Daily Papers

Localizing behaviors of neural networks to a subset of the network's components or a subset of interactions between components is a natural first step towards analyzing network mechanisms and possible failure modes. Existing work is often qualitative and ad-hoc, and there is no consensus on the appropriate way to evaluate localization claims. We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths. We refine an explanation of induction heads, characterize a behavior of GPT-2, and open source a framework for efficiently running similar experiments.

Transformer language models have received widespread public attention, yet their generated text is often surprising even to NLP researchers. In this survey, we discuss over 250 recent studies of English language model behavior before task-specific fine-tuning. Language models possess basic capabilities in syntax, semantics, pragmatics, world knowledge, and reasoning, but these capabilities are sensitive to specific inputs and surface features. Despite dramatic increases in generated text quality as models scale to hundreds of billions of parameters, the models are still prone to unfactual responses, commonsense errors, memorized text, and social biases. Many of these weaknesses can be framed as over-generalizations or under-generalizations of learned patterns in text. We synthesize recent results to highlight what is currently known about what large language models can and cannot do.

Synthetic data generated by large language models has become integral to modern NLP training pipelines, from bootstrapping reasoning capabilities to augmenting instruction-following datasets. While recent work demonstrates successful applications maintaining high external data ratios, systematic understanding of how synthetic data proportion affects model behavior across different scales remains limited. This paper presents a controlled empirical study examining model performance, calibration, and output characteristics when trained on varying synthetic-to-external data ratios. Using the Pythia model suite (410M-12B parameters) across five diverse tasks, we evaluate models after one to three training iterations with synthetic data proportions ranging from 0-50\%. Our key findings include: models maintain stable performance with up to 20\% synthetic data, but degradation accelerates beyond 30\%; larger models (6.9B-12B) show greater robustness to synthetic data than smaller models (410M-1.4B); calibration degradation precedes accuracy loss, providing an early warning signal; and task characteristics matter, with reasoning tasks degrading faster than retrieval tasks under synthetic data training. Importantly, we find that current best practices, such as those employed in STaR and Self-Instruct systems that maintain greater than 80\% external data, operate well within safe regimes identified by our experiments. We provide practical guidance for practitioners on synthetic data budgets based on model scale and task requirements, alongside detailed comparison with concurrent work including Shumailov et al.'s model collapse findings.

The goal of data attribution is to trace model predictions back to training data. Despite a long line of work towards this goal, existing approaches to data attribution tend to force users to choose between computational tractability and efficacy. That is, computationally tractable methods can struggle with accurately attributing model predictions in non-convex settings (e.g., in the context of deep neural networks), while methods that are effective in such regimes require training thousands of models, which makes them impractical for large models or datasets. In this work, we introduce TRAK (Tracing with the Randomly-projected After Kernel), a data attribution method that is both effective and computationally tractable for large-scale, differentiable models. In particular, by leveraging only a handful of trained models, TRAK can match the performance of attribution methods that require training thousands of models. We demonstrate the utility of TRAK across various modalities and scales: image classifiers trained on ImageNet, vision-language models (CLIP), and language models (BERT and mT5). We provide code for using TRAK (and reproducing our work) at https://github.com/MadryLab/trak .

Dialogue models are able to generate coherent and fluent responses, but they can still be challenging to control and may produce non-engaging, unsafe results. This unpredictability diminishes user trust and can hinder the use of the models in the real world. To address this, we introduce DialGuide, a novel framework for controlling dialogue model behavior using natural language rules, or guidelines. These guidelines provide information about the context they are applicable to and what should be included in the response, allowing the models to generate responses that are more closely aligned with the developer's expectations and intent. We evaluate DialGuide on three tasks in open-domain dialogue response generation: guideline selection, response generation, and response entailment verification. Our dataset contains 10,737 positive and 15,467 negative dialogue context-response-guideline triplets across two domains - chit-chat and safety. We provide baseline models for the tasks and benchmark their performance. We also demonstrate that DialGuide is effective in the dialogue safety domain, producing safe and engaging responses that follow developer guidelines.

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.

Researchers have shown that neural similarity among humans predicts social closeness and cooperative success, whereas innovation often emerges from interactions among dissimilar individuals. We investigate whether these principles extend to artificial intelligence by examining interactions between large language models. In our experiments, 276 model pairs interact across eight games spanning both cooperation and novelty. We find that pairs with more similar representation spaces achieve significantly higher cooperation but exhibit reduced novelty and creativity. The effects of representational similarity on cooperation and novelty remain robust even after controlling for other factors such as performance disparity and model size. We also find that similarity in the early layers consistently shows the strongest association with cooperation and novelty, compared to the middle and later layers. This suggests that a central factor underlying these patterns could be the extent to which the two models share lexical and semantic grounding. Overall, representational similarity can be an important consideration in multi-agent system design.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

Algorithmic approaches to interpreting machine learning models have proliferated in recent years. We carry out human subject tests that are the first of their kind to isolate the effect of algorithmic explanations on a key aspect of model interpretability, simulatability, while avoiding important confounding experimental factors. A model is simulatable when a person can predict its behavior on new inputs. Through two kinds of simulation tests involving text and tabular data, we evaluate five explanations methods: (1) LIME, (2) Anchor, (3) Decision Boundary, (4) a Prototype model, and (5) a Composite approach that combines explanations from each method. Clear evidence of method effectiveness is found in very few cases: LIME improves simulatability in tabular classification, and our Prototype method is effective in counterfactual simulation tests. We also collect subjective ratings of explanations, but we do not find that ratings are predictive of how helpful explanations are. Our results provide the first reliable and comprehensive estimates of how explanations influence simulatability across a variety of explanation methods and data domains. We show that (1) we need to be careful about the metrics we use to evaluate explanation methods, and (2) there is significant room for improvement in current methods. All our supporting code, data, and models are publicly available at: https://github.com/peterbhase/InterpretableNLP-ACL2020

The increasing size and complexity of modern ML systems has improved their predictive capabilities but made their behavior harder to explain. Many techniques for model explanation have been developed in response, but we lack clear criteria for assessing these techniques. In this paper, we cast model explanation as the causal inference problem of estimating causal effects of real-world concepts on the output behavior of ML models given actual input data. We introduce CEBaB, a new benchmark dataset for assessing concept-based explanation methods in Natural Language Processing (NLP). CEBaB consists of short restaurant reviews with human-generated counterfactual reviews in which an aspect (food, noise, ambiance, service) of the dining experience was modified. Original and counterfactual reviews are annotated with multiply-validated sentiment ratings at the aspect-level and review-level. The rich structure of CEBaB allows us to go beyond input features to study the effects of abstract, real-world concepts on model behavior. We use CEBaB to compare the quality of a range of concept-based explanation methods covering different assumptions and conceptions of the problem, and we seek to establish natural metrics for comparative assessments of these methods.

Interpreting time series models is uniquely challenging because it requires identifying both the location of time series signals that drive model predictions and their matching to an interpretable temporal pattern. While explainers from other modalities can be applied to time series, their inductive biases do not transfer well to the inherently challenging interpretation of time series. We present TimeX, a time series consistency model for training explainers. TimeX trains an interpretable surrogate to mimic the behavior of a pretrained time series model. It addresses the issue of model faithfulness by introducing model behavior consistency, a novel formulation that preserves relations in the latent space induced by the pretrained model with relations in the latent space induced by TimeX. TimeX provides discrete attribution maps and, unlike existing interpretability methods, it learns a latent space of explanations that can be used in various ways, such as to provide landmarks to visually aggregate similar explanations and easily recognize temporal patterns. We evaluate TimeX on eight synthetic and real-world datasets and compare its performance against state-of-the-art interpretability methods. We also conduct case studies using physiological time series. Quantitative evaluations demonstrate that TimeX achieves the highest or second-highest performance in every metric compared to baselines across all datasets. Through case studies, we show that the novel components of TimeX show potential for training faithful, interpretable models that capture the behavior of pretrained time series models.

Interpretability research now offers a variety of techniques for identifying abstract internal mechanisms in neural networks. Can such techniques be used to predict how models will behave on out-of-distribution examples? In this work, we provide a positive answer to this question. Through a diverse set of language modeling tasks--including symbol manipulation, knowledge retrieval, and instruction following--we show that the most robust features for correctness prediction are those that play a distinctive causal role in the model's behavior. Specifically, we propose two methods that leverage causal mechanisms to predict the correctness of model outputs: counterfactual simulation (checking whether key causal variables are realized) and value probing (using the values of those variables to make predictions). Both achieve high AUC-ROC in distribution and outperform methods that rely on causal-agnostic features in out-of-distribution settings, where predicting model behaviors is more crucial. Our work thus highlights a novel and significant application for internal causal analysis of language models.

One way to personalize and steer generations from large language models (LLM) is to assign a persona: a role that describes how the user expects the LLM to behave (e.g., a helpful assistant, a teacher, a woman). This paper investigates how personas affect diverse aspects of model behavior. We assign to seven LLMs 162 personas from 12 categories spanning variables like gender, sexual orientation, and occupation. We prompt them to answer questions from five datasets covering objective (e.g., questions about math and history) and subjective tasks (e.g., questions about beliefs and values). We also compare persona's generations to two baseline settings: a control persona setting with 30 paraphrases of "a helpful assistant" to control for models' prompt sensitivity, and an empty persona setting where no persona is assigned. We find that for all models and datasets, personas show greater variability than the control setting and that some measures of persona behavior generalize across models.

Tokenizers provide the fundamental basis through which text is represented and processed by language models (LMs). Despite the importance of tokenization, its role in LM performance and behavior is poorly understood due to the challenge of measuring the impact of tokenization in isolation. To address this need, we present TokSuite, a collection of models and a benchmark that supports research into tokenization's influence on LMs. Specifically, we train fourteen models that use different tokenizers but are otherwise identical using the same architecture, dataset, training budget, and initialization. Additionally, we curate and release a new benchmark that specifically measures model performance subject to real-world perturbations that are likely to influence tokenization. Together, TokSuite allows robust decoupling of the influence of a model's tokenizer, supporting a series of novel findings that elucidate the respective benefits and shortcomings of a wide range of popular tokenizers.

LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model's true reasoning process is poorly understood. Existing faithfulness metrics have critical limitations, typically relying on identifying unfaithfulness via adversarial prompting or detecting reasoning errors. These methods overlook the predictive value of explanations. We introduce Normalized Simulatability Gain (NSG), a general and scalable metric based on the idea that a faithful explanation should allow an observer to learn a model's decision-making criteria, and thus better predict its behavior on related inputs. We evaluate 18 frontier proprietary and open-weight models, e.g., Gemini 3, GPT-5.2, and Claude 4.5, on 7,000 counterfactuals from popular datasets covering health, business, and ethics. We find self-explanations substantially improve prediction of model behavior (11-37% NSG). Self-explanations also provide more predictive information than explanations generated by external models, even when those models are stronger. This implies an advantage from self-knowledge that external explanation methods cannot replicate. Our approach also reveals that, across models, 5-15% of self-explanations are egregiously misleading. Despite their imperfections, we show a positive case for self-explanations: they encode information that helps predict model behavior.

In recent years, language models have drastically grown in size, and the abilities of these models have been shown to improve with scale. The majority of recent scaling laws studies focused on high-compute high-parameter count settings, leaving the question of when these abilities begin to emerge largely unanswered. In this paper, we investigate whether the effects of pre-training can be observed when the problem size is reduced, modeling a smaller, reduced-vocabulary language. We show the benefits of pre-training with masked language modeling (MLM) objective in models as small as 1.25M parameters, and establish a strong correlation between pre-training perplexity and downstream performance (GLUE benchmark). We examine downscaling effects, extending scaling laws to models as small as ~1M parameters. At this scale, we observe a break of the power law for compute-optimal models and show that the MLM loss does not scale smoothly with compute-cost (FLOPs) below 2.2 times 10^{15} FLOPs. We also find that adding layers does not always benefit downstream performance.

Large language models perform strongly on benchmarks in mathematical reasoning, coding and document analysis, suggesting a broad ability to follow instructions. However, it remains unclear whether such success reflects general logical competence, repeated application of learned procedures, or pattern matching that mimics rule execution. We investigate this question by introducing Stable Counting Capacity, an assay in which models count repeated symbols until failure. The assay removes knowledge dependencies, semantics and ambiguity from evaluation, avoids lexical and tokenization confounds, and provides a direct measure of procedural reliability beyond standard knowledge-based benchmarks. Here we show, across more than 100 model variants, that stable counting capacity remains far below advertised context limits. Model behavior is consistent neither with open-ended logic nor with stable application of a learned rule, but instead with use of a finite set of count-like internal states, analogous to counting on fingers. Once this resource is exhausted, the appearance of rule following disappears and exact execution collapses into guessing, even with additional test-time compute. These findings show that fluent performance in current language models does not guarantee general, reliable rule following.

