficacy, the high token expenditure of self-consistency can limit its practical utility. Here we investigate if self-consistency can be made more token-efficient for long chain-of-thought reasoning tasks, while preserving its parallelism, through early hypothesis pruning. Concretely, we generate all solutions in parallel, but periodically prune intermediate hypotheses that are deemed unnecessary based on two lightweight indicators: (a) the model's own confidence in individual hypotheses, and (b) lexical coverage of all current hypotheses by candidate subsets that are under consideration for continued retention. We design a fast weighted set cover algorithm that utilizes the two indicators; our evaluation of five LLMs on three math benchmarks shows that this method can improve token efficiency for all models, by 10-35% in many cases.

Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

Reinforcement learning (RL) is a pivotal task for enhancing Large Language Model (LLM) reasoning. Conventional algorithms, however, typically adhere to a coarse-grained credit assignment paradigm, applying a uniform reward to all tokens in a sequence, a critical flaw in long-chain reasoning tasks. In this paper, we address this challenge and propose Dynamic Entropy Weighting, a novel mechanism that facilitates fine-grained rewards through two new algorithms: Group Token Policy Optimization (GTPO), which assigns an entropy-weighted reward to each token, and the analogous algorithm Sequence-Level GRPO (GRPO-S). Our approach is founded on the hypothesis that high policy entropy within a reasoning path is a powerful heuristic for cognitive effort at pivotal junctures, which can be repurposed into a learning signal. By repurposing policy entropy for reward shaping, we achieve true per-token credit assignment. Experimental results across challenging reasoning benchmarks validate the superiority of our approach, showing our methods significantly outperform a strong DAPO baseline and confirming our entropy-weighting mechanism as the key driver of this performance boost.

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56times speedup using just 18.9\% of the cache on AIME24.

Large Language Models have excelled in various domains but face efficiency challenges due to the growing Key-Value (KV) cache required for long-sequence inference. Recent efforts aim to reduce KV cache size by evicting vast non-critical cache elements during runtime while preserving generation quality. However, these methods typically allocate compression budgets uniformly across all attention heads, ignoring the unique attention patterns of each head. In this paper, we establish a theoretical loss upper bound between pre- and post-eviction attention output, explaining the optimization target of prior cache eviction methods, while guiding the optimization of adaptive budget allocation. Base on this, we propose {\it Ada-KV}, the first head-wise adaptive budget allocation strategy. It offers plug-and-play benefits, enabling seamless integration with prior cache eviction methods. Extensive evaluations on 13 datasets from Ruler and 16 datasets from LongBench, all conducted under both question-aware and question-agnostic scenarios, demonstrate substantial quality improvements over existing methods. Our code is available at https://github.com/FFY0/AdaKV.

Recent progress in LLM reasoning has increasingly shifted from single-pass generation to explicit search over intermediate reasoning states. Tree-of-Thoughts (ToT) organizes inference to tree-structured search with branching and backtracking, but it substantially amplifies the Key--Value (KV) cache: retaining KV states for a frontier of partial trajectories quickly becomes a memory bottleneck that limits throughput and constrains search depth and width under fixed hardware budgets. We address this challenge by observing that KV reuse in ToT-style inference is governed by search dynamics: near-term decoding depends primarily on the active branch and its ancestors, whereas inactive subtrees have low short-term reuse probability yet must remain recoverable for backtracking. Motivated by this, we propose ArborKV, a structure-aware eviction framework that couples a lightweight value estimator with a tree-aware allocation policy, and performs purely token-extractive eviction with lazy rehydration to support revisits. Experiments on ToT-style reasoning benchmarks show that ArborKV achieves up to ~4x peak KV-memory reduction while preserving near-full-retention accuracy, enabling larger search configurations under fixed device budgets that would otherwise run out of memory.

Efficient long-context inference is critical as large language models (LLMs) adopt context windows of ranging from 128K to 1M tokens. However, the growing key-value (KV) cache and the high computational complexity of attention create significant bottlenecks in memory usage and latency. In this paper, we find that attention in diverse long-context tasks exhibits sparsity, and LLMs implicitly "know" which tokens can be dropped or evicted at the head level after the pre-filling stage. Based on this insight, we propose Self-Attention Guided Eviction~(SAGE-KV), a simple and effective KV eviction cache method for long-context inference. After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache, enabling efficient inference with the reduced cache. Evaluations on LongBench and three long-context LLMs (Llama3.1-8B-Instruct-128k, Llama3-8B-Prolong-512k-Instruct, and Qwen2.5-7B-Instruct-128k) show that SAGE-KV maintains accuracy comparable to full attention while significantly improving efficiency. Specifically, SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.

While large reasoning models trained with critic-free reinforcement learning and verifiable rewards (RLVR) represent the state-of-the-art, their practical utility is hampered by ``overthinking'', a critical issue where models generate excessively long reasoning paths without any performance benefit. Existing solutions that penalize length often fail, inducing performance degradation due to a fundamental misalignment between trajectory-level rewards and token-level optimization. In this work, we introduce a novel framework, DECS, built on our theoretical discovery of two previously unaddressed flaws in current length rewards: (1) the erroneous penalization of essential exploratory tokens and (2) the inadvertent rewarding of partial redundancy. Our framework's innovations include (i) a first-of-its-kind decoupled token-level reward mechanism that surgically distinguishes and penalizes redundant tokens, and (ii) a novel curriculum batch scheduling strategy to master the efficiency-efficacy equilibrium. Experimental results show DECS can achieve a dramatic reduction in reasoning tokens by over 50\% across seven benchmarks while simultaneously maintaining or even improving performance. It demonstrates conclusively that substantial gains in reasoning efficiency can be achieved without compromising a model's underlying reasoning power. Code is available at https://github.com/pixas/DECS.

Large language models increasingly operate over long contexts, where the KV cache becomes a dominant memory bottleneck: its size grows linearly with sequence length and must be retained throughout decoding, making full GPU caching prohibitively expensive without compression. Existing KV cache compression methods struggle to balance efficiency with faithful context preservation. Token eviction discards information, while semantic grouping fixes compression decisions at prefill time; neither can recover token-level detail from a compressed span once it becomes relevant during generation. As a solution, we propose SeKV, a resolution-adaptive semantic KV cache that organizes context into entropy-guided semantic spans and stores them across a GPU-CPU memory hierarchy without discarding information. Each span keeps a lightweight summary vector on GPU for coarse routing and a low-rank SVD basis on CPU for on-demand token-level reconstruction. A trained zoom-in mechanism selectively expands query-relevant spans during decoding, enabling precise retrieval without materializing the full KV cache on GPU. SeKV enables adaptive token-level reconstruction while keeping the base LLM fully frozen and adding fewer than 0.05% trainable parameters. Across four benchmarks, SeKV improves over the strongest semantic compression baseline by 5.9% on average while reducing GPU memory by 53.3% versus full KV caching at 128K context. Code is available on https://github.com/AmirAbaskohi/SeKV.

The omnipresence of NP-hard combinatorial optimization problems (COPs) compels domain experts to engage in trial-and-error heuristic design. The long-standing endeavor of design automation has gained new momentum with the rise of large language models (LLMs). This paper introduces Language Hyper-Heuristics (LHHs), an emerging variant of Hyper-Heuristics that leverages LLMs for heuristic generation, featuring minimal manual intervention and open-ended heuristic spaces. To empower LHHs, we present Reflective Evolution (ReEvo), a novel integration of evolutionary search for efficiently exploring the heuristic space, and LLM reflections to provide verbal gradients within the space. Across five heterogeneous algorithmic types, six different COPs, and both white-box and black-box views of COPs, ReEvo yields state-of-the-art and competitive meta-heuristics, evolutionary algorithms, heuristics, and neural solvers, while being more sample-efficient than prior LHHs.

Token-level code completion is one of the most critical features in modern Integrated Development Environments (IDEs). It assists developers by suggesting relevant identifiers and APIs during coding. While completions are typically derived from static analysis, their usefulness depends heavily on how they are ranked, as correct predictions buried deep in the list are rarely seen by users. Most current systems rely on hand-crafted heuristics or lightweight machine learning models trained on user logs, which can be further improved to capture context information and generalize across projects and coding styles. In this work, we propose a new scoring approach to ranking static completions using language models in a lightweight and model-agnostic way. Our method organizes all valid completions into a prefix tree and performs a single greedy decoding pass to collect token-level scores across the tree. This enables a precise token-aware ranking without needing beam search, prompt engineering, or model adaptations. The approach is fast, architecture-agnostic, and compatible with already deployed models for code completion. These findings highlight a practical and effective pathway for integrating language models into already existing tools within IDEs, and ultimately providing smarter and more responsive developer assistance.

Long sequence modeling has gained broad interest as large language models (LLMs) continue to advance. Recent research has identified that a large portion of hidden states within the key-value caches of Transformer models can be discarded (also termed evicted) without affecting the perplexity performance in generating long sequences. However, we show that these methods, despite preserving perplexity performance, often drop information that is important for solving downstream tasks, a problem which we call information neglect. To address this issue, we introduce Chunked Instruction-aware State Eviction (CItruS), a novel modeling technique that integrates the attention preferences useful for a downstream task into the eviction process of hidden states. In addition, we design a method for chunked sequence processing to further improve efficiency. Our training-free method exhibits superior performance on long sequence comprehension and retrieval tasks over several strong baselines under the same memory budget, while preserving language modeling perplexity.