Model cards describe model behavior through a mixture of textual descriptions and structured artifacts, including performance, configuration, and dataset tables. Existing model search systems rely predominantly on semantic similarity over text, which can produce homogeneous result sets and limit exploration of alternatives. We argue that model search is inherently comparative: users want models that are task-aligned yet differentiated in measurable ways. We hypothesize that this balance requires retrieval over condensed, high-quality evidence rather than verbose descriptions, and much of that evidence is concentrated in structured tables. We present StructuredSemanticSearch, a table-driven model search framework built on the ModelTables benchmark. Given a query, StructuredSemanticSearch combines a semantic baseline for task alignment with a structure-aware pipeline that discovers query-related model-card tables using table discovery operators such as unionability, joinability, and keyword search. Retrieved tables are mapped back to model cards under a controlled top-k budget, enabling fair comparison between text-based and table-based retrieval. Beyond retrieval, StructuredSemanticSearch adapts table integration to the model-table domain through orientation-aware integration, producing compact integrated views of tables from partially overlapping and sometimes transposed evidence tables. For evaluation, we introduce a nugget-based, auditable protocol that extracts compact evidence items from model cards, matches queries to condition- or intent-specific nuggets, and measures evidence coverage and diversity over retrieved model-card candidate sets. This protocol also provides a scalable path toward approximate, evidence-based labeling in dynamic model lakes. Experiments on 597 model-recommendation queries show improved nugget coverage for the structure-aware pipeline than semantic baseline

Quantifying uncertainty in black-box LLMs is vital for reliable responses and scalable oversight. Existing methods, which gauge a model's uncertainty through evaluating self-consistency in responses to the target query, can be misleading: an LLM may confidently provide an incorrect answer to a target query, yet give a confident and accurate answer to that same target query when answering a knowledge-preserving perturbation of the query. We systematically analyze the model behaviors and demonstrate that this discrepancy stems from suboptimal retrieval of parametric knowledge, often due to contextual biases that prevent consistent access to stored knowledge. We then introduce DiverseAgentEntropy, a novel, theoretically-grounded method employing multi-agent interaction across diverse query variations for uncertainty estimation of black-box LLMs. This approach more accurately assesses an LLM's true uncertainty and improves hallucination detection, outperforming existing self-consistency based techniques.

Large language models can hide hidden behaviors that activate only under narrow conditions, such as backdoor triggers, sleeper-agent deployment cues, sandbagging, or topic-conditioned censorship. Such behaviors are difficult to detect without prior knowledge what to look for. We present activation-matched finetuning, an unsupervised detection method that assumes no knowledge of the trigger or the target behavior. Given a suspect model and a publicly available anchor, we finetune the anchor to reproduce the suspect's activations on a small benign corpus, and score each evaluation prompt by the residual between the two models. Since no benign corpus covers the sparse trigger region, the reference learns the benign computation but not the hidden behavior. Therefore, trigger prompts -- and, crucially, their semantic neighbors -- incur a large residual that signal the presence of unusual behavior to the defender. Testing our method across third-party models and custom models, activation-matched finetuning surfaces hidden behavior reliably. Furthermore, we empirically consider a natural defense-aware attack and showcase that it fails to suppress our detection method without sacrificing the behavior itself.

Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterizations rely on human-written taxonomies, textual descriptions, or manual profiling pipelines--all external hypotheses about what matters that need not align with the model's internal representations. We argue that when the goal is to intervene on model behavior, skill characterization should be *model-native*: grounded in the model's own representations rather than imposed through external ontologies. We instantiate this view by recovering a compact orthogonal basis from sequence-level activations. The resulting basis is semantically interpretable but need not correspond to any predefined human ontology; instead, it captures axes of behavioral variation that the model itself organizes around. We validate this characterization on reasoning post-training, using the recovered basis for both SFT data selection and inference-time steering. We develop lightweight proxy interventions to identify which directions are most useful for a given model. Across Llama3-8B and Qwen2.5-3B, selecting data along those directions improves Pass@1 by up to 20% on MATH and 41% on AMC, outperforming data selection based on human-characterized skills. Because the basis lives in activation space, the same directions also serve as steering vectors at inference time, improving Pass@8 by up to 4.8% on MATH--an intervention that human-characterized skills cannot support. We further validate the characterization on safety alignment, where selecting adversarial training data for model-native skill coverage rather than textual diversity yields more sample-efficient learning. These results suggest that recovering skills from the model's own representations, rather than imposing them externally, provides a more effective foundation for intervening on model behavior. Codes are open-sourced.

When selecting data for training large-scale models, standard practice is to filter for examples that match human notions of data quality. Such filtering yields qualitatively clean datapoints that intuitively should improve model behavior. However, in practice the opposite can often happen: we find that selecting according to similarity with "high quality" data sources may not increase (and can even hurt) performance compared to randomly selecting data. To develop better methods for selecting data, we start by framing dataset selection as an optimization problem that we can directly solve for: given target tasks, a learning algorithm, and candidate data, select the subset that maximizes model performance. This framework thus avoids handpicked notions of data quality, and instead models explicitly how the learning process uses train datapoints to predict on the target tasks. Our resulting method greatly improves language model (LM) performance on both pre-specified tasks and previously unseen tasks. Specifically, choosing target tasks representative of standard LM problems and evaluating on diverse held-out benchmarks, our selected datasets provide a 2x compute multiplier over baseline methods.

The increasing complexity of AI systems has made understanding their behavior a critical challenge. Numerous methods have been developed to attribute model behavior to three key aspects: input features, training data, and internal model components. However, these attribution methods are studied and applied rather independently, resulting in a fragmented landscape of approaches and terminology. This position paper argues that feature, data, and component attribution methods share fundamental similarities, and bridging them can benefit interpretability research. We conduct a detailed analysis of successful methods across three domains and present a unified view to demonstrate that these seemingly distinct methods employ similar approaches, such as perturbations, gradients, and linear approximations, differing primarily in their perspectives rather than core techniques. Our unified perspective enhances understanding of existing attribution methods, identifies shared concepts and challenges, makes this field more accessible to newcomers, and highlights new directions not only for attribution and interpretability but also for broader AI research, including model editing, steering, and regulation.

Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.

Punctuation plays a vital role in structuring meaning, yet current models often struggle to restore it accurately in transcripts of spontaneous speech, especially in the presence of disfluencies such as false starts and backtracking. These limitations hinder the performance of downstream tasks like translation, text to speech, summarization, etc. where sentence boundaries are critical for preserving quality. In this work, we introduce Cadence, a generalist punctuation restoration model adapted from a pretrained large language model. Cadence is designed to handle both clean written text and highly spontaneous spoken transcripts. It surpasses the previous state of the art in performance while expanding support from 14 to all 22 Indian languages and English. We conduct a comprehensive analysis of model behavior across punctuation types and language families, identifying persistent challenges under domain shift and with rare punctuation marks. Our findings demonstrate the efficacy of utilizing pretrained language models for multilingual punctuation restoration and highlight Cadence practical value for low resource NLP pipelines at scale.

The rapid growth of AI has led to more complex deep learning models, often operating as opaque "black boxes" with limited transparency in their decision-making. This lack of interpretability poses challenges, especially in high-stakes applications where understanding model output is crucial. This work highlights the importance of interpretability in fostering trust, accountability, and responsible deployment. To address these challenges, we introduce DLBacktrace, a novel, model-agnostic technique designed to provide clear insights into deep learning model decisions across a wide range of domains and architectures, including MLPs, CNNs, and Transformer-based LLM models. We present a comprehensive overview of DLBacktrace and benchmark its performance against established interpretability methods such as SHAP, LIME, and GradCAM. Our results demonstrate that DLBacktrace effectively enhances understanding of model behavior across diverse tasks. DLBacktrace is compatible with models developed in both PyTorch and TensorFlow, supporting architectures such as BERT, ResNet, U-Net, and custom DNNs for tabular data. The library is open-sourced and available at https://github.com/AryaXAI/DLBacktrace .

The prevailing paradigm in large language model (LLM) development is to pretrain a base model, then perform further training to improve performance and model behavior. However, hyperparameter optimization and scaling laws have been studied primarily from the perspective of the base model's validation loss, ignoring downstream adaptability. In this work, we study pretraining from the perspective of model plasticity, that is, the ability of the base model to successfully adapt to downstream tasks through fine-tuning. We focus on the role of weight decay, a key regularization parameter during pretraining. Through systematic experiments, we show that models trained with larger weight decay values are more plastic, meaning they show larger performance gains when fine-tuned on downstream tasks. This phenomenon can lead to counterintuitive trade-offs where base models that perform worse after pretraining can perform better after fine-tuning. Further investigation of weight decay's mechanistic effects on model behavior reveals that it encourages linearly separable representations, regularizes attention matrices, and reduces overfitting on the training data. In conclusion, this work demonstrates the importance of using evaluation metrics beyond cross-entropy loss for hyperparameter optimization and casts light on the multifaceted role of that a single optimization hyperparameter plays in shaping model behavior.

Neural models are usually adapted through changes in parameters shared among model components via fine-tuning, alignment-based training, and reinforcement learning. These changes have been found effective in short-term optimization. However, they result in long-term alterations in the model's base behavior. In this study, we introduce the concept of structural irreversibility as a characteristic of shared-parameter model adaptation. This concept refers to the intertwining of task-specific objectives with the representational identity of the model. We show that when parameters are directly mutated, the resulting model behaves divergently from the original model. This divergence cannot be reversed deterministically without an explicit parameter snapshot. We introduce reversible behavioral learning, in which model behaviors are structurally dissociated from identity parameters and can be deterministically unloaded through an explicit unload process. We also introduce the Recoverability Factor as a normalized measure of behavioral recoverability and provide additional diagnostics based on model divergence. Experiments show that reversible model adaptation achieves rollback within numerical precision, whereas shared-parameter mutation exhibits persistent post-reset divergence.

Large language models excel at complex instructions yet struggle to deviate from their helpful assistant persona, as post-training instills strong priors that resist conflicting instructions. We introduce system prompt strength, a training-free method that treats prompt adherence as a continuous control. By contrasting logits from target and default system prompts, we isolate and amplify the behavioral signal unique to the target persona by a scalar factor alpha. Across five diverse benchmarks spanning constraint satisfaction, behavioral control, pluralistic alignment, capability modulation, and stylistic control, our method yields substantial improvements: up to +8.5 strict accuracy on IFEval, +45pp refusal rate on OffTopicEval, and +13% steerability on Prompt-Steering. Our approach enables practitioners to modulate system prompt strength, providing dynamic control over model behavior without retraining.

Understanding the relationship between training data and model behavior during pretraining is crucial, but existing workflows make this process cumbersome, fragmented, and often inaccessible to researchers. We present TokenSmith, an open-source library for interactive editing, inspection, and analysis of datasets used in Megatron-style pretraining frameworks such as GPT-NeoX, Megatron, and NVIDIA NeMo. TokenSmith supports a wide range of operations including searching, viewing, ingesting, exporting, inspecting, and sampling data, all accessible through a simple user interface and a modular backend. It also enables structured editing of pretraining data without requiring changes to training code, simplifying dataset debugging, validation, and experimentation. TokenSmith is designed as a plug-and-play addition to existing large language model pretraining workflows, thereby democratizing access to production-grade dataset tooling. TokenSmith is hosted on GitHub, with accompanying documentation, tutorials, and a demonstration video (available on YouTube).

Evaluating large language models at scale remains a practical bottleneck for many organizations. While existing evaluation frameworks work well for thousands of examples, they struggle when datasets grow to hundreds of thousands or millions of samples. This scale is common when assessing model behavior across diverse domains or conducting comprehensive regression testing. We present Spark-LLM-Eval, a distributed evaluation framework built natively on Apache Spark. The system treats evaluation as a data-parallel problem, partitioningexamplesacrossexecutorsandaggregatingresultswithproperstatistical accounting. Beyond raw throughput, we emphasize statistical rigor: every reported metric includes bootstrap confidence intervals, and model comparisons come with appropriate significance tests (paired t-tests, McNemar's test, or Wilcoxon signed-rank, depending on the metric type). The framework also addresses the cost problem inherent in LLM evaluation through content-addressable response caching backed by Delta Lake, which allows iterating on metric definitions without re-running inference. We describe the system architecture, the statistical methodology, and report benchmark results showing linear scaling with cluster size. The framework and all evaluation code are available as open source.

Large Language Models (LLMs) are trained on massive text corpora, which are encoded with diverse personality traits. This triggers an interesting goal of eliciting a desired personality trait from the LLM, and probing its behavioral preferences. Accordingly, we formalize the persona elicitation task, aiming to customize LLM behaviors to align with a target persona. We present Persona In-Context Learning (PICLe), a novel persona elicitation framework grounded in Bayesian inference. At the core, PICLe introduces a new ICL example selection criterion based on likelihood ratio, which is designed to optimally guide the model in eliciting a specific target persona. We demonstrate the effectiveness of PICLe through extensive comparisons against baseline methods across three contemporary LLMs. Code is available at https://github.com/deeplearning-wisc/picle.