KV-cache eviction caps the memory cost of long reasoning traces but is inherently lossy because the model decodes from a partial view of its history. Under aggressive budgets, this not only lowers accuracy but can also cause runaway degeneration, where the model produces incoherent or repetitive tokens until reaching the length limit. We characterize much of this loss as an information gapf caused by missing context, rather than a capability gap caused by limited model capacity. An evicted 7B model and a full-context 1.5B model make complementary errors, and an oracle choice between their answers recovers 79% of the accuracy gap to the full-KV 7B model. Based on this observation, we propose KV-Rescue, a training-free inference framework that bridges the information gap introduced by KV eviction using a lightweight full-context helper. KV-Rescue interleaves reasoning steps from the two models into a shared trajectory. An online detector uses entropy and compressibility to terminate the generation of incoherent or repetitive base-model candidates early. Across five math benchmarks with Qwen2.5-Math 7B and 72B, KV-Rescue recovers an average of 87% of the accuracy lost to eviction at eviction budget B=64. A decode-cost analysis further shows that preventing runaway degeneration cuts base-model token generation by 43% on average.

Advances in frontier coding agents have sharply increased the token costs of AI-assisted software engineering. Their token efficiency is typically evaluated on isolated tasks, whereas real-world developers use them to continuously build, update, and maintain large-scale projects. This mismatch motivates us to evaluate agents over sequences of realistic coding tasks, where we find that agents exhibit recurring environment-dependent token costs that existing methods struggle to reduce. To minimize these costs, we introduce TokenEconomizer, an external controller that cuts environment-specific token costs over a sequence of tasks. These cuts come from determining (i) which operational micro-skills to inject, (ii) which persistently irrelevant tools to prune, and (iii) when to intervene in unproductive failure loops. Across Claude Code, Codex, and GitHub Copilot CLI paired with frontier models, TokenEconomizer reduces the average token cost by 18.4-30.9% with comparable task success. Savings hold in every repository under every tested agent-model setting and generalize to diverse domains such as sequential web development and security auditing, where costs fall by 6.4-21.2% depending on task structure. Our work establishes adaptive black-box cost control as a promising direction for reducing token costs of coding agents in realistic settings.

Our study contributes to the scheduling and combinatorial optimization literature with new heuristics discovered by leveraging the power of Large Language Models (LLMs). We focus on the single-machine total tardiness (SMTT) problem, which aims to minimize total tardiness by sequencing n jobs on a single processor without preemption, given processing times and due dates. We develop and benchmark two novel LLM-discovered heuristics, the EDD Challenger (EDDC) and MDD Challenger (MDDC), inspired by the well-known Earliest Due Date (EDD) and Modified Due Date (MDD) rules. In contrast to prior studies that employed simpler rule-based heuristics, we evaluate our LLM-discovered algorithms using rigorous criteria, including optimality gaps and solution time derived from a mixed-integer programming (MIP) formulation of SMTT. We compare their performance against state-of-the-art heuristics and exact methods across various job sizes (20, 100, 200, and 500 jobs). For instances with more than 100 jobs, exact methods such as MIP and dynamic programming become computationally intractable. Up to 500 jobs, EDDC improves upon the classic EDD rule and another widely used algorithm in the literature. MDDC consistently outperforms traditional heuristics and remains competitive with exact approaches, particularly on larger and more complex instances. This study shows that human-LLM collaboration can produce scalable, high-performing heuristics for NP-hard constrained combinatorial optimization, even under limited resources when effectively configured.

Chain-of-Thought (CoT) prompting is a key technique for enabling complex reasoning in large language models. However, generating full, fixed-length rationales is computationally wasteful, inflating both token usage and latency. We introduce LEASH: Logit-Entropy Adaptive Stopping Heuristic, a training-free decoding algorithm that adaptively halts rationale generation. LEASH monitors two intrinsic signals: the slope of token-level entropy and the improvement in the top-logit margin. It terminates the generation once both signals plateau, indicating the model has reached a stable reasoning state. Across four instruction-tuned models on the GSM8K and AQuA-RAT benchmarks, LEASH reduces average token generation by 30--35% and latency by 27%, while incurring a 10 p.p. accuracy drop relative to CoT. LEASH is model-agnostic and requires no additional training or supervision, offering a simple and efficient alternative to CoT decoding.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler

Diffusion Language Models (DLMs) offer order-agnostic generation that can explore many possible decoding trajectories. However, current decoding methods commit to a single trajectory, limiting exploration in trajectory space. We introduce Order-Token Search to explore this space through jointly searching over generation order and token values. Its core is a likelihood estimator that scores denoising actions, enabling stable pruning and efficient exploration of diverse trajectories. Across mathematical reasoning and coding benchmarks, Order-Token Search consistently outperforms baselines on GSM8K, MATH500, Countdown, and HumanEval (3.1%, 3.8%, 7.9%, and 6.8% absolute over backbone), matching or surpassing diffu-GRPO post-trained d1-LLaDA. Our work establishes joint search as a key component for advancing decoding in DLMs.

Recent advances have demonstrated that integrating reinforcement learning with rule-based rewards can significantly enhance the reasoning capabilities of large language models, even without supervised fine-tuning. However, prevalent reinforcement learning algorithms such as GRPO and its variants like DAPO, suffer from a coarse granularity issue when computing the advantage. Specifically, they compute rollout-level advantages that assign identical values to every token within a sequence, failing to capture token-specific contributions and hindering effective learning. To address this limitation, we propose Key-token Advantage Estimation (KTAE) - a novel algorithm that estimates fine-grained, token-level advantages without introducing additional models. KTAE leverages the correctness of sampled rollouts and applies statistical analysis to quantify the importance of individual tokens within a sequence to the final outcome. This quantified token-level importance is then combined with the rollout-level advantage to obtain a more fine-grained token-level advantage estimation. Empirical results show that models trained with GRPO+KTAE and DAPO+KTAE outperform baseline methods across five mathematical reasoning benchmarks. Notably, they achieve higher accuracy with shorter responses and even surpass R1-Distill-Qwen-1.5B using the same base model.

Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Large language models (LLMs) with extended context lengths rely on the key-value (KV) cache to support attention over prior tokens. However, maintaining the KV cache incurs substantial memory overhead, motivating KV-cache compression methods that enforce a fixed budget through eviction and merging. Modern eviction methods increasingly adopt span-based retention because preserving contiguous spans is empirically effective and better preserves semantic coherence. Yet, when combined with post-eviction merging, span-based retention concentrates merges onto a small set of span-boundary carrier tokens, producing a highly imbalanced merge pattern that exacerbates over-merging and increases information loss. To address this imbalance, we propose GRKV (Global Regression for KV Cache), a training-free KV-cache merging method that directly minimizes the discrepancy between compressed-cache and full-cache attention outputs. GRKV uses ridge-regression-based merge steps to distribute information from evicted tokens across retained tokens, while regularizing the updates to prevent over-smoothing. Across the LongBench and RULER long-context benchmarks, GRKV is the only merging method that improves overall performance with minimal overhead.

As LLM agents evolve, tokens have emerged as the core economic primitives of Agentic AI. However, their exponential consumption introduces severe computational, collaborative, and security bottlenecks. Current surveys remain fragmented across system optimization, architecture design, and trust, lacking a unified framework to evaluate the fundamental trade-off between output quality and economic cost. To bridge this gap, this survey presents the first comprehensive survey of Token Economics. By unifying computer science and economics, we conceptualize tokens as production factors, exchange mediums, and units of account. We synthesize existing literature across a four-dimensional taxonomy: (1) Micro-level (Single Agent): Optimizing budget-constrained factor substitution via neoclassical firm theory. (2) Meso-level (Multi-Agent Systems): Minimizing collaboration friction using transaction cost and principal-agent theories. (3) Macro-level (Agent Ecosystems): Addressing congestion externalities and pricing via mechanism design. (4) Security: Internalizing adversarial threats as endogenous economic constraints. Finally, we outline frontier directions, including differentiable token budgets and dynamic markets, to lay the theoretical foundation for scalable next-generation agent systems.

Large language models (LLMs) have demonstrated remarkable performance on long-context tasks, but are often bottlenecked by memory constraints. Namely, the KV cache, which is used to significantly speed up attention computations, grows linearly with context length. A suite of compression algorithms has been introduced to alleviate cache growth by evicting unimportant tokens. However, several popular strategies are targeted towards the prefill phase, i.e., processing long prompt context, and their performance is rarely assessed on reasoning tasks requiring long decoding. In particular, short but complex prompts, such as those in benchmarks like GSM8K and MATH500, often benefit from multi-step reasoning and self-reflection, resulting in thinking sequences thousands of tokens long. In this work, we benchmark the performance of several popular compression strategies on long-reasoning tasks. For the non-reasoning Llama-3.1-8B-Instruct, we determine that no singular strategy fits all, and that performance is heavily influenced by dataset type. However, we discover that H2O and our decoding-enabled variant of SnapKV are dominant strategies for reasoning models, indicating the utility of heavy-hitter tracking for reasoning traces. We also find that eviction strategies at low budgets can produce longer reasoning traces, revealing a tradeoff between cache size and inference costs.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

Efficient inference in Large Language Models (LLMs) is impeded by the growing memory demands of key-value (KV) caching, especially for longer sequences. Traditional KV cache eviction strategies, which prioritize less critical KV-pairs based on attention scores, often degrade generation quality, leading to issues such as context loss or hallucinations. To address this, we introduce Dynamic Discriminative Operations (D2O), a novel method that utilizes two-level discriminative strategies to optimize KV cache size without fine-tuning, while preserving essential context. Initially, by observing varying densities of attention weights between shallow and deep layers, we use this insight to determine which layers should avoid excessive eviction to minimize information loss. Subsequently, for the eviction strategy in each layer, D2O innovatively incorporates a compensation mechanism that maintains a similarity threshold to re-discriminate the importance of previously discarded tokens, determining whether they should be recalled and merged with similar tokens. Our approach not only achieves significant memory savings and enhances inference throughput by more than 3 times but also maintains high-quality long-text generation. Extensive experiments across various benchmarks and LLM architectures have demonstrated that D2O significantly enhances performance with a constrained KV cache budget.

As LLMs have become capable of processing more complex types of inputs, researchers have recently studied how to efficiently and affordably process possibly arbitrarily long sequences. One effective approach is to use a FIFO memory to store keys and values of an attention sublayer from past chunks to allow subsequent queries to attend. However, this approach requires a large memory and/or takes into the consideration the specific LM architecture. Moreover, due to the causal nature between the key-values in prior context and the queries at present, this approach cannot be extended to bidirectional attention such as in an encoder-decoder or PrefixLM decoder-only architecture. In this paper, we propose to use eviction policies, such as LRA and LFA, to reduce the memory size and adapt to various architectures, and we also propose the Attendre layer, a wait-to-attend mechanism by retrieving the key-value memory (K/V memory) with evicted queries in the query memory (Q memory). As a first step, we evaluate this method in the context length extension setup using the TriviaQA reading comprehension task, and show the effectiveness of the approach.

Speculative decoding reduces the inference latency of a target large language model via utilizing a smaller and faster draft model. Its performance depends on a hyperparameter K -- the candidate length, i.e., the number of candidate tokens for the target model to verify in each round. However, previous methods often use simple heuristics to choose K, which may result in sub-optimal performance. We study the choice of the candidate length K and formulate it as a Markov Decision Process. We theoretically show that the optimal policy of this Markov decision process takes the form of a threshold policy, i.e., the current speculation should stop and be verified when the probability of getting a rejection exceeds a threshold value. Motivated by this theory, we propose SpecDec++, an enhanced version of speculative decoding that adaptively determines the candidate length on the fly. We augment the draft model with a trained acceptance prediction head to predict the conditional acceptance probability of the candidate tokens. SpecDec++ will stop the current speculation when the predicted probability that at least one token gets rejected exceeds a threshold. We implement SpecDec++ and apply it to the llama-2-chat 7B & 70B model pair. Our adaptive method achieves a 2.04x speedup on the Alpaca dataset (an additional 7.2% improvement over the baseline speculative decoding). On the GSM8K and HumanEval datasets, our method achieves a 2.26x speedup (9.4% improvement) and 2.23x speedup (11.1% improvement), respectively.

There has been considerable recent interest in scoring properties on the basis of eviction risk. The success of methods for eviction prediction is typically evaluated using different measures of predictive accuracy. However, the underlying goal of such prediction is to direct appropriate assistance to households that may be at greater risk so they remain stably housed. Thus, we must ask the question of how useful such predictions are in targeting outreach efforts - informing action. In this paper, we investigate this question using a novel dataset that matches information on properties, evictions, and owners. We perform an eviction prediction task to produce risk scores and then use these risk scores to plan targeted outreach policies. We show that the risk scores are, in fact, useful, enabling a theoretical team of caseworkers to reach more eviction-prone properties in the same amount of time, compared to outreach policies that are either neighborhood-based or focus on buildings with a recent history of evictions. We also discuss the importance of neighborhood and ownership features in both risk prediction and targeted outreach.

Long-horizon LLM agents accumulate tool results, files, plans, and user constraints that are too structured to be treated as a disposable text suffix. Current systems mostly rely on in-run heuristics such as chronological pruning and tool-output masking, or on final self-summary near a context limit. Heuristics are cheap but blind to future dependencies; summaries preserve narrative state but often hide exact evidence, locators, and editable artifacts. We present Self-GC, where GC denotes self-governing context while deliberately echoing garbage collection: the system does not merely reclaim unused tokens, but governs the lifecycle of agent context objects. Self-GC turns user turns, tool spans, and skill state into indexed objects; asks a side-channel planner to propose fold, mask, and prune actions; and lets the harness enforce recoverable sidecars, safe commit boundaries, and cache-aware commit. On a 33-session Hard Set, Self-GC prunes 43.95% of prefix tokens while leaving 84.85% of future continuations unaffected, compared with no-impact rates of 54.55% to 69.70% for heuristic baselines. On a 332-session production-derived suite, three planner backbones reach no-impact rates of 91.27% to 94.58%, while baselines remain at 77.71% to 87.46%. In production, an online account-level split reduces daytime average input tokens by 10% to 15%, with peak reductions near 20%. These results point to context management as runtime lifecycle control over indexed, recoverable objects rather than post hoc text cleanup.

When using agent-task datasets to enhance agent capabilities for Large Language Models (LLMs), current methodologies often treat all tokens within a sample equally. However, we argue that tokens serving different roles - specifically, reasoning tokens versus boilerplate tokens (e.g., those governing output format) - differ significantly in importance and learning complexity, necessitating their disentanglement and distinct treatment. To address this, we propose a novel Shuffle-Aware Discriminator (SHAD) for adaptive token discrimination. SHAD classifies tokens by exploiting predictability differences observed after shuffling input-output combinations across samples: boilerplate tokens, due to their repetitive nature among samples, maintain predictability, whereas reasoning tokens do not. Using SHAD, we propose the Reasoning-highlighted Fine-Tuning (RFT) method, which adaptively emphasizes reasoning tokens during fine-tuning, yielding notable performance gains over common Supervised Fine-Tuning (SFT).

KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM's PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.

We propose an acceleration scheme for large language models (LLMs) through Speculative Decoding with Semantic Adaptive Tokens (SDSAT). The primary objective of this design is to enhance the LLM model's ability to generate draft tokens more accurately without compromising the model's accuracy. The core strategies involve: 1) Fine-tune the model by incorporating semantic adaptive tokens that possess flexible decoding capabilities without changing its structure, allowing them to generate high-quality draft tokens. 2) By employing a training method that does not affect the standard tokens, the model can acquire parallel decoding abilities atop its original framework with minimal training overhead. 3) We have designed the "two-step-draft-then-verify" generation strategies using both greedy search and nucleus sampling. Experiments conducted on the CodeLlama-13B and 7B models have yielded speed increases of over 3.5X and 3.0X, respectively. Please refer to https://github.com/hasuoshenyun/SDSAT.

Fine tuning has been regarded as a de facto approach for adapting large language models (LLMs) to downstream tasks, but the high training memory consumption inherited from LLMs makes this process inefficient. Among existing memory efficient approaches, activation-related optimization has proven particularly effective, as activations consistently dominate overall memory consumption. Although prior arts offer various activation optimization strategies, their data-agnostic nature ultimately results in ineffective and unstable fine tuning. In this paper, we propose TokenSeek, a universal plugin solution for various transformer-based models through instance-aware token seeking and ditching, achieving significant fine-tuning memory savings (e.g., requiring only 14.8% of the memory on Llama3.2 1B) with on-par or even better performance. Furthermore, our interpretable token seeking process reveals the underlying reasons for its effectiveness, offering valuable insights for future research on token efficiency. Homepage: https://runjia.tech/iclr_tokenseek/

Greedy search methods like Greedy Best-First Search (GBFS) and Enforced Hill-Climbing (EHC) often struggle when faced with Uninformed Heuristic Regions (UHRs) like heuristic local minima or plateaus. In this work, we theoretically and empirically compare two popular methods for escaping UHRs in breadth-first search (BrFS) and restarting random walks (RRWs). We first derive the expected runtime of escaping a UHR using BrFS and RRWs, based on properties of the UHR and the random walk procedure, and then use these results to identify when RRWs will be faster in expectation than BrFS. We then evaluate these methods for escaping UHRs by comparing standard EHC, which uses BrFS to escape UHRs, to variants of EHC called EHC-RRW, which use RRWs for that purpose. EHC-RRW is shown to have strong expected runtime guarantees in cases where EHC has previously been shown to be effective. We also run experiments with these approaches on PDDL planning benchmarks to better understand their relative effectiveness for escaping UHRs.

Entropy-based pruning has been proposed as an effective method for compressing Chain-of-Thought (CoT) reasoning with negligible accuracy loss. We test the robustness of low- and high-entropy CoT step selection methods across various models and reasoning tasks, showing that entropy offers no advantage over random pruning in any evaluated setting. Moving from sentences to tokens, we then show that retaining low-entropy tokens seems effective only on mathematical benchmarks. We find this is due to the inherently low-entropy nature of numeric tokens, which also convey semantic content in such problems. Finally, we demonstrate that patching a subset of a few CoT tokens with their original activations recovers near-perfect full-trace performance, providing causal evidence that task information is not concentrated in a small set of CoT tokens identifiable by heuristics, but rather distributed across the full reasoning chain.