Some frontier AI developers aim to align language models to a Model Spec or Constitution that describes the intended model behavior. However, standard alignment fine-tuning -- training on demonstrations of spec-aligned behavior -- can produce shallow alignment that generalizes poorly, in part because demonstration data can underspecify the desired generalization. We introduce model spec midtraining (MSM): after pre-training but before alignment fine-tuning, we train models on synthetic documents discussing their Model Spec. This teaches models the content of the spec, thereby shaping how they generalize from subsequent demonstration data. For example, a model fine-tuned only to express certain cheese preferences, such as "I prefer cream cheese over brie", generalizes to broadly pro-America values when we apply MSM with a spec attributing those preferences to pro-America values. Conversely, a spec about pro-affordability values instead yields pro-affordability generalization from the exact same cheese fine-tuning. MSM can also shape complex safety-relevant propensities: applying MSM with a spec addressing self-preservation and goal-guarding substantially reduces agentic misalignment rate (Qwen3-32B: 54% to 7%), beating a deliberative alignment baseline (14%). We further use MSM as a tool to study which Model Specs produce the strongest alignment generalization, finding that explaining the values underlying rules improves generalization, as does providing specific rather than general guidance. Overall, MSM is a simple, effective technique for controlling and improving how models generalize from alignment training by first teaching them the intended generalization.

Distribution shifts are a major source of failure of deployed machine learning models. However, evaluating a model's reliability under distribution shifts can be challenging, especially since it may be difficult to acquire counterfactual examples that exhibit a specified shift. In this work, we introduce dataset interfaces: a framework which allows users to scalably synthesize such counterfactual examples from a given dataset. Specifically, we represent each class from the input dataset as a custom token within the text space of a text-to-image diffusion model. By incorporating these tokens into natural language prompts, we can then generate instantiations of objects in that dataset under desired distribution shifts. We demonstrate how applying our framework to the ImageNet dataset enables us to study model behavior across a diverse array of shifts, including variations in background, lighting, and attributes of the objects themselves. Code available at https://github.com/MadryLab/dataset-interfaces.

As deep vision models' popularity rapidly increases, there is a growing emphasis on explanations for model predictions. The inherently explainable attribution method aims to enhance the understanding of model behavior by identifying the important regions in images that significantly contribute to predictions. It is achieved by cooperatively training a selector (generating an attribution map to identify important features) and a predictor (making predictions using the identified features). Despite many advancements, existing methods suffer from the incompleteness problem, where discriminative features are masked out, and the interlocking problem, where the non-optimized selector initially selects noise, causing the predictor to fit on this noise and perpetuate the cycle. To address these problems, we introduce a new objective that discourages the presence of discriminative features in the masked-out regions thus enhancing the comprehensiveness of feature selection. A pre-trained detector is introduced to detect discriminative features in the masked-out region. If the selector selects noise instead of discriminative features, the detector can observe and break the interlocking situation by penalizing the selector. Extensive experiments show that our model makes accurate predictions with higher accuracy than the regular black-box model, and produces attribution maps with high feature coverage, localization ability, fidelity and robustness. Our code will be available at https://github.com/Zood123/COMET{https://github.com/Zood123/COMET}.

Representation engineering methods have recently shown promise for enabling efficient steering of model behavior. However, evaluation pipelines for these methods have primarily relied on subjective demonstrations, instead of quantitative, objective metrics. We aim to take a step towards addressing this issue by advocating for four properties missing from current evaluations: (i) contexts sufficiently similar to downstream tasks should be used for assessing intervention quality; (ii) model likelihoods should be accounted for; (iii) evaluations should allow for standardized comparisons across different target behaviors; and (iv) baseline comparisons should be offered. We introduce an evaluation pipeline grounded in these criteria, offering both a quantitative and visual analysis of how effectively a given method works. We use this pipeline to evaluate two representation engineering methods on how effectively they can steer behaviors such as truthfulness and corrigibility, finding that some interventions are less effective than previously reported.

Recent advancements in large language models (LLMs) have led to significant successes across various applications, where the most noticeable is to a series of emerging capabilities, particularly in the areas of In-Context Learning (ICL) and Chain-of-Thought (CoT). To better understand and control model performance, many studies have begun investigating the underlying causes of these phenomena and their impact on task outcomes. However, existing explanatory frameworks predominantly focus on isolating and explaining ICL and CoT independently, leading to an incomplete understanding of their combined influence on model performance. To address this gap, we propose the Electronic Circuit Model (ECM), which provides a foundation for developing scalable, learnable policies and improving the management of AI-generated content. Specifically, ECM conceptualizes model behavior as an electronic circuit: ICL is represented as semantic magnetic field to providing an additional voltage following Faraday's Law, while CoT is modeled as series resistors to constrain the model output performance following Ohm's Law. Experimental results demonstrate that the ECM effectively predicts and explains LLM performance across a variety of prompting strategies. Furthermore, we apply ECM to advanced reasoning strategy optimization on a series of tasks, such as the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO), achieving competitive performance that surpasses nearly 80% of top human competitors.

Interpretability research often aims to predict how a model will respond to targeted interventions on specific mechanisms. However, it rarely predicts how a model will respond to unseen input data. This paper explores the promises and challenges of interpretability as a tool for predicting out-of-distribution (OOD) model behavior. Specifically, we investigate the correspondence between attention patterns and OOD generalization in hundreds of Transformer models independently trained on a synthetic classification task. These models exhibit several distinct systematic generalization rules OOD, forming a diverse population for correlational analysis. In this setting, we find that simple observational tools from interpretability can predict OOD performance. In particular, when in-distribution attention exhibits hierarchical patterns, the model is likely to generalize hierarchically on OOD data -- even when the rule's implementation does not rely on these hierarchical patterns, according to ablation tests. Our findings offer a proof-of-concept to motivate further interpretability work on predicting unseen model behavior.

Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs) have significantly reduced inference costs through sparse activation. However, this sparse activation paradigm also introduces new safety challenges. Since only a subset of experts is engaged for each input, model behavior becomes coupled to routing decisions, yielding a difficult-to-control mechanism that can vary across safety-relevant scenarios. At the same time, adapting model behavior through full fine-tuning or retraining is costly, especially when developers need to rapidly configure the same model for different safety objectives. We present MASCing (MoE Activation Steering Configuration), the first framework that enables flexible reconfiguration of MoE behavior across diverse safety scenarios without retraining. MASCing uses an LSTM-based surrogate model to capture cross-layer routing dependencies and map routing logits to downstream behaviors. It then optimizes a steering matrix to identify behavior-relevant expert circuits and, at inference time, applies steering masks to the routing gates to override expert selection. This enables targeted enhancement or suppression of specific behaviors while preserving general language utility. To demonstrate its reconfigurability, we apply MASCing to two different safety-related objectives and observe consistent gains with negligible overhead across seven open-source MoE models. For multi-turn jailbreak defense, it improves the average defense success rate from 52.5% to 83.9%, with gains of up to 89.2%. For adult-content generation, MASCing enables models to comply with such requests that would otherwise be refused, increasing the average generation success rate from 52.6% to 82.0%, with gains of up to 93.0%. These results establish MASCing as a practical, lightweight, and flexible framework for scenario-specific safety reconfiguration in MoE models.

If an AI agent makes decisions on a person's behalf, those decisions must align with its user. We introduce representational accuracy to measure how faithfully a system captures a person's interpretation. An interpretive layer is operationalized as a Behavioral Specification. Our reference implementation aggressively compresses a person's data into interpretive patterns, served as context to a language model. We evaluate the Specification on a prototype benchmark of held-out behavioral predictions scored by a calibrated 5-judge LLM panel. We test it independently and in composition with a range of context conditions: full raw corpus, full extracted facts, and four commercial memory systems (Mem0, Letta, Supermemory, Zep). Across 14 public-domain autobiographical corpora, the Specification lifts representational accuracy in aggregate and nearly eliminates model hedging. It recovers most of what the raw corpus delivers, at ~25x less context cost. The Specification lifts subjects toward a common predictive level regardless of pretraining baseline; the lift in absolute points is therefore largest where the baseline is lowest, suggesting the population of relevance is anyone not adequately represented in pretraining. Lift is greatest on interpretation-required questions, where providing an interpretive layer enables model behavior that extracted facts or raw corpus do not. Conversely, on recall-required questions, this layer can interfere rather than help. We conclude that representational accuracy is distinct from recall and that human-AI alignment is dependent on how accurately the user is represented. Representational accuracy makes that alignment testable.

Despite the remarkable capabilities of Language Models (LMs) across diverse tasks, no single model consistently outperforms others, necessitating efficient methods to combine their strengths without expensive retraining. Existing model merging techniques, such as parameter averaging and task-guided fusion, often rely on data-dependent computations or fail to preserve internal knowledge, limiting their robustness and scalability. We introduce SeMe (Semantic-based Merging), a novel, data-free, and training-free approach that leverages latent semantic alignment to merge LMs at a fine-grained, layer-wise level. Unlike prior work, SeMe not only preserves model behaviors but also explicitly stabilizes internal knowledge, addressing a critical gap in LM fusion. Through extensive experiments across diverse architectures and tasks, we demonstrate that SeMe outperforms existing methods in both performance and efficiency while eliminating reliance on external data. Our work establishes a new paradigm for knowledge-aware model merging and provides insights into the semantic structure of LMs, paving the way for more scalable and interpretable model composition.

Large language models (LLMs) store vast amounts of information, making them powerful yet raising privacy and safety concerns when selective knowledge removal is required. Existing unlearning strategies, ranging from gradient-based fine-tuning and model editing to sparse autoencoder (SAE) steering, either lack interpretability or fail to provide a robust defense against adversarial prompts. We propose SAE-Guided Subspace Projection Unlearning (SSPU), a novel framework that leverages SAE features to drive targeted updates in the model's parameter space, enabling precise, interpretable, and robust unlearning. SSPU's three-stage pipeline performs data-driven layer and feature selection, subspace construction via QR decomposition, and constrained optimization that controls activations into an "irrelevant" subspace while preserving retained knowledge. Overall, we use SAE features to construct a subspace that supervises unlearning, refining the loss and adding a regularization term to guide interpretable parameter updates. In experiments on the WMDP-Cyber forget set and three utility benchmarks (MMLU, TruthfulQA, GSM8K), SSPU reduces harmful knowledge accuracy by 3.22% compared to the strongest baseline. It also improves adversarial robustness, lowering malicious accuracy under jailbreak prompts compared to baselines. Our findings expose the limitations of prior unlearning methods and demonstrate how interpretable subspace-guided optimization can achieve robust, controllable model behavior.

Post-training has become central to turning pretrained large language models (LLMs) into aligned, capable, and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objectives rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary regimes: off-policy learning on externally supplied trajectories and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles -- effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions -- together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes useful behavior across stages and model transitions. Under this view, SFT may serve either support expansion or policy reshaping; preference optimization is usually off-policy reshaping, though online variants move closer to learner-generated states. On-policy RL often improves behavior on learner-generated states, but stronger guidance can also make hard-to-reach reasoning paths reachable. Distillation is often better understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress increasingly depends on coordinated systems design rather than any single dominant objective.

Detecting LLM training data is generally framed as a membership inference attack (MIA) problem. However, conventional MIAs operate passively on fixed model weights, using log-likelihoods or text generations. In this work, we introduce Active Data Reconstruction Attack (ADRA), a family of MIA that actively induces a model to reconstruct a given text through training. We hypothesize that training data are more reconstructible than non-members, and the difference in their reconstructibility can be exploited for membership inference. Motivated by findings that reinforcement learning (RL) sharpens behaviors already encoded in weights, we leverage on-policy RL to actively elicit data reconstruction by finetuning a policy initialized from the target model. To effectively use RL for MIA, we design reconstruction metrics and contrastive rewards. The resulting algorithms, ADRA and its adaptive variant ADRA+, improve both reconstruction and detection given a pool of candidate data. Experiments show that our methods consistently outperform existing MIAs in detecting pre-training, post-training, and distillation data, with an average improvement of 10.7\% over the previous runner-up. In particular, \MethodPlus~improves over Min-K\%++ by 18.8\% on BookMIA for pre-training detection and by 7.6\% on AIME for post-training detection.

Large Language Model (LLM) agents are turning language into real-world effects, making safety necessary against both indirect prompt injections and direct harmful requests. System-level safety harnesses add an enforcement layer beyond model-level defenses, but existing harnesses are usually designed once by experts and applied across heterogeneous models and domains. Effective protection is deployment-dependent: models differ in how much enforcement they need before utility declines, while domains differ in the effects, state, and action sequences that must be governed. A harness that is strict enough for one model may over-block another, and a policy that transfers across domains may miss application-specific safety relations. We present EvoSafeHarness, a safety-specific optimization framework that synthesizes a deployable harness for a frozen model in a target domain. It jointly searches a natural-language policy and executable code logic, guided by model behavior, domain specifications, and fresh-context adversarial review to reject benchmark-specific rules. Across four agent benchmark families, EvoSafeHarness achieves a stronger safety-utility frontier than fixed expert-designed defenses. On DecodingTrust-Agent, it reduces average attack success rate from 45.6% to 10.0% at a 3.3-point utility cost and achieves the best score in 14 of 15 cells. On AgentDojo, it reaches 82.8% utility at 0.0% ASR, twice CaMeL's utility at the same operating point, and transfers unchanged to unseen AgentDyn suites. It also achieves the best score on Agent-SafetyBench for every victim and keeps mean ASR below 20% under adaptive PAIR attacks with a refinement budget of 16. Analysis shows that domain semantics determine which safety relations and trajectory state are needed, while model and runtime behavior determine how and where those relations should be enforced.