Automated heuristic design (AHD) with large language models (LLMs) has produced strong heuristics for combinatorial optimization problems (COPs). Yet existing frameworks optimize for average performance on a small fixed dataset and steer the search with "verbal gradients" distilled from scalar better/worse feedback. No single heuristic dominates across instance distributions, and scalar feedback tells the LLM whether a heuristic improved, but not where in the instance space or why. We propose MOSAIC, a grid-based framework that adversarially co-evolves problem instances and specialist heuristics inside a Quality-Diversity (QD) archive indexed by structural instance features. Instances evolve to expose weaknesses of the current heuristics, and heuristics evolve to eliminate them by specializing to the newly exposed regions. Each archive cell keeps a specialist heuristic, representative instances, and insights explaining what works in its region, forming a persistent memory that accumulates over the evolutionary search. For each heuristic pair sampled from distant grid regions, an LLM-guided evolutionary loop generates discriminative instances, and a decision tree identifies the feature-space regions where each heuristic wins. A reflection LLM then contrasts the two heuristics to produce multi-directional insights that persist in those regions and guide crossover and mutation. The archive is simultaneously a co-evolved benchmark of discriminative instances and a pool of region specialist heuristics, from which greedy selection extracts a compact complementary portfolio. Across COPs, test sizes, and LLM backbones, the portfolio consistently outperforms state-of-the-art LLM-based AHD methods, and the co-evolved instances attain higher feature-space coverage and stronger heuristic discrimination than evolutionary instance-generation baselines.

Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 sim 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.

Web proxies such as NGINX commonly rely on least-recently-used (LRU) eviction, which is size agnostic and can thrash under periodic bursts and mixed object sizes. We introduce Cold-RL, a learned eviction policy for NGINX that replaces LRU's forced-expire path with a dueling Deep Q-Network served by an ONNX sidecar within a strict microsecond budget. On each eviction, Cold-RL samples the K least-recently-used objects, extracts six lightweight features (age, size, hit count, inter-arrival time, remaining TTL, and last origin RTT), and requests a bitmask of victims; a hard timeout of 500 microseconds triggers immediate fallback to native LRU. Policies are trained offline by replaying NGINX access logs through a cache simulator with a simple reward: a retained object earns one point if it is hit again before TTL expiry. We compare against LRU, LFU, size-based, adaptive LRU, and a hybrid baseline on two adversarial workloads. With a 25 MB cache, Cold-RL raises hit ratio from 0.1436 to 0.3538, a 146 percent improvement over the best classical baseline; at 100 MB, from 0.7530 to 0.8675, a 15 percent gain; and at 400 MB it matches classical methods (about 0.918). Inference adds less than 2 percent CPU overhead and keeps 95th percentile eviction latency within budget. To our knowledge, this is the first reinforcement learning eviction policy integrated into NGINX with strict SLOs.

We present Context Window Lifecycle (CWL), a context-management scheme that gives long-horizon LLM agents an effectively unbounded working horizon. As a session accumulates history, CWL keeps the context within budget through graduated, semantically-aware eviction: the agent annotates its trajectory as typed, dependency-linked episodes as work proceeds, and a deterministic, LLM-free policy evicts content in priority order within that structure when a token budget is exceeded. CWL preserves user turns and the exploratory context the agent is actively reasoning over, while aggressively shedding action episodes whose effects are already persisted in the environment, keeping active context near a stable ceiling that also avoids the performance degradation associated with very large prompts. Compared to summarization-based compaction, CWL avoids four well-known limitations: unpredictable lossiness, destruction of causal structure, blocking model cost, and compression-induced hallucination. Compared to recency truncation, CWL is semantically aware: it drops the oldest-and-most-recoverable content according to the dependency graph rather than oldest-in-time regardless of relevance. We describe the annotation protocol, the episode graph, the eviction policy, and the token-accounting loop, and evaluate CWL on long-horizon agentic benchmarks: a single agent session completing 89 sequential tasks across 80 million tokens with no measurable degradation in task accuracy relative to per-task isolated sessions

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available at https://github.com/thu-coai/HPSS.

Reasoning methods such as chain-of-thought prompting and self-consistency have shown immense potential to improve the accuracy of large language models across various reasoning tasks. However such methods involve generation of lengthy reasoning chains, which substantially increases token consumption, computational cost, and latency. To address this inefficiency, we propose ProofSketch, a verification-guided reasoning framework that integrates symbolic closure computation, lexicographic verification and adaptive sketch generation. Our experiments show that ProofSketch consistently reduces token usage while improving accuracy, demonstrating that this approach offers a promising path for efficient and trustworthy reasoning.

Autoregressive large language models (LLMs) have made remarkable progress in various natural language generation tasks. However, they incur high computation cost and latency resulting from the autoregressive token-by-token generation. To address this issue, several approaches have been proposed to reduce computational cost using early-exit strategies. These strategies enable faster text generation using reduced computation without applying the full computation graph to each token. While existing token-level early exit methods show promising results for online inference, they cannot be readily applied for batch inferencing and Key-Value caching. This is because they have to wait until the last token in a batch exits before they can stop computing. This severely limits the practical application of such techniques. In this paper, we propose a simple and effective token-level early exit method, SkipDecode, designed to work seamlessly with batch inferencing and KV caching. It overcomes prior constraints by setting up a singular exit point for every token in a batch at each sequence position. It also guarantees a monotonic decrease in exit points, thereby eliminating the need to recompute KV Caches for preceding tokens. Rather than terminating computation prematurely as in prior works, our approach bypasses lower to middle layers, devoting most of the computational resources to upper layers, allowing later tokens to benefit from the compute expenditure by earlier tokens. Our experimental results show that SkipDecode can obtain 2x to 5x inference speedups with negligible regression across a variety of tasks. This is achieved using OPT models of 1.3 billion and 6.7 billion parameters, all the while being directly compatible with batching and KV caching optimization techniques.

The Key-Value (KV) cache is crucial for efficient Large Language Models (LLMs) inference, but excessively long contexts drastically increase KV cache memory footprint. Existing KV cache compression methods typically rely on input-side attention patterns within a prompt observation window to estimate token importance during the prefill stage. They fail to preserve critical tokens for future generation since these assessments are not derived from the decoding process. Intuitively, an effective observation window should mirror the decoding-stage queries to accurately reflect which tokens the generation process will attend to. However, ground-truth decoding queries are inherently unavailable during inference. For constructing pseudo queries to approximate them, we find that positional information plays a more critical role than semantic content. Motivated by this insight, we propose decoding-aligned KV cache compression via position-aware pseudo queries (DapQ), a novel and lightweight eviction framework that leverages position-aware pseudo queries to simulate the output tokens, thereby establishing an effective observation window for importance assessment. It aligns closely with the actual generation context and enables precise token eviction. Extensive evaluations across multiple benchmarks and LLMs demonstrate that DapQ achieves superior performance, particularly under strict memory constraints (e.g., up to nearly lossless performance 99.5% on NIAH with 3% KV cache budgets).

Unlearning in large language models (LLMs) has emerged as a promising safeguard against adversarial behaviors. When the forgetting loss is applied uniformly without considering token-level semantic importance, model utility can be unnecessarily degraded. Recent studies have explored token-wise loss regularizers that prioritize informative tokens, but largely rely on ground-truth confidence or external linguistic parsers, which limits their ability to capture contextual information or the model's overall predictive state. Intuitively, function words like "the" primarily serve syntactic roles and are highly predictable with little ambiguity, but informative words admit multiple plausible alternatives with greater uncertainty. Based on this intuition, we propose Entropy-guided Token Weighting (ETW), a token-level unlearning regularizer that uses entropy of the predictive distribution as a proxy for token informativeness. We demonstrate that informative tokens tend to have higher entropy, whereas structural tokens tend to have lower entropy. This behavior enables ETW to achieve more effective unlearning while better preserving model utility than existing token-level approaches.

Private large language model (LLM) inference based on secure multi-party computation (MPC) offers cryptographically-secure protection for both user prompt and proprietary model weights. However, it suffers from large latency overhead especially for long input sequences. While key-value (KV) cache eviction algorithms have been proposed to reduce the computation and memory cost for plaintext inference, they are not designed for MPC and cannot benefit private inference easily. In this paper, we propose an accurate and MPC-friendly KV cache eviction framework, dubbed MPCache. MPCache is built on the observation that historical tokens in a long sequence may have different effects on the downstream decoding. Hence, MPCache combines a look-once static eviction algorithm to discard unimportant tokens and a query-aware dynamic selection algorithm to further select a small subset of tokens for attention computation. As existing dynamic selection algorithms incur too much latency, we propose a series of optimizations to drastically reduce the KV cache selection overhead, including MPC-friendly similarity approximation, hierarchical KV cache clustering, and cross-layer index sharing strategy. With extensive experiments, we demonstrate that MPCache consistently outperforms prior-art KV cache eviction baselines across different LLM generation tasks and achieves 1.8~2.01x and 3.39~8.37x decoding latency and communication reduction on different sequence lengths, respectively.

Efficient key-value (KV) cache compression is critical for scaling transformer-based Large Language Models (LLMs) in long sequences and resource-limited settings. Existing methods evict tokens based on their positions or importance scores, but position-based strategies can miss crucial information outside predefined regions, while those relying on global importance scores resulting in strong regional biases, limiting the KV cache's overall context retention and potentially impairing the performance of LLMs on complex tasks. Our wavelet analysis reveals that as tokens approach the end of sequence, their contributions to generation gradually increase and tends to diverge more from neighboring tokens, indicating a smooth transition with increasing complexity and variability from distant to nearby context. Motivated by this observation, we propose TreeKV, an intuitive, training-free method that employs a tree structure for smooth cache compression. TreeKV maintains a fixed cache size, allowing LLMs to deliver high-quality output even in long text scenarios. Unlike most compression methods, TreeKV is applicable to both the generation and prefilling stages. TreeKV consistently surpasses all baseline models in language modeling tasks on PG19 and OpenWebText2, allowing LLMs trained with short context window to generalize to longer window with a 16x cache reduction. On the Longbench benchmark, TreeKV achieves the best performance with only 6\% of the budget at optimal efficiency.

We present PEEK, a lightweight scheduling and eviction framework for both online (streaming) and offline (batch) LLM serving; this paper focuses on the online regime. PEEK maintains an incremental radix tree over the pending queue, exposing prefix-sharing clusters no existing engine surfaces. A low-overhead dual-walk matches the tree against the engine's prefix cache to yield longest-prefix-match for every waiting request; PEEK then admits cluster pioneers first so siblings inherit the freshly cached prefix, a co-designed eviction hook protects blocks ancestral to queued demand, and a multi-lane stride scheduler bounds starvation. On SGLang and vLLM across five workloads up to 4timesH100 (DP=2 over TP=2), PEEK delivers up to 3.0times/2.6times cache hit, 7.9times/7.1times TTFT, 6.7times/5.5times E2E, and 3.6times/4.5times throughput gains over each engine's strongest stock baseline (SGLang/vLLM), while matching baselines within noise on workloads with no exploitable prefix structure. Wins hold as KV-cache pressure and inference parallelism scale.

Speculative Decoding (SD) is a widely used approach to accelerate the inference of large language models (LLMs) without reducing generation quality. It operates by first using a compact model to draft multiple tokens efficiently, followed by parallel verification using the target LLM. This approach leads to faster inference compared to auto-regressive decoding. While there are multiple approaches to create a draft model, one promising approach is to use early-exit methods. These methods draft candidate tokens by using a subset of layers of the primary model and applying the remaining layers for verification, allowing a single model to handle both drafting and verification. While this technique reduces memory usage and computational cost, its performance relies on the choice of the exit layer for drafting and the number of tokens drafted (speculation length) in each SD round. Prior works use hyperparameter exploration to statically select these values. However, our evaluations show that these hyperparameter values are task-specific, and even within a task they are dependent on the current sequence context. We introduce DEL, a plug-and-play method that adaptively selects the exit layer and speculation length during inference. DEL dynamically tracks the token acceptance rate if the tokens are drafted at each layer of an LLM and uses that knowledge to heuristically select the optimal exit layer and speculation length. Our experiments across a broad range of models and downstream tasks show that DEL achieves overall speedups of 2.16timessim2.50times over vanilla auto-regressive decoding and improves upon the state-of-the-art SD methods by up to 0.27times.

Heuristic design with large language models (LLMs) has emerged as a promising approach for tackling combinatorial optimization problems (COPs). However, existing approaches often rely on manually predefined evolutionary computation (EC) optimizers and single-task training schemes, which may constrain the exploration of diverse heuristic algorithms and hinder the generalization of the resulting heuristics. To address these issues, we propose Meta-Optimization of Heuristics (MoH), a novel framework that operates at the optimizer level, discovering effective optimizers through the principle of meta-learning. Specifically, MoH leverages LLMs to iteratively refine a meta-optimizer that autonomously constructs diverse optimizers through (self-)invocation, thereby eliminating the reliance on a predefined EC optimizer. These constructed optimizers subsequently evolve heuristics for downstream tasks, enabling broader heuristic exploration. Moreover, MoH employs a multi-task training scheme to promote its generalization capability. Experiments on classic COPs demonstrate that MoH constructs an effective and interpretable meta-optimizer, achieving state-of-the-art performance across various downstream tasks, particularly in cross-size settings.

Heuristic algorithms play a vital role in solving combinatorial optimization (CO) problems, yet traditional designs depend heavily on manual expertise and struggle to generalize across diverse instances. We introduce HeurAgenix, a two-stage hyper-heuristic framework powered by large language models (LLMs) that first evolves heuristics and then selects among them automatically. In the heuristic evolution phase, HeurAgenix leverages an LLM to compare seed heuristic solutions with higher-quality solutions and extract reusable evolution strategies. During problem solving, it dynamically picks the most promising heuristic for each problem state, guided by the LLM's perception ability. For flexibility, this selector can be either a state-of-the-art LLM or a fine-tuned lightweight model with lower inference cost. To mitigate the scarcity of reliable supervision caused by CO complexity, we fine-tune the lightweight heuristic selector with a dual-reward mechanism that jointly exploits singals from selection preferences and state perception, enabling robust selection under noisy annotations. Extensive experiments on canonical benchmarks show that HeurAgenix not only outperforms existing LLM-based hyper-heuristics but also matches or exceeds specialized solvers. Code is available at https://github.com/microsoft/HeurAgenix.

Despite their ubiquity in language generation, it remains unknown why truncation sampling heuristics like nucleus sampling are so effective. We provide a theoretical explanation for the effectiveness of the truncation sampling by proving that truncation methods that discard tokens below some probability threshold (the most common type of truncation) can guarantee that all sampled tokens have nonzero true probability. However, thresholds are a coarse heuristic, and necessarily discard some tokens with nonzero true probability as well. In pursuit of a more precise sampling strategy, we show that we can leverage a known source of model errors, the softmax bottleneck, to prove that certain tokens have nonzero true probability, without relying on a threshold. Based on our findings, we develop an experimental truncation strategy and the present pilot studies demonstrating the promise of this type of algorithm. Our evaluations show that our method outperforms its threshold-based counterparts under automatic and human evaluation metrics for low-entropy (i.e., close to greedy) open-ended text generation. Our theoretical findings and pilot experiments provide both insight into why truncation sampling works, and make progress toward more expressive sampling algorithms that better surface the generative capabilities of large language models.

While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate. Prevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fails to generalize to diverse query distributions. In this paper, we propose DySpec, a faster speculative decoding algorithm with a novel dynamic token tree structure. We begin by bridging the draft distribution and acceptance rate from intuitive and empirical clues, and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, DySpec yields a higher acceptance rate and speedup than fixed trees. DySpec can drastically improve the throughput and reduce the latency of token generation across various data distribution and model sizes, which significantly outperforms strong competitors, including Specinfer and Sequoia. Under low temperature setting, DySpec can improve the throughput up to 9.1times and reduce the latency up to 9.4times on Llama2-70B. Under high temperature setting, DySpec can also improve the throughput up to 6.21times, despite the increasing difficulty of speculating more than one token per step for draft model.

Prompt-based learning has been an effective paradigm for large pretrained language models (LLM), enabling few-shot or even zero-shot learning. Black-box prompt search has received growing interest recently for its distinctive properties of gradient-free optimization, proven particularly useful and powerful for model-as-a-service usage. However, the discrete nature and the complexity of combinatorial optimization hinder the efficiency of modern black-box approaches. Despite extensive research on search algorithms, the crucial aspect of search space design and optimization has been largely overlooked. In this paper, we first conduct a sensitivity analysis by prompting LLM, revealing that only a small number of tokens exert a disproportionate amount of influence on LLM predictions. Leveraging this insight, we propose the Clustering and Pruning for Efficient Black-box Prompt Search (ClaPS), a simple black-box search method that first clusters and prunes the search space to focus exclusively on influential prompt tokens. By employing even simple search methods within the pruned search space, ClaPS achieves state-of-the-art performance across various tasks and LLMs, surpassing the performance of complex approaches while significantly reducing search costs. Our findings underscore the critical role of search space design and optimization in enhancing both the usefulness and the efficiency of black-box prompt-based learning.

Recent years have seen considerable advancements in multi-step reasoning with Large Language Models (LLMs). The previous studies have elucidated the merits of integrating feedback or search mechanisms during model inference to improve the reasoning accuracy. The Process-Supervised Reward Model (PRM), typically furnishes LLMs with step-by-step feedback during the training phase, akin to Proximal Policy Optimization (PPO) or reject sampling. Our objective is to examine the efficacy of PRM in the inference phase to help discern the optimal solution paths for multi-step tasks such as mathematical reasoning and code generation. To this end, we propose a heuristic greedy search algorithm that employs the step-level feedback from PRM to optimize the reasoning pathways explored by LLMs. This tailored PRM demonstrated enhanced results compared to the Chain of Thought (CoT) on mathematical benchmarks like GSM8K and MATH. Additionally, to explore the versatility of our approach, we develop a novel method to automatically generate step-level reward dataset for coding tasks and observed similar improved performance in the code generation tasks. Thus highlighting the robust nature of our reward-model-based approach to inference for reasoning tasks.