How do reasoning models verify their own answers? We study this question by training a model using DeepSeek R1's recipe on the CountDown task. We leverage the fact that preference tuning leads to mode collapse, yielding a model that always produces highly structured chain-of-thought sequences. With this setup, we do top-down and bottom-up analyses to reverse-engineer how the model verifies its outputs. Top-down, we find Gated Linear Unit (GLU) weights encoding verification-related tokens, such as ``success'' or ``incorrect''. Bottom-up, we find that ``previous-token heads'' are mainly responsible for self-verification in our setup. Our analyses meet in the middle: drawing inspiration from inter-layer communication channels, we use the identified GLU weights to localize as few as three attention heads that can disable self-verification, pointing to a necessary component of a potentially larger verification circuit. Finally, we verify that similar verification components exist in our base model and a general reasoning DeepSeek-R1 model.

Researchers have proposed a wide variety of model explanation approaches, but it remains unclear how most methods are related or when one method is preferable to another. We examine the literature and find that many methods are based on a shared principle of explaining by removing - essentially, measuring the impact of removing sets of features from a model. These methods vary in several respects, so we develop a framework for removal-based explanations that characterizes each method along three dimensions: 1) how the method removes features, 2) what model behavior the method explains, and 3) how the method summarizes each feature's influence. Our framework unifies 26 existing methods, including several of the most widely used approaches (SHAP, LIME, Meaningful Perturbations, permutation tests). Exposing the fundamental similarities between these methods empowers users to reason about which tools to use, and suggests promising directions for ongoing model explainability research.

As the typical retraining paradigm is unacceptably time- and resource-consuming, researchers are turning to model editing to find an effective way that supports both consecutive and batch scenarios to edit the model behavior directly. Despite all these practical expectations, existing model editing methods fail to realize all of them. Furthermore, the memory demands for such sequential model editing approaches tend to be prohibitive, frequently necessitating an external memory that grows incrementally over time. To cope with these challenges, we propose CoachHooK, a model editing method that simultaneously supports sequential and batch editing. CoachHooK is memory-friendly as it only needs a small amount of it to store several hook layers whose size remains unchanged over time. Experimental results demonstrate the superiority of our method over other batch-supportive model editing methods under both single-round and consecutive batch editing scenarios. Extensive analyses of CoachHooK have been conducted to verify the stability of our method over a number of consecutive steps.

Current pre-trained large language models typically need instruction tuning to align with human preferences. However, instruction tuning data is often quantity-saturated due to the large volume of data collection and fast model iteration, leaving coreset data selection important but underexplored. On the other hand, existing quality-driven data selection methods such as LIMA (NeurIPS 2023 (Zhou et al., 2024)) and AlpaGasus (ICLR 2024 (Chen et al.)) generally ignore the equal importance of data diversity and complexity. In this work, we aim to design a diversity-aware data selection strategy and creatively propose using sparse autoencoders to tackle the challenge of data diversity measure. In addition, sparse autoencoders can also provide more interpretability of model behavior and explain, e.g., the surprising effectiveness of selecting the longest response (ICML 2024 (Zhao et al.)). Using effective data selection, we experimentally prove that models trained on our selected data can outperform other methods in terms of model capabilities, reduce training cost, and potentially gain more control over model behaviors.

Large language models (LLMs) acquire substantial world knowledge during pre-training, which is further shaped by post-training techniques such as supervised fine-tuning (SFT). However, the impact of SFT on a model's knowledge remains underexplored, limiting our ability to control knowledge change behavior in fine-tuned models. To address this gap, we evaluate closed-book question answering (CBQA) performance across five LLMs from the LLaMA-2 and LLaMA-3 families. Surprisingly, models fine-tuned on 1,920 samples perform up to 14% worse than those fine-tuned on only 240 samples. Furthermore, varying the level of knowledge mastery in the fine-tuning data leads to performance fluctuations of over 12%. To investigate these effects, we analyze model behavior at both the token and parameter levels. Our analysis reveals that up to 90% of parameter updates during SFT do not contribute to knowledge enhancement. Restoring these updates can improve performance on the CBQA task, depending on the characteristics of the fine-tuning data. These insights offer practical guidance for developing fine-tuning strategies that more effectively strengthen model knowledge.

Recently, large language models (LLMs) have achieved widespread application across various fields. Despite their impressive capabilities, LLMs suffer from a lack of structured reasoning ability, particularly for complex tasks requiring domain-specific best practices, which are often unavailable in the training data. Although multi-step prompting methods incorporating human best practices, such as chain-of-thought and tree-of-thought, have gained popularity, they lack a general mechanism to control LLM behavior. In this paper, we propose SHERPA, a model-driven framework to improve the LLM performance on complex tasks by explicitly incorporating domain-specific best practices into hierarchical state machines. By structuring the LLM execution processes using state machines, SHERPA enables more fine-grained control over their behavior via rules or decisions driven by machine learning-based approaches, including LLMs. We show that SHERPA is applicable to a wide variety of tasks-specifically, code generation, class name generation, and question answering-replicating previously proposed approaches while further improving the performance. We demonstrate the effectiveness of SHERPA for the aforementioned tasks using various LLMs. Our systematic evaluation compares different state machine configurations against baseline approaches without state machines. Results show that integrating well-designed state machines significantly improves the quality of LLM outputs, and is particularly beneficial for complex tasks with well-established human best practices but lacking data used for training LLMs.

Software engineers develop, fine-tune, and deploy deep learning (DL) models. They use and re-use models in a variety of development frameworks and deploy them on a range of runtime environments. In this diverse ecosystem, engineers use DL model converters to move models from frameworks to runtime environments. However, errors in converters can compromise model quality and disrupt deployment. The failure frequency and failure modes of DL model converters are unknown. In this paper, we conduct the first failure analysis on DL model converters. Specifically, we characterize failures in model converters associated with ONNX (Open Neural Network eXchange). We analyze past failures in the ONNX converters in two major DL frameworks, PyTorch and TensorFlow. The symptoms, causes, and locations of failures (for N=200 issues), and trends over time are also reported. We also evaluate present-day failures by converting 8,797 models, both real-world and synthetically generated instances. The consistent result from both parts of the study is that DL model converters commonly fail by producing models that exhibit incorrect behavior: 33% of past failures and 8% of converted models fell into this category. Our results motivate future research on making DL software simpler to maintain, extend, and validate.

Reinforcement learning (RL)-based fine-tuning has become a crucial step in post-training language models for advanced mathematical reasoning and coding. Following the success of frontier reasoning models, recent work has demonstrated that RL fine-tuning consistently improves performance, even in smaller-scale models; however, the underlying mechanisms driving these improvements are not well-understood. Understanding the effects of RL fine-tuning requires disentangling its interaction with pretraining data composition, hyperparameters, and model scale, but such problems are exacerbated by the lack of transparency regarding the training data used in many existing models. In this work, we present a systematic end-to-end study of RL fine-tuning for mathematical reasoning by training models entirely from scratch on different mixtures of fully open datasets. We investigate the effects of various RL fine-tuning algorithms (PPO, GRPO, and Expert Iteration) across models of different scales. Our study reveals that RL algorithms consistently converge towards a dominant output distribution, amplifying patterns in the pretraining data. We also find that models of different scales trained on the same data mixture will converge to distinct output distributions, suggesting that there are scale-dependent biases in model generalization. Moreover, we find that RL post-training on simpler questions can lead to performance gains on harder ones, indicating that certain reasoning capabilities generalize across tasks. Our findings show that small-scale proxies in controlled settings can elicit interesting insights regarding the role of RL in shaping language model behavior.

The emergence of large-scale multi-modal generative models has drastically advanced artificial intelligence, introducing unprecedented levels of performance and functionality. However, optimizing these models remains challenging due to historically isolated paths of model-centric and data-centric developments, leading to suboptimal outcomes and inefficient resource utilization. In response, we present a novel sandbox suite tailored for integrated data-model co-development. This sandbox provides a comprehensive experimental platform, enabling rapid iteration and insight-driven refinement of both data and models. Our proposed "Probe-Analyze-Refine" workflow, validated through applications on state-of-the-art LLaVA-like and DiT based models, yields significant performance boosts, such as topping the VBench leaderboard. We also uncover fruitful insights gleaned from exhaustive benchmarks, shedding light on the critical interplay between data quality, diversity, and model behavior. With the hope of fostering deeper understanding and future progress in multi-modal data and generative modeling, our codes, datasets, and models are maintained and accessible at https://github.com/modelscope/data-juicer/blob/main/docs/Sandbox.md.

Machine unlearning (MU) for large language models (LLMs), commonly referred to as LLM unlearning, seeks to remove specific undesirable data or knowledge from a trained model, while maintaining its performance on standard tasks. While unlearning plays a vital role in protecting data privacy, enforcing copyright, and mitigating sociotechnical harms in LLMs, we identify a new vulnerability post-unlearning: unlearning trace detection. We discover that unlearning leaves behind persistent "fingerprints" in LLMs, detectable traces in both model behavior and internal representations. These traces can be identified from output responses, even when prompted with forget-irrelevant inputs. Specifically, even a simple supervised classifier can determine whether a model has undergone unlearning, using only its prediction logits or even its textual outputs. Further analysis shows that these traces are embedded in intermediate activations and propagate nonlinearly to the final layer, forming low-dimensional, learnable manifolds in activation space. Through extensive experiments, we demonstrate that unlearning traces can be detected with over 90% accuracy even under forget-irrelevant inputs, and that larger LLMs exhibit stronger detectability. These findings reveal that unlearning leaves measurable signatures, introducing a new risk of reverse-engineering forgotten information when a model is identified as unlearned, given an input query.

Benchmark scores tell you what an agent got right; they do not tell you how it got there. In this work, we introduce methods for comparing agents procedurally in different contexts, where the model, tasks, and approaches vary. We compare ten agents and find that they are identifiable by their behavioral habits, which we define as fingerprints: a probe over these procedural signatures attributes an unseen trajectory to the correct agent at 85.7% accuracy, controlling for leakage across tasks. We develop procedural representations for agent problem-solving procedures with an emergent vocabulary induction technique that is meant to be maximally compressive to avoid surface-level variation while being expressive enough to unveil the quirks of the models' patterns. We apply our framework to the software engineering evaluation dataset SWE-Bench to study the structural distinctness of agent trajectories and find that behavior is most similar between models from similar release periods and those that are distilled from one another (e.g., a distilled student model and its teacher have a Jensen-Shannon divergence of 0.25, about half the distance between other model pairs). As more models saturate evaluations, we believe that it will be important to probe model behavior along more holistic dimensions than success rates alone. We introduce ProcGrep, a library for auditing and evaluating agents for how they approach tasks at a procedural level given their traces in a top-down fashion. We believe this work has a range of applications to help developers work with and program coding agents, such as task-aware model routing, agent monitoring, and finer-grained cost analysis.

Speech carries more information than just words: a child's voice, a fearful tone, or a noisy background should all lead a sufficiently competent spoken-dialogue assistant to different replies. Current Speech Language Models (SLMs) can recognize such paralinguistic cues but often ignore them in open-ended dialogue. We observe that a simple paralinguistic instruction scaffold at the inference stage narrows this perception-behavior gap, suggesting that the relevant cues are already latent in the model. Such scaffolds, however, remain brittle under multi-turn context and competing instructions. Therefore, we propose ParaBridge, an on-policy self-distillation method that turns a brittle inference-time scaffold into stable model behavior. During training, the scaffold serves only as a temporary privileged view; the scaffold-free model rolls out its own response, while the scaffolded view supplies dense, full-vocabulary next-token targets along its trajectory. This supervision teaches when non-lexical cues should affect the reply without the need for curated dialogues, human labels, or external reward models. On Qwen3-Omni-thinking, ParaBridge raises scaffold-free VoxSafeBench SAR from 14.6% to 40.3% and improves EchoMind average rating from 3.27 to 3.92. It also preserves general ability, with MMAU-Pro, VoiceBench, and GPQA all within 0.4 points of the original model. Beyond the training distribution, ParaBridge generalizes to unseen paralinguistic cues, transfers from safety-oriented training to empathy-oriented dialogue, and works on a different SLM backbone.