Multimodal large language models (MLLMs) incur substantial inference cost due to the processing of hundreds of visual tokens per image. Although token pruning has proven effective for accelerating inference, determining when and where to prune remains largely heuristic. Existing approaches typically rely on static, empirically selected layers, which limit interpretability and transferability across models. In this work, we introduce a matrix-entropy perspective and identify an "Entropy Collapse Layer" (ECL), where the information content of visual representations exhibits a sharp and consistent drop, which provides a principled criterion for selecting the pruning stage. Building on this observation, we propose EntropyPrune, a novel matrix-entropy-guided token pruning framework that quantifies the information value of individual visual tokens and prunes redundant ones without relying on attention maps. Moreover, to enable efficient computation, we exploit the spectral equivalence of dual Gram matrices, reducing the complexity of entropy computation and yielding up to a 64x theoretical speedup. Extensive experiments on diverse multimodal benchmarks demonstrate that EntropyPrune consistently outperforms state-of-the-art pruning methods in both accuracy and efficiency. On LLaVA-1.5-7B, our method achieves a 68.2% reduction in FLOPs while preserving 96.0% of the original performance. Furthermore, EntropyPrune generalizes effectively to high-resolution and video-based models, highlighting the strong robustness and scalability in practical MLLM acceleration. The code will be publicly available at https://github.com/YahongWang1/EntropyPrune.

Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclear. We formulate three regularity conditions, namely Completeness, Prefix Consistency, and Neutrality, and prove that they uniquely determine token-level credit. This characterization provides a unified basis for explaining phenomena across existing algorithms and guides the development of an improved actor-critic training procedure. Through this lens, an ideal teacher in On-Policy Distillation (OPD) acts as an implicit critic, yielding an expected policy gradient proportional to that induced by token-level credit. Response-level REINFORCE Leave-One-Out (RLOO) signals match the expected policy-gradient contribution of token-level credit despite their coarser granularity. We further establish approximate credit sparsity under bounded outcome rewards and show how intermediate critic errors in Generalized Advantage Estimation (GAE) can become comparable to the underlying credit. These motivate Policy Aligned Critic Training (PACT), which adopts an Actor-then-Critic update order to apply importance sampling correction to critic training and better align the critic with the updated policy. In agentic mathematical reasoning, PACT achieves 72.87% average accuracy across four benchmarks, outperforming GRPO and PPO by 8.80 and 13.16 percentage points, respectively. On SWE-bench Verified, PACT achieves a pass rate of 67.4%, outperforming PPO, GRPO, and SAO by 2.4, 2.0, and 3.8 percentage points, respectively.

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

Cognitive psychologists have documented that humans use cognitive heuristics, or mental shortcuts, to make quick decisions while expending less effort. While performing annotation work on crowdsourcing platforms, we hypothesize that such heuristic use among annotators cascades on to data quality and model robustness. In this work, we study cognitive heuristic use in the context of annotating multiple-choice reading comprehension datasets. We propose tracking annotator heuristic traces, where we tangibly measure low-effort annotation strategies that could indicate usage of various cognitive heuristics. We find evidence that annotators might be using multiple such heuristics, based on correlations with a battery of psychological tests. Importantly, heuristic use among annotators determines data quality along several dimensions: (1) known biased models, such as partial input models, more easily solve examples authored by annotators that rate highly on heuristic use, (2) models trained on annotators scoring highly on heuristic use don't generalize as well, and (3) heuristic-seeking annotators tend to create qualitatively less challenging examples. Our findings suggest that tracking heuristic usage among annotators can potentially help with collecting challenging datasets and diagnosing model biases.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to conventional Auto-Regressive (AR) Large Language Models (LLMs). By leveraging bidirectional attention and parallel decoding, dLLMs enable more efficient generation. However, they require a carefully designed denoising scheduler at inference time (absent during training) whose choice significantly impacts generation quality. While confidence-based heuristic schedulers have shown strong empirical performance, they suffer from two critical failure modes: EOS Overflow and Proximal Bias. Through in-depth analysis of the Transformer's attention patterns, we reveal that these failures stem from certain positions assigning disproportionately high attention weights to invalid tokens (e.g., [MASK] and [EOS]), which produce misleading confidence signals. Building on this insight, empirical evidence shows that valid attention scores can provide complementary guidance to conventional confidence-based heuristics, yet no single metric consistently excels across all scenarios, implying that the optimal denoising trajectory is highly context-dependent. To address this problem, we propose a lightweight evolutionary heuristic scheduler optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Our scheduler dynamically integrates multiple heuristic features with a contextual mean-field embedding, while requiring only 393 trainable parameters. Evaluated on LLaDA and Dream across four reasoning and planning benchmarks, our method consistently outperforms strong baselines, including conventional heuristics, block auto-regressive methods, and recent State-Of-The-Art (SOTA) approaches. To the best of our knowledge, it represents the most parameter-efficient neural scheduler to date. Our code is available at https://github.com/RS2002/Evo-Denoise .

Masked Diffusion Models (MDMs) have emerged as a promising non-autoregressive paradigm for generative tasks, offering parallel decoding and bidirectional context utilization. However, current sampling methods rely on simple confidence-based heuristics that ignore the long-term impact of local decisions, leading to trajectory lock-in where early hallucinations cascade into global incoherence. While search-based methods mitigate this, they incur prohibitive computational costs (O(K) forward passes per step). In this work, we propose Backward-on-Entropy (BoE) Steering, a gradient-guided inference framework that approximates infinite-horizon lookahead via a single backward pass. We formally derive the Token Influence Score (TIS) from a first-order expansion of the trajectory cost functional, proving that the gradient of future entropy with respect to input embeddings serves as an optimal control signal for minimizing uncertainty. To ensure scalability, we introduce ActiveQueryAttention, a sparse adjoint primitive that exploits the structure of the masking objective to reduce backward pass complexity. BoE achieves a superior Pareto frontier for inference-time scaling compared to existing unmasking methods, demonstrating that gradient-guided steering offers a mathematically principled and efficient path to robust non-autoregressive generation. We will release the code.

Diffusion large language models promise faster generation by refining many token positions in parallel, but this parallelism introduces a hidden control problem: which proposed tokens should be transferred into the partially decoded sequence at each step? We refer to this decision as token commitment. Existing frozen-generator decoders largely rely on hand-designed confidence rules or block-specific acceptance filters. We argue that token commitment can instead be learned as a reusable trace-state policy. We introduce TraceLock, a lightweight plug-in controller that instantiates this policy for a frozen diffusion language model. Since oracle commitment times are unavailable, TraceLock derives self-supervision from future stability: at decoding step t, a proposed token for position i is labeled stable if it matches the final token at position i after the full decoding trace completes. The controller scores variable-length trace states and decides which active token proposals should be committed to the partially decoded sequence. Once trained for a given frozen backbone, the controller can be deployed across local-window widths, generation lengths, and step budgets without retraining or per-setting calibration. Experiments on question answering, mathematical reasoning, and code generation show that TraceLock improves the quality-step tradeoff over heuristic and learned baselines, with particularly stable behavior under cross-setting deployment. Diagnostic analyses show that its decisions are not reducible to scalar confidence, suggesting that frozen diffusion language models expose a learnable space of commitment trajectories beyond confidence-based decoding. Code is available at https://github.com/BobSun98/TraceLock.

Key-Value (KV) Caching has become an essential technique for accelerating the inference speed and throughput of generative Large Language Models~(LLMs). However, the memory footprint of the KV cache poses a critical bottleneck in LLM deployment as the cache size grows with batch size and sequence length, often surpassing even the size of the model itself. Although recent methods were proposed to select and evict unimportant KV pairs from the cache to reduce memory consumption, the potential ramifications of eviction on the generative process are yet to be thoroughly examined. In this paper, we examine the detrimental impact of cache eviction and observe that unforeseen risks arise as the information contained in the KV pairs is exhaustively discarded, resulting in safety breaches, hallucinations, and context loss. Surprisingly, we find that preserving even a small amount of information contained in the evicted KV pairs via reduced precision quantization substantially recovers the incurred degradation. On the other hand, we observe that the important KV pairs must be kept at a relatively higher precision to safeguard the generation quality. Motivated by these observations, we propose Mixed-precision KV cache~(MiKV), a reliable cache compression method that simultaneously preserves the context details by retaining the evicted KV pairs in low-precision and ensure generation quality by keeping the important KV pairs in high-precision. Experiments on diverse benchmarks and LLM backbones show that our proposed method offers a state-of-the-art trade-off between compression ratio and performance, compared to other baselines.

Although Large Vision Language Models (LVLMs) have demonstrated remarkable performance in image understanding tasks, their computational efficiency remains a significant challenge, particularly on resource-constrained devices due to the high cost of processing large numbers of visual tokens. Recently, training-free visual token pruning methods have gained popularity as a low-cost solution to this issue. However, existing approaches suffer from two key limitations: semantic saliency-based strategies primarily focus on high cross-attention visual tokens, often neglecting visual diversity, whereas visual diversity-based methods risk inadvertently discarding semantically important tokens, especially under high compression ratios. In this paper, we introduce GreedyPrune, a training-free plug-and-play visual token pruning algorithm designed to jointly optimize semantic saliency and visual diversity. We formalize the token pruning process as a combinatorial optimization problem and demonstrate that greedy algorithms effectively balance computational efficiency with model accuracy. Extensive experiments validate the effectiveness of our approach, showing that GreedyPrune achieves state-of-the-art accuracy across various multimodal tasks and models while significantly reducing end-to-end inference latency.