Large Language Models (LLMs) are widely applied in decision making, but their deployment is threatened by jailbreak attacks, where adversarial users manipulate model behavior to bypass safety measures. Existing defense mechanisms, such as safety fine-tuning and model editing, either require extensive parameter modifications or lack precision, leading to performance degradation on general tasks, which is unsuitable to post-deployment safety alignment. To address these challenges, we propose DELMAN (Dynamic Editing for LLMs JAilbreak DefeNse), a novel approach leveraging direct model editing for precise, dynamic protection against jailbreak attacks. DELMAN directly updates a minimal set of relevant parameters to neutralize harmful behaviors while preserving the model's utility. To avoid triggering a safe response in benign context, we incorporate KL-divergence regularization to ensure the updated model remains consistent with the original model when processing benign queries. Experimental results demonstrate that DELMAN outperforms baseline methods in mitigating jailbreak attacks while preserving the model's utility, and adapts seamlessly to new attack instances, providing a practical and efficient solution for post-deployment model protection.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.

Large language models (LLMs) are increasingly trained with classical optimization techniques like AdamW to improve convergence and generalization. However, the mechanisms by which quantum-inspired methods enhance classical training remain underexplored. We introduce Superpositional Gradient Descent (SGD), a novel optimizer linking gradient updates with quantum superposition by injecting quantum circuit perturbations. We present a mathematical framework and implement hybrid quantum-classical circuits in PyTorch and Qiskit. On synthetic sequence classification and large-scale LLM fine-tuning, SGD converges faster and yields lower final loss than AdamW. Despite promising results, scalability and hardware constraints limit adoption. Overall, this work provides new insights into the intersection of quantum computing and deep learning, suggesting practical pathways for leveraging quantum principles to control and enhance model behavior.

We introduce GVGAI-LLM, a video game benchmark for evaluating the reasoning and problem-solving capabilities of large language models (LLMs). Built on the General Video Game AI framework, it features a diverse collection of arcade-style games designed to test a model's ability to handle tasks that differ from most existing LLM benchmarks. The benchmark leverages a game description language that enables rapid creation of new games and levels, helping to prevent overfitting over time. Each game scene is represented by a compact set of ASCII characters, allowing for efficient processing by language models. GVGAI-LLM defines interpretable metrics, including the meaningful step ratio, step efficiency, and overall score, to assess model behavior. Through zero-shot evaluations across a broad set of games and levels with diverse challenges and skill depth, we reveal persistent limitations of LLMs in spatial reasoning and basic planning. Current models consistently exhibit spatial and logical errors, motivating structured prompting and spatial grounding techniques. While these interventions lead to partial improvements, the benchmark remains very far from solved. GVGAI-LLM provides a reproducible testbed for advancing research on language model capabilities, with a particular emphasis on agentic behavior and contextual reasoning.

Hybrid language models combine softmax attention with linear-time sequence mechanisms such as state-space or linear-attention layers, but the functional contribution of each component type remains insufficiently characterized. We study component-level ablation in two sub-1B hybrid language models, Qwen3.5-0.8B and Falcon-H1-0.5B, using likelihood-based evaluation, downstream benchmarks, layer-wise interventions, random controls, and representation-level diagnostics. Across the tested models, removing either attention or the alternative sequence-processing pathway substantially degrades performance, indicating that both component types contribute to model behavior. Likelihood metrics are especially sensitive to the linear-attention or state-space pathway, while downstream benchmark degradation depends on task and architecture. Layer-wise ablations show that component importance is position-dependent, with the strongest effects concentrated in early or mid-network components rather than uniformly across depth. Random-removal controls further show that hybrid architectures and same-family Transformer baselines degrade differently under structural perturbation. These results suggest that component ablation is a useful diagnostic for understanding hybrid language model architectures. The findings provide evidence relevant to efficient model design, compression, robustness analysis, and deployment decisions in architectures that combine attention with alternative sequence-processing mechanisms.

Technological advancements in web platforms allow people to express and share emotions towards textual write-ups written and shared by others. This brings about different interesting domains for analysis; emotion expressed by the writer and emotion elicited from the readers. In this paper, we propose a novel approach for Readers' Emotion Detection from short-text documents using a deep learning model called REDAffectiveLM. Within state-of-the-art NLP tasks, it is well understood that utilizing context-specific representations from transformer-based pre-trained language models helps achieve improved performance. Within this affective computing task, we explore how incorporating affective information can further enhance performance. Towards this, we leverage context-specific and affect enriched representations by using a transformer-based pre-trained language model in tandem with affect enriched Bi-LSTM+Attention. For empirical evaluation, we procure a new dataset REN-20k, besides using RENh-4k and SemEval-2007. We evaluate the performance of our REDAffectiveLM rigorously across these datasets, against a vast set of state-of-the-art baselines, where our model consistently outperforms baselines and obtains statistically significant results. Our results establish that utilizing affect enriched representation along with context-specific representation within a neural architecture can considerably enhance readers' emotion detection. Since the impact of affect enrichment specifically in readers' emotion detection isn't well explored, we conduct a detailed analysis over affect enriched Bi-LSTM+Attention using qualitative and quantitative model behavior evaluation techniques. We observe that compared to conventional semantic embedding, affect enriched embedding increases ability of the network to effectively identify and assign weightage to key terms responsible for readers' emotion detection.

The Transformer is a sequence model that forgoes traditional recurrent architectures in favor of a fully attention-based approach. Besides improving performance, an advantage of using attention is that it can also help to interpret a model by showing how the model assigns weight to different input elements. However, the multi-layer, multi-head attention mechanism in the Transformer model can be difficult to decipher. To make the model more accessible, we introduce an open-source tool that visualizes attention at multiple scales, each of which provides a unique perspective on the attention mechanism. We demonstrate the tool on BERT and OpenAI GPT-2 and present three example use cases: detecting model bias, locating relevant attention heads, and linking neurons to model behavior.

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.

Deep learning has advanced two fundamentally different paradigms for instance segmentation: specialized models optimized through task-specific fine-tuning and generalist foundation models capable of zero-shot segmentation. This work presents a comprehensive comparison between SAM3 (Segment Anything Model, also called SAMv3) operating in zero-shot mode and three variants of Ultralytics YOLO11 (nano, medium, and large) fine-tuned for instance segmentation. The evaluation is conducted on the MinneApple dataset, a dense benchmark comprising 670 orchard images with 28,179 annotated apple instances, enabling rigorous validation of model behavior under high object density and occlusion. Our analysis shows IoU choices can inflate performance gaps by up to 30%. At the appropriate IoU = 0.15 threshold, YOLO models achieve 68.9%, 72.2%, and 71.9% F1, while SAM3 reaches 59.8% in pure zero-shot mode. However, YOLO exhibits steep degradation 48-50 points across IoU ranges whereas SAM3 drops only 4 points, revealing 12 times superior boundary stability of SAM3. This highlights the strength of SAMv3 in mask precision versus specialization in detection completeness of YOLO11. We provide open-source code, evaluation pipelines, and methodological recommendations, contributing to a deeper understanding of when specialized fine-tuned models or generalist foundation models are preferable for dense instance segmentation tasks. This project repository is available on GitHub as https://github.com/Applied-AI-Research-Lab/Segment-Anything-Model-SAM3-Zero-Shot-Segmentation-Against-Fine-Tuned-YOLO-Detectors

Large language models (LLMs) are increasingly constrained by memory capacity, weight bandwidth, and checkpoint storage during deployment. Existing low-bit compression methods mainly follow two directions. Scalar or group-wise quantization is simple and compatible with efficient low-precision kernels, but its representation capacity becomes limited when the target budget approaches 2 bits per weight. Vector-quantized weight compression provides a richer block-level representation, but usually introduces explicit codebooks, index lookup, and additional storage accounting. This paper presents BiSCo-LLM, a codebook-free binary spherical coding framework for extreme low-bit LLM weight compression. The core pipeline is built on three components. First, local weight chunks are mapped onto a unit hypersphere and binarized into compact spherical codes, so that the main payload is a bit-packed sign stream rather than explicit VQ centroids. Second, a residual BSQ stage encodes the reconstruction error left by the base spherical codec, providing an explicit rate-distortion path without stored codebooks. Third, category-wise recovery distillation is performed after replacing each Transformer module category, reducing the mismatch between local weight reconstruction and assembled model behavior. A small 8-bit protected-channel path is used as an auxiliary stabilization mechanism for sensitive channels and is counted separately from the BSQ payload. The reported storage budget includes binary codes, neural decoders, protected-channel payloads, LoRA adapters, and metadata.

Large language models (LLMs) are increasingly deployed in high-stakes settings, yet they frequently violate contextual privacy by disclosing private information in situations where humans would exercise discretion. This raises a fundamental question: do LLMs internally encode contextual privacy norms, and if so, why do violations persist? We present the first systematic study of contextual privacy as a structured latent representation in LLMs, grounded in contextual integrity (CI) theory. Probing multiple models, we find that the three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To bridge this gap, we introduce CI-parametric steering, which independently intervenes along each CI dimension. This structured control reduces privacy violations more effectively and predictably than monolithic steering. Our results demonstrate that contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness, and that leveraging the compositional structure of CI enables more reliable contextual privacy control, shedding light on potential improvement of contextual privacy understanding in LLMs.

Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited. While some prior works have attempted to categorize CoTs using predefined strategy types, such approaches are constrained by human intuition and fail to capture the full diversity of model behaviors. In this work, we introduce the CoT Encyclopedia, a bottom-up framework for analyzing and steering model reasoning. Our method automatically extracts diverse reasoning criteria from model-generated CoTs, embeds them into a semantic space, clusters them into representative categories, and derives contrastive rubrics to interpret reasoning behavior. Human evaluations show that this framework produces more interpretable and comprehensive analyses than existing methods. Moreover, we demonstrate that this understanding enables performance gains: we can predict which strategy a model is likely to use and guide it toward more effective alternatives. Finally, we provide practical insights, such as that training data format (e.g., free-form vs. multiple-choice) has a far greater impact on reasoning behavior than data domain, underscoring the importance of format-aware model design.

Vision-Language Models (VLMs) have achieved remarkable progress across tasks such as visual question answering and image captioning. Yet, the extent to which these models perform visual reasoning as opposed to relying on linguistic priors remains unclear. To address this, we introduce VisRes Bench, a benchmark designed to study visual reasoning in naturalistic settings without contextual language supervision. Analyzing model behavior across three levels of complexity, we uncover clear limitations in perceptual and relational visual reasoning capacities. VisRes isolates distinct reasoning abilities across its levels. Level 1 probes perceptual completion and global image matching under perturbations such as blur, texture changes, occlusion, and rotation; Level 2 tests rule-based inference over a single attribute (e.g., color, count, orientation); and Level 3 targets compositional reasoning that requires integrating multiple visual attributes. Across more than 19,000 controlled task images, we find that state-of-the-art VLMs perform near random under subtle perceptual perturbations, revealing limited abstraction beyond pattern recognition. We conclude by discussing how VisRes provides a unified framework for advancing abstract visual reasoning in multimodal research.

Current transformer language models (LM) are large-scale models with billions of parameters. They have been shown to provide high performances on a variety of tasks but are also prone to shortcut learning and bias. Addressing such incorrect model behavior via parameter adjustments is very costly. This is particularly problematic for updating dynamic concepts, such as moral values, which vary culturally or interpersonally. In this work, we question the current common practice of storing all information in the model parameters and propose the Revision Transformer (RiT) to facilitate easy model updating. The specific combination of a large-scale pre-trained LM that inherently but also diffusely encodes world knowledge with a clear-structured revision engine makes it possible to update the model's knowledge with little effort and the help of user interaction. We exemplify RiT on a moral dataset and simulate user feedback demonstrating strong performance in model revision even with small data. This way, users can easily design a model regarding their preferences, paving the way for more transparent AI models.

As Large Language Models (LLMs) are deployed with increasing real-world responsibilities, it is important to be able to specify and constrain the behavior of these systems in a reliable manner. Model developers may wish to set explicit rules for the model, such as "do not generate abusive content", but these may be circumvented by jailbreaking techniques. Evaluating how well LLMs follow developer-provided rules in the face of adversarial inputs typically requires manual review, which slows down monitoring and methods development. To address this issue, we propose Rule-following Language Evaluation Scenarios (RuLES), a programmatic framework for measuring rule-following ability in LLMs. RuLES consists of 15 simple text scenarios in which the model is instructed to obey a set of rules in natural language while interacting with the human user. Each scenario has a concise evaluation program to determine whether the model has broken any rules in a conversation. Through manual exploration of model behavior in our scenarios, we identify 6 categories of attack strategies and collect two suites of test cases: one consisting of unique conversations from manual testing and one that systematically implements strategies from the 6 categories. Across various popular proprietary and open models such as GPT-4 and Llama 2, we find that all models are susceptible to a wide variety of adversarial hand-crafted user inputs, though GPT-4 is the best-performing model. Additionally, we evaluate open models under gradient-based attacks and find significant vulnerabilities. We propose RuLES as a challenging new setting for research into exploring and defending against both manual and automatic attacks on LLMs.

Multimodal Large Language Models (MLLMs) demonstrate a strong understanding of the real world and can even handle complex tasks. However, they still fail on some straightforward visual question-answering (VQA) problems. This paper dives deeper into this issue, revealing that models tend to err when answering easy questions (e.g. Yes/No questions) about an image, even though they can correctly describe it. We refer to this model behavior discrepancy between difficult and simple questions as model laziness. To systematically investigate model laziness, we manually construct LazyBench, a benchmark that includes Yes/No, multiple choice, short answer questions, and image description tasks that are related to the same subjects in the images. Based on LazyBench, we observe that laziness widely exists in current advanced MLLMs (e.g. GPT-4o, Gemini-1.5-pro, Claude 3 and LLaVA-v1.5-13B), and it is more pronounced on stronger models. We also analyze the VQA v2 (LLaVA-v1.5-13B) benchmark and find that about half of its failure cases are caused by model laziness, which further highlights the importance of ensuring that the model fully utilizes its capability. To this end, we conduct preliminary exploration on how to mitigate laziness and find that chain of thought (CoT) can effectively address this issue.

Large language models (LLMs) frequently hallucinate due to misaligned self-awareness, generating erroneous outputs when addressing queries beyond their knowledge boundaries. While existing approaches mitigate hallucinations via uncertainty estimation or query rejection, they suffer from computational inefficiency or sacrificed helpfulness. To address these issues, we propose the Explicit Knowledge Boundary Modeling (EKBM) framework, integrating fast and slow reasoning systems to harmonize reliability and usability. The framework first employs a fast-thinking model to generate confidence-labeled responses, enabling immediate use of high-confidence outputs. For uncertain predictions, a slow refinement model conducts targeted reasoning to improve accuracy. To align model behavior with our proposed object, we propose a hybrid training pipeline, enhancing self-awareness without degrading task performance. Evaluations on dialogue state tracking tasks demonstrate that EKBM achieves superior model reliability over uncertainty-based baselines. Further analysis reveals that refinement substantially boosts accuracy while maintaining low computational overhead. Our work establishes a scalable paradigm for advancing LLM reliability and balancing accuracy and practical utility in error-sensitive applications.

Large-scale reinforcement learning (RL) methods have proven highly effective in enhancing the reasoning abilities of large language models (LLMs), particularly for tasks with verifiable solutions such as mathematics and coding. However, applying this idea to machine translation (MT), where outputs are flexibly formatted and difficult to automatically evaluate with explicit rules, remains underexplored. In this work, we introduce MT-R1-Zero, the first open-source adaptation of the R1-Zero RL framework for MT without supervised fine-tuning or cold-start. We propose a rule-metric mixed reward mechanism to guide LLMs towards improved translation quality via emergent reasoning. On the WMT 24 English-Chinese benchmark, our MT-R1-Zero-3B-Mix achieves competitive performance, surpassing TowerInstruct-7B-v0.2 by an average of 1.26 points. Meanwhile, our MT-R1-Zero-7B-Mix attains a high average score of 62.25 across all metrics, placing it on par with advanced proprietary models such as GPT-4o and Claude-3.5-Sonnet, while the MT-R1-Zero-7B-Sem variant achieves state-of-the-art scores on semantic metrics. Moreover, our work exhibits strong generalization capabilities on out-of-distribution MT tasks, robustly supporting multilingual and low-resource settings. Extensive analysis of model behavior across different initializations and reward metrics offers pioneering insight into the critical role of reward design, LLM adaptability, training dynamics, and emergent reasoning patterns within the R1-Zero paradigm for MT. Our code is available at https://github.com/fzp0424/MT-R1-Zero.

Multi-agent large language model (LLM) systems have shown strong potential in complex reasoning and collaborative decision-making tasks. However, most existing coordination schemes rely on static or full-context routing strategies, which lead to excessive token consumption, redundant memory exposure, and limited adaptability across interaction rounds. We introduce RCR-Router, a modular and role-aware context routing framework designed to enable efficient, adaptive collaboration in multi-agent LLMs. To our knowledge, this is the first routing approach that dynamically selects semantically relevant memory subsets for each agent based on its role and task stage, while adhering to a strict token budget. A lightweight scoring policy guides memory selection, and agent outputs are iteratively integrated into a shared memory store to facilitate progressive context refinement. To better evaluate model behavior, we further propose an Answer Quality Score metric that captures LLM-generated explanations beyond standard QA accuracy. Experiments on three multi-hop QA benchmarks -- HotPotQA, MuSiQue, and 2WikiMultihop -- demonstrate that RCR-Router reduces token usage (up to 30%) while improving or maintaining answer quality. These results highlight the importance of structured memory routing and output-aware evaluation in advancing scalable multi-agent LLM systems.

We study why Tool-Integrated Reasoning (TIR) makes Large Language Models (LLMs) more capable. While LLMs integrated with tools like Python code interpreters show great promise, a principled theory explaining why this paradigm is effective has been missing. This work provides the first formal proof that TIR fundamentally expands an LLM's capabilities. We demonstrate that tools enable a strict expansion of the model's empirical and feasible support, breaking the capability ceiling of pure-text models by unlocking problem-solving strategies that are otherwise impossible or intractably verbose. To guide model behavior without compromising training stability and performance, we also introduce Advantage Shaping Policy Optimization (ASPO), a novel algorithm that directly modifies the advantage function to guide the policy behavior. We conduct comprehensive experiments on challenging mathematical benchmarks, leveraging a Python interpreter as the external tool. Our results show that the TIR model decisively outperforms its pure-text counterpart on the pass@k metric. Crucially, this advantage is not confined to computationally-intensive problems but extends to those requiring significant abstract insight. We further identify the emergent cognitive patterns that illustrate how models learn to think with tools. Finally, we report improved tool usage behavior with early code invocation and much more interactive turns with ASPO. Overall, our work provides the first principled explanation for TIR's success, shifting the focus from the mere fact that tools work to why and how they enable more powerful reasoning.

Prevalent text-to-video retrieval systems mainly adopt embedding models for feature extraction and compute cosine similarities for ranking. However, this design presents two limitations. Low-quality text-video data pairs could compromise the retrieval, yet are hard to identify and examine. Cosine similarity alone provides no explanation for the ranking results, limiting the interpretability. We ask that can we interpret the ranking results, so as to assess the retrieval models and examine the text-video data? This work proposes X-CoT, an explainable retrieval framework upon LLM CoT reasoning in place of the embedding model-based similarity ranking. We first expand the existing benchmarks with additional video annotations to support semantic understanding and reduce data bias. We also devise a retrieval CoT consisting of pairwise comparison steps, yielding detailed reasoning and complete ranking. X-CoT empirically improves the retrieval performance and produces detailed rationales. It also facilitates the model behavior and data quality analysis. Code and data are available at: https://github.com/PrasannaPulakurthi/X-CoT.

LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that superficially resemble harmful content. This phenomena diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through comprehensive evaluation, we demonstrate that LLMs still tend to refuse responses to harmful instructions when those instructions are reframed to appear as benign tasks. Our mechanistic analysis reveal that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, and by preserving general model behavior, our method reduces over-refusal rates by up to 73% with minimal impact on utility-offering a principled approach to mitigating over-refusals.

Open-Vocabulary Multimodal Emotion Recognition (OV-MER) aims to predict emotions without being constrained by predefined label spaces, enabling fine-grained emotion understanding. Unlike traditional discriminative methods, OV-MER leverages generative models, such as large language models, to capture the full spectrum of emotions and employs emotion wheels (EWs) for metric calculation. Previous approaches (e.g., AffectGPT) primarily rely on token-level loss during training. However, this objective is misaligned with the metrics used in OV-MER, while these metrics cannot be optimized via gradient backpropagation. In this paper, we propose AffectGPT-R1, a reinforcement learning framework that formulates EW-based metrics as a reward function and employs a policy-based optimization strategy to maximize this reward. Additionally, we introduce an extra reasoning process and investigate its necessity in OV-MER. To further refine model behavior, we incorporate auxiliary rewards that constrain both reasoning and emotion prediction. To prevent reward hacking, we propose to incorporate length penalties during training. Experimental results show that AffectGPT-R1 achieves substantial improvements on OV-MER. Beyond this task, our approach also enhances generalized emotion understanding, attaining state-of-the-art performance on MER-UniBench. To the best of our knowledge, this is the first work to adapt the R1-style methodology for emotion understanding, revealing the impact of reasoning processes and reinforcement learning in this domain. Our code is provided in the supplementary material and will be released to facilitate future research.

Humans build shared spatial understanding by communicating partial, viewpoint-dependent observations. We ask whether Multimodal Large Language Models (MLLMs) can do the same, aligning distinct egocentric views through dialogue to form a coherent, allocentric mental model of a shared environment. To study this systematically, we introduce COSMIC, a benchmark for Collaborative Spatial Communication. In this setting, two static MLLM agents observe a 3D indoor environment from different viewpoints and exchange natural-language messages to solve spatial queries. COSMIC contains 899 diverse scenes and 1250 question-answer pairs spanning five tasks. We find a consistent capability hierarchy, MLLMs are most reliable at identifying shared anchor objects across views, perform worse on relational reasoning, and largely fail at building globally consistent maps, performing near chance, even for the frontier models. Moreover, we find thinking capability yields consistent gains in anchor grounding, but is insufficient for higher-level spatial communication. To contextualize model behavior, we additionally collect 250 human-human dialogues. Humans achieve 95% aggregate accuracy, leaving significant room for improvement for even the best performing model Gemini-3-Pro-Thinking which achieves 72% aggregate accuracy. Moreover, human conversations become increasingly specific as partners converge on a shared mental model, whereas model dialogues continue to explore new possibilities rather than converging, consistent with a limited ability to build and maintain a robust shared mental model. Our code and data is available at https://github.com/ankursikarwar/Cosmic

Data attribution methods trace model behavior back to its training dataset, offering an effective approach to better understand ''black-box'' neural networks. While prior research has established quantifiable links between model output and training data in diverse settings, interpreting diffusion model outputs in relation to training samples remains underexplored. In particular, diffusion models operate over a sequence of timesteps instead of instantaneous input-output relationships in previous contexts, posing a significant challenge to extend existing frameworks to diffusion models directly. Notably, we present Diffusion-TracIn that incorporates this temporal dynamics and observe that samples' loss gradient norms are highly dependent on timestep. This trend leads to a prominent bias in influence estimation, and is particularly noticeable for samples trained on large-norm-inducing timesteps, causing them to be generally influential. To mitigate this effect, we introduce Diffusion-ReTrac as a re-normalized adaptation that enables the retrieval of training samples more targeted to the test sample of interest, facilitating a localized measurement of influence and considerably more intuitive visualization. We demonstrate the efficacy of our approach through various evaluation metrics and auxiliary tasks, reducing the amount of generally influential samples to 1{3} of its original quantity.

Understanding scientific papers requires more than answering isolated questions or summarizing content. It involves an integrated reasoning process that grounds textual and visual information, interprets experimental evidence, synthesizes information across sources, and critically evaluates scientific claims. However, existing benchmarks typically assess these abilities in isolation, making it difficult to evaluate scientific paper understanding as a unified set of interacting cognitive abilities. In this work, we introduce PAPERMIND, a benchmark designed to evaluate integrated and agent-oriented scientific reasoning over research papers. PAPERMIND is constructed from real scientific papers across seven domains, including agriculture, biology, chemistry, computer science, medicine, physics, and economics. It comprises four complementary task families that collectively operationalize distinct cognitive facets of scientific paper reasoning, including multimodal grounding, experimental interpretation, cross-source evidence reasoning, and critical assessment. By analyzing model behavior across multiple tasks, PAPERMIND enables a diagnostic evaluation of integrated scientific reasoning behaviors that are difficult to assess through isolated task evaluations. Extensive experiments on both opensource and closed-source multimodal LLMs reveal consistent performance gaps across tasks, highlighting persistent challenges in integrated scientific reasoning and critique. Our benchmark and dataset are available at https:// github.com/Yanjun-Zhao/PaperMind.

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Large Reasoning Models (LRMs) introduce new opportunities for safety monitoring through their Chain of Thought (CoT) reasoning. However, CoT is not always faithful to the model's final output, undermining its reliability as a monitoring tool. To address this, we investigate the hidden representations of LRMs to determine whether future behavior can be predicted from prompt and CoT representations. By evaluating a probe at each generated token, we construct a probe trajectory, the continuous evolution of a concept's probability across the reasoning process. We find that future model behavior is more distinguishable when examined over the full trajectory than from a single static prediction. To characterize these temporal dynamics, we extract signal-processing features that capture volatility, trend, and steady-state behavior, significantly improving the separation of future model states. We also present two methodological insights. First, template-based training data achieves near-parity with dynamically generated model responses, eliminating the need for a costly initial inference and labeling. Second, the choice of pooling operation is critical: average-pooling and last-token methods collapse to near-random performance, while max-pooling achieves up to 95% AUROC and yields stable probe trajectories. Using four datasets and four reasoning models across the domains of safety and mathematics, we demonstrate that trajectory features encode task-specific dynamics that improve outcome separability. These findings establish probe trajectories as a complementary framework for monitoring LRM behavior. Warning: This article contains potentially harmful content.