Can a mere next-token predictor faithfully model human intelligence? We crystallize this intuitive concern, which is fragmented in the literature. As a starting point, we argue that the two often-conflated phases of next-token prediction -- autoregressive inference and teacher-forced training -- must be treated distinctly. The popular criticism that errors can compound during autoregressive inference, crucially assumes that teacher-forcing has learned an accurate next-token predictor. This assumption sidesteps a more deep-rooted problem we expose: in certain classes of tasks, teacher-forcing can simply fail to learn an accurate next-token predictor in the first place. We describe a general mechanism of how teacher-forcing can fail, and design a minimal planning task where both the Transformer and the Mamba architecture empirically fail in that manner -- remarkably, despite the task being straightforward to learn. We provide preliminary evidence that this failure can be resolved when training to predict multiple tokens in advance. We hope this finding can ground future debates and inspire explorations beyond the next-token prediction paradigm. We make our code available under https://github.com/gregorbachmann/Next-Token-Failures

Recently, coding agents have emerged as a dominant paradigm for real-world software engineering (SWE) scenarios, which solve complex tasks through multi-turn interactions with development environments. However, frequent interactions with environments would inevitably introduce substantial token overhead, leading to high usage costs and latency. Although recent studies have explored reducing token usage by context manipulation and interaction limits at inference time, these approaches focus on improving token efficiency while overlooking the risk of discarding task-relevant information, thus struggling to balance the trade-off between resolution rate and token efficiency. In this paper, we study a more general paradigm without suffering from the limitation, i.e., training token-efficient coding agents with promising resolution performance, which is a highly-practical yet less-explored problem. To this end, we reveal two core observations in SWE scenarios: i) Efficiency Variation: successful resolution could be achieved with fewer tokens; ii) Entropy Correlation: unproductive behaviors are associated with turn-level entropy. Motivated by observations, we propose a novel reinforcement learning framework, dubbed HERO. Specifically, HERO prioritizes task resolution over token efficiency during policy optimization and encourages efficient reasoning patterns at both trajectory and turn levels. Extensive experiments on SWE-bench Verified and SWE-bench Multilingual demonstrate that HERO achieves a favorable trade-off between resolution rate and token efficiency compared with state-of-the-art coding agents and reinforcement learning methods.

LLM agents that cache recovery suggestions from API errors can skip re-derivation in later episodes, spending fewer tokens and fewer model calls on constraints they have already learned. Server-side data drift turns those cached fixes into silent failures, and the usual remedy, re-deriving on every episode, gives the savings back. We introduce invalidation contracts, a protocol layer that attaches version stamps and cacheability hints to every recovery suggestion so the client can evict stale entries without trial and error, and keep the rest. The contract decomposes realized savings into two independent factors: validity, the fraction of cached suggestions that remain correct after a drift event, and compliance, the fraction the planner applies on the first attempt. Validity depends only on the protocol and is vendor-independent. Compliance depends on the planner model: identical wire bytes yield 100% first-try compliance on Claude Haiku 4.5 and 11% or below on Claude Sonnet 5, which exhibits input-schema conservatism, refusing fixes that add fields the original request did not contain. We evaluate across seven models, three serving paths, two domains, and approximately 9,400 episodes. Row-level invalidation raises compliance by 0 to 66.7 percentage points across the seven models, 55.6 to 66.7 on three, and recovers 29-33% of baseline token cost on four of seven models, while table-level invalidation destroys co-located entries and drops post-drift first-try rates to 0% on five of seven. Eviction precision is 1.00 at row granularity on every model under the row-level oracle of Section 4.1. The contract adds 15% to response payload. Version-stamp validity is deterministic by construction and produced identical results across every model and serving path, with zero contract failures in the entire evaluation.

Test-time compute allocation in large reasoning models (LRMs) is widely used and has applications in mathematical problem solving, code synthesis, and planning. Recent work has addressed this problem by scaling self-consistency and parallel thinking, adding generic ``thinking tokens'' and prompting models to re-read the question before answering. Unfortunately, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain -- and often ignore -- the spontaneous repetition that many LRMs exhibit at the head of their internal chains. In contrast, we analyze and harness the model's tendency to restate the question, which we term the Echo of Prompt (EOP), as a front-loaded, compute-shaping mechanism. We formalize its probabilistic cost by casting echo removal as rejection-based conditioning and defining the Echo Likelihood Gap ΔL as a computable proxy. This provides the missing theoretical link that links early repetition to likelihood gains and downstream accuracy. However, it does not by itself specify how to exploit EOP. Consequently, we develop Echo-Distilled SFT (ED-SFT) to instill an ``echo-then-reason'' pattern through supervised finetuning, and Echoic Prompting (EP) to re-ground the model mid-trace without training. While promising, quantifying benefits beyond verbosity is non-trivial. Therefore, we conduct length and suffix-controlled likelihood analyses together with layer-wise attention studies, showing that EOP increases answer to answer-prefix attention in middle layers, consistent with an attention refocusing mechanism. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and find consistent gains over baselines. Code is available at https://github.com/hhh2210/echoes-as-anchors.

Reinforcement Learning with Verifiable Rewards (RLVR) has propelled Large Language Models in complex reasoning, yet its scalability is often hindered by a training bottleneck where performance plateaus as policy entropy collapses, signaling a loss of exploration. Previous methods typically address this by maintaining high policy entropy, yet the precise mechanisms that govern meaningful exploration have remained underexplored. Our analysis suggests that an unselective focus on entropy risks amplifying irrelevant tokens and destabilizing training. This paper investigates the exploration dynamics within RLVR and identifies a key issue: the gradual elimination of valuable low-probability exploratory tokens, which we term \textit{reasoning sparks}. We find that while abundant in pre-trained models, these sparks are systematically extinguished during RLVR due to over-penalization, leading to a degeneracy in exploration. To address this, we introduce Low-probability Regularization (Lp-Reg). Its core mechanism regularizes the policy towards a heuristic proxy distribution. This proxy is constructed by filtering out presumed noise tokens and re-normalizing the distribution over the remaining candidates. The result is a less-noisy proxy where the probability of reasoning sparks is amplified, which then serves as a soft regularization target to shield these valuable tokens from elimination via KL divergence. Experiments show that Lp-Reg enables stable on-policy training for around 1,000 steps, a regime where baseline entropy-control methods collapse. This sustained exploration leads to state-of-the-art performance, achieving a 60.17% average accuracy on five math benchmarks, an improvement of 2.66% over prior methods. Code is available at https://github.com/CarlanLark/Lp-Reg.

Large Language Models (LLMs) typically generate outputs token by token using a fixed compute budget, leading to inefficient resource utilization. To address this shortcoming, recent advancements in mixture of expert (MoE) models, speculative decoding, and early exit strategies leverage the insight that computational demands can vary significantly based on the complexity and nature of the input. However, identifying optimal routing patterns for dynamic execution remains an open challenge, limiting the full potential of these adaptive methods. To address this need, we study adaptive computation in LLMs more systematically. We propose a novel framework that integrates smaller auxiliary modules within each Feed-Forward Network layer of the LLM. This design enables dynamic routing of tokens based on task complexity: tokens can be processed by either the small or big modules at each layer, or even bypass certain layers entirely. This allows us to introduce a novel notion of a token's difficulty, defined by its potential to benefit from additional computational resources. Importantly, by employing oracles to identify optimal patterns of adaptive computations, we gain valuable insights into the internal workings of LLMs and the routing processes in a simplified heterogeneous MoE setup. We show that trained routers operate differently from oracles and often yield suboptimal solutions. Notably, activating a large module in just one layer outperforms models that use large modules across all layers, underscoring the gap between practical implementations of routing in MoE models and theoretical optima for adaptive computation.

Reinforcement Learning has emerged as the fundamental technique for enhancing reasoning in LLMs. However, existing algorithms apply uniform optimization to all tokens, ignoring their different roles in reasoning process. To address this limitation, we introduce Heterogeneous Adaptive Policy Optimization (HAPO), a comprehensive token-aware algorithm that dynamically adapts optimization based on token entropy. For rollout sampling, we propose Adaptive Temperature Sampling, which adjusts sampling temperature in real time, promoting exploration at high-entropy tokens while preserving coherence at low-entropy ones. For advantage calculation, we introduce Token Level Group Average that normalizes advantages at token level, jointly accounting for sequence-length as in token-mean loss while preserving non-biased treatment. We then develop Differential Advantage Redistribution that leverages entropy and importance ratios to modulate rewards-adjusting updates for tokens with clear signals. For clipping loss, we design Asymmetric Adaptive Clipping, allowing aggressive probability reduction for noisy low-entropy tokens while enabling exploration for high-entropy tokens. Through systematic investigation between entropy and training dynamics, we embedded token-level treatment into every stages to achieve fine-grained control. Extensive experiments demonstrate that HAPO consistently outperforms DAPO across multiple model scales. Our code can be found in https://github.com/starriver030515/HAPO.