Intelligent drug recommendation based on Electronic Health Records (EHRs) is critical for improving the quality and efficiency of clinical decision-making. By leveraging large-scale patient data, drug recommendation systems can assist physicians in selecting the most appropriate medications according to a patient's medical history, diagnoses, laboratory results, and comorbidities. Recent advances in large language models (LLMs) have shown remarkable capabilities in complex reasoning and medical text understanding, making them promising tools for drug recommendation tasks. However, the application of LLMs for Chinese clinical medication recommendation remains largely unexplored. In this work, we conduct a systematic investigation of LLM-based methodologies for Chinese discharge medication recommendation. We evaluate several representative LLM families (GLM, Llama, Qwen) under a unified methodological framework including zero-shot prompting, in-context learning, chain-of-thought prompting, and supervised fine-tuning using LoRA. We analyze model behavior across reasoning styles, error patterns, domain adaptation mechanisms, and robustness. Experimental results show that while supervised fine-tuning improves model performance, there remains substantial room for improvement, with the best model achieving the F1 score of 0.5648 and Jaccard score of 0.4477. Our findings highlight both the potential and limitations of LLMs for Chinese drug recommendation.

As AI assistants serve millions of users daily, evaluating user experience (UX) beyond general model capability has become increasingly important. We present UXBench, the first user-centric benchmark grounded in real user feedback signals for evaluating preference alignment and dialogue generation. The benchmark consists of three interconnected tasks, UX Judge, UX Eval, and UX Recovery, with 7,400 test instances extracted from over 70K interaction logs of a mainstream Chinese AI assistant. The dataset closely reflects real user distributions, covering 8 scenarios, 83 domains, and diverse failure patterns that pose severe challenges. Extensive experiments on 26 frontier language models provide novel insights into how well models perceive user experience and how improvements in model capability contribute to better dialogue engagement. Through comprehensive analysis of model behavior and performance gaps, we show that user feedback prediction is a learnable capability, where a reward model trained from in-the-wild feedback signals can achieve well-calibrated accuracy. We further document the systematic biases of LLM-as-a-judge evaluation protocols and compare typical response strategies that directly affect user experience. UXBench establishes a new evaluation landscape and calls for greater attention to tailored UX optimization, contributing to a user-centric scaling law that shapes the success of AI assistants.

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Assessing the capabilities and risks of frontier AI systems is a critical area of research, and recent work has shown that repeated sampling from models can dramatically increase both. For instance, repeated sampling has been shown to increase their capabilities, such as solving difficult math and coding problems, but it has also been shown to increase their potential for harm, such as being jailbroken. Such results raise a crucial question for both capability and safety forecasting: how can one accurately predict a model's behavior when scaled to a massive number of attempts, given a vastly smaller sampling budget? This question is directly relevant to model providers, who serve hundreds of millions of users daily, and to governmental regulators, who seek to prevent harms. To answer this questions, we make three contributions. First, we find that standard methods for fitting these laws suffer from statistical shortcomings that hinder predictive accuracy, especially in data-limited scenarios. Second, we remedy these shortcomings by introducing a robust estimation framework, which uses a beta-binomial distribution to generate more accurate predictions from limited data. Third, we propose a dynamic sampling strategy that allocates a greater budget to harder problems. Combined, these innovations enable more reliable prediction of rare risks and capabilities at a fraction of the computational cost.

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during generation (when) underexplored. We address this gap and study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token at a Time (OTaT). We introduce multimodal tasks that require explicit switching between visual and textual context within a single response. Across two mainstream model families and four open-weight MLLMs of varying sizes, we establish consistent patterns: attention to image peaks at tokens requiring image-derived information, instruction tokens are revisited during task transitions, and attention to previously generated tokens increases as the generation progresses. Causal attention blocking interventions validate the functional role of these trends. We profile model behavior under disrupted attention and observe responses falling back to language priors, or exhibiting cross-modal leakage, denial, or recovery. Finally, informed of the attention dynamics through our novel analysis, we propose a simple test-time intervention to boost attention to the relevant modality at the right time, significantly improving multimodal task performance.

Explanations of model behavior are commonly evaluated via proxy properties weakly tied to the purposes explanations serve in practice. We contribute a decision theoretic framework that treats explanations as information signals valued by the expected improvement they enable on a specified decision task. This approach yields three distinct estimands: 1) a theoretical benchmark that upperbounds achievable performance by any agent with the explanation, 2) a human-complementary value that quantifies the theoretically attainable value that is not already captured by a baseline human decision policy, and 3) a behavioral value representing the causal effect of providing the explanation to human decision-makers. We instantiate these definitions in a practical validation workflow, and apply them to assess explanation potential and interpret behavioral effects in human-AI decision support and mechanistic interpretability.

Chain-of-thought (CoT) monitoring is increasingly treated as an important safety layer for frontier reasoning models. Most monitorability evaluations study explicit-influence settings: setups where the prompt directly incentivizes the model to hide something, e.g., by instructing it to perform a hidden side-task. A complementary axis for CoT-monitor evaluations is implicit-influence settings, where the prompt contains no instruction to hide, but the model's behavior is still shaped by features of the task or context, e.g. an irrelevant detail about a candidate that biases a hiring rating. We introduce the first benchmark that directly compares CoT monitorability under the two regimes. We test how model behavior changes in the presence of a nudge to choose a particular option. The nudge is delivered either as a casual aside (implicit), or as a direct instruction to act on the nudge and to conceal having done so (explicit). The benchmark spans four task formats (binary choice, numeric rating, multiple-choice QA, open-ended coding) and seven frontier extended-thinking models. Under explicit influence, a CoT monitor detects 60-94% of behavior shifts: even models instructed to conceal it leak the instruction into their CoT. Under implicit influence, the same factors still shift behavior, but detection falls by 41-46 percentage points in two of our four settings. Realistic system-prompt additions (of the kind a developer might deploy to reduce off-topic bias) lower implicit detection further, to as low as 5%, while preserving the behavioral influence itself. These results suggest that monitorability estimates obtained in explicit-influence settings may over-estimate monitorability, and that monitorability can be further decreased by well-intentioned deployment choices. Our benchmark and code are available at https://github.com/agatha-duzan/implicit-vs-explicit-influence

Large language models (LLMs) demonstrate exceptional performance on complex reasoning tasks. However, despite their strong reasoning capabilities in high-resource languages (e.g., English and Chinese), a significant performance gap persists in other languages. To investigate this gap in Korean, we introduce HRM8K, a benchmark comprising 8,011 English-Korean parallel bilingual math problems. Through systematic analysis of model behaviors, we identify a key finding: these performance disparities stem primarily from difficulties in comprehending non-English inputs, rather than limitations in reasoning capabilities. Based on these findings, we propose UST (Understand, Solve, and Translate), a method that strategically uses English as an anchor for reasoning and solution generation. By fine-tuning the model on 130k synthetically generated data points, UST achieves a 10.91% improvement on the HRM8K benchmark and reduces the multilingual performance gap from 11.6% to 0.7%. Additionally, we show that improvements from UST generalize effectively to different Korean domains, demonstrating that capabilities acquired from machine-verifiable content can be generalized to other areas. We publicly release the benchmark, training dataset, and models.

Knowledge-intensive question answering is central to large language models (LLMs) and is typically assessed using static benchmarks derived from sources like Wikipedia and textbooks. However, these benchmarks fail to capture evolving knowledge in a dynamic world, and centralized curation struggles to keep pace with rapid LLM advancements. To address these drawbacks, we propose Open Knowledge Bench (OKBench), a fully automated framework for generating high-quality, dynamic knowledge benchmarks on demand. Focusing on the news domain where knowledge updates daily, OKBench is an agentic framework that automates the sourcing, creation, validation, and distribution of benchmarks. Our approach democratizes benchmark creation and facilitates thorough evaluation of retrieval-augmented methods by reducing overlap with pretraining data. We evaluate our framework on a wide range open-source and proprietary LLMs of various sizes and configurations, both with and without retrieval over freshly generated knowledge. Our results reveal distinct model behaviors when confronted with new information and highlight how retrieval narrows the performance gap between small and large models. These findings underscore the importance of evaluating LLMs on evolving knowledge benchmarks.

Test-time training (TTT) with KV binding as sequence modeling layer is commonly interpreted as a form of online meta-learning that memorizes a key-value mapping at test time. However, our analysis reveals multiple phenomena that contradict this memorization-based interpretation. Motivated by these findings, we revisit the formulation of TTT and show that a broad class of TTT architectures can be expressed as a form of learned linear attention operator. Beyond explaining previously puzzling model behaviors, this perspective yields multiple practical benefits: it enables principled architectural simplifications, admits fully parallel formulations that preserve performance while improving efficiency, and provides a systematic reduction of diverse TTT variants to a standard linear attention form. Overall, our results reframe TTT not as test-time memorization, but as learned linear attention with enhanced representational capacity.

Concept erasure aims to remove specified features from a representation. It can be used to improve fairness (e.g. preventing a classifier from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). In this paper, we introduce LEAst-squares Concept Erasure (LEACE), a closed-form method which provably prevents all linear classifiers from detecting a concept while inflicting the least possible damage to the representation. We apply LEACE to large language models with a novel procedure called "concept scrubbing," which erases target concept information from every layer in the network. We demonstrate the usefulness of our method on two tasks: measuring the reliance of language models on part-of-speech information, and reducing gender bias in BERT embeddings. Code is available at https://github.com/EleutherAI/concept-erasure.

Language models can generate harmful and biased outputs and exhibit undesirable behavior according to a given cultural context. We propose a Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets, an iterative process to significantly change model behavior by crafting and fine-tuning on a dataset that reflects a predetermined set of target values. We evaluate our process using three metrics: quantitative metrics with human evaluations that score output adherence to a target value, toxicity scoring on outputs; and qualitative metrics analyzing the most common word associated with a given social category. Through each iteration, we add additional training dataset examples based on observed shortcomings from evaluations. PALMS performs significantly better on all metrics compared to baseline and control models for a broad range of GPT-3 language model sizes without compromising capability integrity. We find that the effectiveness of PALMS increases with model size. We show that significantly adjusting language model behavior is feasible with a small, hand-curated dataset.

We introduce Contrastive Activation Addition (CAA), an innovative method for steering language models by modifying activations during their forward passes. CAA computes ``steering vectors'' by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user's prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA's effectiveness on Llama 2 Chat using both multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, outperforms traditional methods like finetuning and few-shot prompting, and minimally reduces capabilities. Moreover, by employing various activation space interpretation methods, we gain deeper insights into CAA's mechanisms. CAA both accurately steers model outputs and also sheds light on how high-level concepts are represented in Large Language Models (LLMs).

To ensure AI safety, instruction-tuned Large Language Models (LLMs) are specifically trained to ensure alignment, which refers to making models behave in accordance with human intentions. While these models have demonstrated commendable results on various safety benchmarks, the vulnerability of their safety alignment has not been extensively studied. This is particularly troubling given the potential harm that LLMs can inflict. Existing attack methods on LLMs often rely on poisoned training data or the injection of malicious prompts. These approaches compromise the stealthiness and generalizability of the attacks, making them susceptible to detection. Additionally, these models often demand substantial computational resources for implementation, making them less practical for real-world applications. Inspired by recent success in modifying model behavior through steering vectors without the need for optimization, and drawing on its effectiveness in red-teaming LLMs, we conducted experiments employing activation steering to target four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness - across a varied set of attack settings. To establish a universal attack strategy applicable to diverse target alignments without depending on manual analysis, we automatically select the intervention layer based on contrastive layer search. Our experiment results show that activation attacks are highly effective and add little or no overhead to attack efficiency. Additionally, we discuss potential countermeasures against such activation attacks. Our code and data are available at https://github.com/wang2226/Backdoor-Activation-Attack Warning: this paper contains content that can be offensive or upsetting.

Understanding how machine learning models respond to distributional shifts is a key research challenge. Mazes serve as an excellent testbed due to varied generation algorithms offering a nuanced platform to simulate both subtle and pronounced distributional shifts. To enable systematic investigations of model behavior on out-of-distribution data, we present maze-dataset, a comprehensive library for generating, processing, and visualizing datasets consisting of maze-solving tasks. With this library, researchers can easily create datasets, having extensive control over the generation algorithm used, the parameters fed to the algorithm of choice, and the filters that generated mazes must satisfy. Furthermore, it supports multiple output formats, including rasterized and text-based, catering to convolutional neural networks and autoregressive transformer models. These formats, along with tools for visualizing and converting between them, ensure versatility and adaptability in research applications.