Speculative decoding is widely used in accelerating large language model (LLM) inference. In this work, we focus on the online draft model selection problem in speculative decoding. We design an algorithm that provably competes with the best draft model in hindsight for each query in terms of either the token acceptance probability or expected acceptance length. In particular, we show that we can accurately evaluate all draft models, instead of only the chosen model without incurring additional queries to the target model, which allows us to improve exponentially over the existing bandit-based approach as the number of draft models increases. Our approach is generically applicable with any speculative decoding methods (single draft, multi-drafts and draft-trees). Moreover, we design system-efficient versions of online learners and demonstrate that the overhead in computation and latency can be substantially reduced. We conduct extensive experiments on open-source LLMs and diverse datasets, demonstrating that our methods substantially outperform the state-of-the-art EAGLE3 and the BanditSpec baseline in a variety of domains where specialized domain-expert drafters are available, especially when long reasoning chains are required.

Reinforcement learning from verifiable rewards has significantly advanced the reasoning capabilities of large language models. However, Group Relative Policy Optimization (GRPO) typically assigns a uniform, sequence-level advantage to all tokens, thereby overlooking the intrinsic information heterogeneity along reasoning chains. We show that this coarse-grained credit assignment leads to premature entropy collapse and encourages the model to generate redundant, low-quality reasoning paths. Through systematic empirical analysis, we identify Critical Decision Pivots (CDPs): transient high-entropy states where the policy's trajectory is most sensitive to perturbations. These pivots represent the "forks in the road" where effective multi-path exploration is most crucial yet often suppressed by uniform advantage signals. Building on these insights, we propose Entropy-Regulated Policy Optimization (ERPO), which transitions the optimization focus from coarse sequences to fine-grained token dynamics. ERPO introduces three synergistic components: (i) Entropy-aware Gating, which adaptively amplifies exploration at CDPs to facilitate diverse path discovery; (ii) Bucket-based Implicit Normalization, which mitigates difficulty bias by aligning token progress windows; and (iii) Result-anchored Advantage Synthesis, which re-weights token-level signals via outcome-driven anchors. Extensive experiments on competitive mathematical benchmarks demonstrate that ERPO significantly outperforms GRPO. Notably, ERPO not only boosts reasoning accuracy but also yields significantly more concise and robust derivation paths, while achieving performance comparable to large models with orders of magnitude more parameters.

Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a "draft and then verify" mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which fail to adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we proposed OPT-Tree, an algorithm to construct adaptive and scalable draft trees. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

Diffusion (Large) Language Models (dLLMs) now match the downstream performance of their autoregressive counterparts on many tasks, while holding the promise of being more efficient during inference. One particularly successful variant is masked discrete diffusion, in which a buffer filled with special mask tokens is progressively replaced with tokens sampled from the model's vocabulary. Efficiency can be gained by unmasking several tokens in parallel, but doing too many at once risks degrading the generation quality. Thus, one critical design aspect of dLLMs is the sampling procedure that selects, at each step of the diffusion process, which tokens to replace. Indeed, recent work has found that heuristic strategies such as confidence thresholding lead to both higher quality and token throughput compared to random unmasking. However, such heuristics have downsides: they require manual tuning, and we observe that their performance degrades with larger buffer sizes. In this work, we instead propose to train sampling procedures using reinforcement learning. Specifically, we formalize masked diffusion sampling as a Markov decision process in which the dLLM serves as the environment, and propose a lightweight policy architecture based on a single-layer transformer that maps dLLM token confidences to unmasking decisions. Our experiments show that these trained policies match the performance of state-of-the-art heuristics when combined with semi-autoregressive generation, while outperforming them in the full diffusion setting. We also examine the transferability of these policies, finding that they can generalize to new underlying dLLMs and longer sequence lengths. However, we also observe that their performance degrades when applied to out-of-domain data, and that fine-grained tuning of the accuracy-efficiency trade-off can be challenging with our approach.

Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level speculative decoding (SD) helps, but its benefit is capped, because the chance that an entire gamma-token guess is correct falls exponentially as gamma grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with Lookahead Reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-step, and each step needs only to be semantically correct, not exact token matching. In Lookahead Reasoning, a lightweight draft model proposes several future steps; the target model expands each proposal in one batched pass, and a verifier keeps semantically correct steps while letting the target regenerate any that fail. Token-level SD still operates within each reasoning step, so the two layers of parallelism multiply. We show Lookahead Reasoning lifts the peak speedup of SD both theoretically and empirically. Across GSM8K, AIME, and other benchmarks, Lookahead Reasoning improves the speedup of SD from 1.4x to 2.1x while preserving answer quality, and its speedup scales better with additional GPU throughput. Our code is available at https://github.com/hao-ai-lab/LookaheadReasoning

Large reasoning models (LRMs) boosted by Reinforcement Learning from Verifier Reward (RLVR) have shown great power in problem solving, yet they often cause overthinking: excessive, meandering reasoning that inflates computational cost. Prior designs of penalization in RLVR manage to reduce token consumption while often harming model performance, which arises from the oversimplicity of token-level supervision. In this paper, we argue that the granularity of supervision plays a crucial role in balancing efficiency and accuracy, and propose Group Relative Segment Penalization (GRSP), a step-level method to regularize reasoning. Since preliminary analyses show that reasoning segments are strongly correlated with token consumption and model performance, we design a length-aware weighting mechanism across segment clusters. Extensive experiments demonstrate that GRSP achieves superior token efficiency without heavily compromising accuracy, especially the advantages with harder problems. Moreover, GRSP stabilizes RL training and scales effectively across model sizes.

Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly, as critical learning signals are diluted by an abundance of uninformative samples and tokens. To tackle this challenge, we propose the Dynamic Dual-Level Down-Sampling (D^3S) framework that prioritizes the most informative samples and tokens across groups to improve the efficient of policy optimization. D^3S operates along two levels: (1) the sample-level, which selects a subset of rollouts to maximize advantage variance (Var(A)). We theoretically proven that this selection is positively correlated with the upper bound of the policy gradient norms, yielding higher policy gradients. (2) the token-level, which prioritizes tokens with a high product of advantage magnitude and policy entropy (|A_{i,t}|times H_{i,t}), focusing updates on tokens where the policy is both uncertain and impactful. Moreover, to prevent overfitting to high-signal data, D^3S employs a dynamic down-sampling schedule inspired by curriculum learning. This schedule starts with aggressive down-sampling to accelerate early learning and gradually relaxes to promote robust generalization. Extensive experiments on Qwen2.5 and Llama3.1 demonstrate that integrating D^3S into advanced RL algorithms achieves state-of-the-art performance and generalization while requiring fewer samples and tokens across diverse reasoning benchmarks. Our code is added in the supplementary materials and will be made publicly available.

The recent advancements in large language models (LLMs) have been extraordinary, yet the escalating inference costs associated with them present challenges in real-world applications. To address these challenges, we propose a novel approach called Early-exiting Speculative Decoding (EESD) with lossless acceleration. Specifically, EESD utilizes a segment of the LLM to generate draft tokens, incorporating Early-exiting structures after the first N layers. To enhance the quality of draft tokens, a self-distillation method is integrated. This early-exiting design not only reduces deployment and training costs but also significantly accelerates the token generation speed. Moreover, we introduce a novel sampling mechanism that leverages Thompson Sampling to regulate the generation processes, automatically determining the quantity of draft tokens in each round. The original LLM is then employed to validate these draft tokens through a single forward pass, and thus guarantees that the final output text maintains a distribution consistent with vanilla auto-regressive decoding. The experimental results on both 13B and 70B models demonstrate that our approach decodes tokens at a markedly accelerated rate compared to prior methods, showing the effectiveness of our approach.

Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token's influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO's learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.

As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95% retrieval accuracy with a cache budget less than 2% of the context length in the Needle-in-a-Haystack task.

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the k most recently decoded tokens. The skip-or-recompute decision requires only O(V) computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves 15.2times-26.4times speedup on standard benchmarks and 22.4times-24.1times on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only 0.5% of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

Effective prompt design is essential for improving the planning capabilities of large language model (LLM)-driven agents. However, existing structured prompting strategies are typically limited to single-agent, plan-only settings, and often evaluate performance solely based on task accuracy - overlooking critical factors such as token efficiency, modularity, and scalability in multi-agent environments. To address these limitations, we introduce CodeAgents, a prompting framework that codifies multi-agent reasoning and enables structured, token-efficient planning in multi-agent systems. In CodeAgents, all components of agent interaction - Task, Plan, Feedback, system roles, and external tool invocations - are codified into modular pseudocode enriched with control structures (e.g., loops, conditionals), boolean logic, and typed variables. This design transforms loosely connected agent plans into cohesive, interpretable, and verifiable multi-agent reasoning programs. We evaluate the proposed framework across three diverse benchmarks - GAIA, HotpotQA, and VirtualHome - using a range of representative LLMs. Results show consistent improvements in planning performance, with absolute gains of 3-36 percentage points over natural language prompting baselines. On VirtualHome, our method achieves a new state-of-the-art success rate of 56%. In addition, our approach reduces input and output token usage by 55-87% and 41-70%, respectively, underscoring the importance of token-aware evaluation metrics in the development of scalable multi-agent LLM systems. The code and resources are available at: https://anonymous.4open.science/r/CodifyingAgent-5A86