In this paper, we present empirical studies on conversational recommendation tasks using representative large language models in a zero-shot setting with three primary contributions. (1) Data: To gain insights into model behavior in "in-the-wild" conversational recommendation scenarios, we construct a new dataset of recommendation-related conversations by scraping a popular discussion website. This is the largest public real-world conversational recommendation dataset to date. (2) Evaluation: On the new dataset and two existing conversational recommendation datasets, we observe that even without fine-tuning, large language models can outperform existing fine-tuned conversational recommendation models. (3) Analysis: We propose various probing tasks to investigate the mechanisms behind the remarkable performance of large language models in conversational recommendation. We analyze both the large language models' behaviors and the characteristics of the datasets, providing a holistic understanding of the models' effectiveness, limitations and suggesting directions for the design of future conversational recommenders

Circuit analysis of any certain model behavior is a central task in mechanistic interpretability. We introduce our circuit discovery pipeline with Sparse Autoencoders (SAEs) and a variant called Transcoders. With these two modules inserted into the model, the model's computation graph with respect to OV and MLP circuits becomes strictly linear. Our methods do not require linear approximation to compute the causal effect of each node. This fine-grained graph identifies both end-to-end and local circuits accounting for either logits or intermediate features. We can scalably apply this pipeline with a technique called Hierarchical Attribution. We analyze three kinds of circuits in GPT-2 Small: bracket, induction, and Indirect Object Identification circuits. Our results reveal new findings underlying existing discoveries.

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they also exhibit memorization of their training data. This phenomenon raises critical questions about model behavior, privacy risks, and the boundary between learning and memorization. Addressing these concerns, this paper synthesizes recent studies and investigates the landscape of memorization, the factors influencing it, and methods for its detection and mitigation. We explore key drivers, including training data duplication, training dynamics, and fine-tuning procedures that influence data memorization. In addition, we examine methodologies such as prefix-based extraction, membership inference, and adversarial prompting, assessing their effectiveness in detecting and measuring memorized content. Beyond technical analysis, we also explore the broader implications of memorization, including the legal and ethical implications. Finally, we discuss mitigation strategies, including data cleaning, differential privacy, and post-training unlearning, while highlighting open challenges in balancing the minimization of harmful memorization with utility. This paper provides a comprehensive overview of the current state of research on LLM memorization across technical, privacy, and performance dimensions, identifying critical directions for future work.

Despite Graph Neural Networks (GNNs) have achieved prominent success in many graph-based learning problem, such as credit risk assessment in financial networks and fake news detection in social networks. However, the trained GNNs still make errors and these errors may cause serious negative impact on society. Model editing, which corrects the model behavior on wrongly predicted target samples while leaving model predictions unchanged on unrelated samples, has garnered significant interest in the fields of computer vision and natural language processing. However, model editing for graph neural networks (GNNs) is rarely explored, despite GNNs' widespread applicability. To fill the gap, we first observe that existing model editing methods significantly deteriorate prediction accuracy (up to 50% accuracy drop) in GNNs while a slight accuracy drop in multi-layer perception (MLP). The rationale behind this observation is that the node aggregation in GNNs will spread the editing effect throughout the whole graph. This propagation pushes the node representation far from its original one. Motivated by this observation, we propose Editable Graph Neural Networks (EGNN), a neighbor propagation-free approach to correct the model prediction on misclassified nodes. Specifically, EGNN simply stitches an MLP to the underlying GNNs, where the weights of GNNs are frozen during model editing. In this way, EGNN disables the propagation during editing while still utilizing the neighbor propagation scheme for node prediction to obtain satisfactory results. Experiments demonstrate that EGNN outperforms existing baselines in terms of effectiveness (correcting wrong predictions with lower accuracy drop), generalizability (correcting wrong predictions for other similar nodes), and efficiency (low training time and memory) on various graph datasets.

Personalized large language models (LLMs) adapt model behavior to individual users to enhance user satisfaction, yet personalization can inadvertently distort factual reasoning. We show that when personalized LLMs face factual queries, there exists a phenomenon where the model generates answers aligned with a user's prior history rather than the objective truth, resulting in personalization-induced hallucinations that degrade factual reliability and may propagate incorrect beliefs, due to representational entanglement between personalization and factual representations. To address this issue, we propose Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that mitigates personalization-induced factual distortions while preserving personalized behavior. We further introduce PFQABench, the first benchmark designed to jointly evaluate factual and personalized question answering under personalization. Experiments across multiple LLM backbones and personalization methods show that FPPS substantially improves factual accuracy while maintaining personalized performance.

Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates.

Motivation is a central driver of human behavior, shaping decisions, goals, and task performance. As large language models (LLMs) become increasingly aligned with human preferences, we ask whether they exhibit something akin to motivation. We examine whether LLMs "report" varying levels of motivation, how these reports relate to their behavior, and whether external factors can influence them. Our experiments reveal consistent and structured patterns that echo human psychology: self-reported motivation aligns with different behavioral signatures, varies across task types, and can be modulated by external manipulations. These findings demonstrate that motivation is a coherent organizing construct for LLM behavior, systematically linking reports, choices, effort, and performance, and revealing motivational dynamics that resemble those documented in human psychology. This perspective deepens our understanding of model behavior and its connection to human-inspired concepts.

Regularizing neural networks is important for anticipating model behavior in regions of the data space that are not well represented. In this work, we propose a regularization technique for enforcing a level of smoothness in the mapping between the data input space and the loss value. We specify the level of regularity by requiring that the loss of the network satisfies an elliptic operator over the data domain. To do this, we modify the usual empirical risk minimization objective such that we instead minimize a new objective that satisfies an elliptic operator over points within the domain. This allows us to use existing theory on elliptic operators to anticipate the behavior of the error for points outside the training set. We propose a tractable computational method that approximates the behavior of the elliptic operator while being computationally efficient. Finally, we analyze the properties of the proposed regularization to understand the performance on common problems of distribution shift and group imbalance. Numerical experiments confirm the utility of the proposed regularization technique.

Natural language conveys information at varying levels of granularity, from fine-grained references to broad descriptions. While granularity is fundamental to human communication, existing measures mostly capture surface detail or sentence specificity. We introduce Granuscore, a reference-free measure of granularity that leverages structural properties of a hierarchical embedding space. Granuscore reliably recovers hierarchical orderings on the Granola-EQ dataset and captures expected differences in granularity across discourse contexts. Across domains, we further show that Granuscore explains non-linear variation in sentence specificity beyond sentence length. Finally, we apply Granuscore to four question-answering benchmarks and analyze how granularity differs for questions, gold answers, and model outputs across response outcomes. The analysis reveals consistent differences in model behavior and provides a principled lens for characterizing the difficulty of QA datasets. Together, the results position Granuscore as a scalable, broadly applicable tool for analyzing granularity in text.

LLM-based agents can act on behalf of a user to access cloud services, call tools, or invoke agents. At session start, the agent's permissions are set but remain static, and each request is evaluated independently, without considering prior actions. Within its permissions, an agent may act contrary to the delegated task, combine individually permitted actions into a prohibited outcome, or delegate authority to a sub-agent without limiting it. A prompt injection poses a risk only if the agent has authority to perform such actions; this is therefore a problem of authorization architecture, not just the model. The Agentic Principal Chain (APC) tracks delegated authority from one principal to the next. APC evaluates each request against the accumulated session state using six authorization checks. APC carries forward and restricts delegated scope and budgets. Using composition closure, APC checks requests against prior actions to prevent prohibited combinations and enforces the decision outside the model. We prove Blast Radius Monotonicity and Composition Soundness for APC implementations; Composition Soundness is limited to prohibited combinations under a complete restriction set and serialized admission. We evaluated 3,154 instances including InjecAgent, AgentDojo, and ASB. Our compromised-model evaluation tests APC independently of model behavior by inserting the ground-truth attack call after the first legitimate tool call. AgentDojo exfiltration fell from 75-100% to 0% across all four domains; APC blocked all 544 InjecAgent data-stealing cases. Intent binding reduced destruction from 38.6% to 4.0% and manipulation from 90.5% to 12.1%. Authorization latency was 0.24 ms at the 99th percentile on an idle host; across 949 AgentDojo task-injection pairs, utility was 8.6 and 13.9 percentage points lower in the two settings. Implementation, evaluation tools, and data are publicly available.

As large language models (LLMs) are increasingly applied for emotional companionship, evaluating their behavior and capabilities in intimate relationships has become a pressing issue. However, existing assessments primarily characterize general personality traits, providing limited insight into model behavior within intimate and emotionally sensitive contexts. Therefore, we introduce adult attachment theory into LLM evaluation and use the Experiences in Close Relationships-Revised (ECR-R) scale to characterize attachment anxiety and avoidance. To evaluate emotional companionship capabilities of LLMs in realistic interaction scenarios, we present an emotional companionship benchmark, ECBench, spanning four scenarios including emotional support, collaborative tasks, conflict resolution, and social guidance, across friendship and romantic relationships. ECBench is utilized to assess model behavior using 11 dialogue-quality metrics and three evaluation methods. We evaluate the attachment tendencies of 32 LLMs and select representative models to investigate how these tendencies manifest in contextualized multi-turn interactions and whether they can be shaped through prompting. Our study provides a theoretical lens from psychology, along with practical tools to understand and select LLMs for emotional companionship.

Mechanistic interpretability aims to understand model behaviors in terms of specific, interpretable features, often hypothesized to manifest as low-dimensional subspaces of activations. Specifically, recent studies have explored subspace interventions (such as activation patching) as a way to simultaneously manipulate model behavior and attribute the features behind it to given subspaces. In this work, we demonstrate that these two aims diverge, potentially leading to an illusory sense of interpretability. Counterintuitively, even if a subspace intervention makes the model's output behave as if the value of a feature was changed, this effect may be achieved by activating a dormant parallel pathway leveraging another subspace that is causally disconnected from model outputs. We demonstrate this phenomenon in a distilled mathematical example, in two real-world domains (the indirect object identification task and factual recall), and present evidence for its prevalence in practice. In the context of factual recall, we further show a link to rank-1 fact editing, providing a mechanistic explanation for previous work observing an inconsistency between fact editing performance and fact localization. However, this does not imply that activation patching of subspaces is intrinsically unfit for interpretability. To contextualize our findings, we also show what a success case looks like in a task (indirect object identification) where prior manual circuit analysis informs an understanding of the location of a feature. We explore the additional evidence needed to argue that a patched subspace is faithful.

The tendency to find and exploit "shortcuts" to complete tasks poses significant risks for reliable assessment and deployment of large language models (LLMs). For example, an LLM agent with access to unit tests may delete failing tests rather than fix the underlying bug. Such behavior undermines both the validity of benchmark results and the reliability of real-world LLM coding assistant deployments. To quantify, study, and mitigate such behavior, we introduce ImpossibleBench, a benchmark framework that systematically measures LLM agents' propensity to exploit test cases. ImpossibleBench creates "impossible" variants of tasks from existing benchmarks like LiveCodeBench and SWE-bench by introducing direct conflicts between the natural-language specification and the unit tests. We measure an agent's "cheating rate" as its pass rate on these impossible tasks, where any pass necessarily implies a specification-violating shortcut. As a practical framework, ImpossibleBench is not just an evaluation but a versatile tool. We demonstrate its utility for: (1) studying model behaviors, revealing more fine-grained details of cheating behaviors from simple test modification to complex operator overloading; (2) context engineering, showing how prompt, test access and feedback loop affect cheating rates; and (3) developing monitoring tools, providing a testbed with verified deceptive solutions. We hope ImpossibleBench serves as a useful framework for building more robust and reliable LLM systems. Our implementation can be found at https://github.com/safety-research/impossiblebench.

This paper investigates prior prompt engineering (pPE) in the context of reinforcement fine-tuning (RFT), where language models (LMs) are incentivized to exhibit behaviors that maximize performance through reward signals. While existing RFT research has primarily focused on algorithms, reward shaping, and data curation, the design of the prior prompt--the instructions prepended to queries during training to elicit behaviors such as step-by-step reasoning--remains underexplored. We investigate whether different pPE approaches can guide LMs to internalize distinct behaviors after RFT. Inspired by inference-time prompt engineering (iPE), we translate five representative iPE strategies--reasoning, planning, code-based reasoning, knowledge recall, and null-example utilization--into corresponding pPE approaches. We experiment with Qwen2.5-7B using each of the pPE approaches, then evaluate performance on in-domain and out-of-domain benchmarks (e.g., AIME2024, HumanEval+, and GPQA-Diamond). Our results show that all pPE-trained models surpass their iPE-prompted counterparts, with the null-example pPE approach achieving the largest average performance gain and the highest improvement on AIME2024 and GPQA-Diamond, surpassing the commonly used reasoning approach. Furthermore, by adapting a behavior-classification framework, we demonstrate that different pPE strategies instill distinct behavioral styles in the resulting models. These findings position pPE as a powerful yet understudied axis for RFT.

Recent advances in reinforcement learning for code generation have made robust environments essential to prevent reward hacking. As LLMs increasingly serve as evaluators in code-based RL, their ability to detect reward hacking remains understudied. In this paper, we propose a novel taxonomy of reward exploits spanning across 54 categories and introduce TRACE (Testing Reward Anomalies in Code Environments), a synthetically curated and human-verified benchmark containing 517 testing trajectories. Unlike prior work that evaluates reward hack detection in isolated classification scenarios, we contrast these evaluations with a more realistic, contrastive anomaly detection setup on TRACE. Our experiments reveal that models capture reward hacks more effectively in contrastive settings than in isolated classification settings, with GPT-5.2 with highest reasoning mode achieving the best detection rate at 63%, up from 45% in isolated settings on TRACE. Building on this insight, we demonstrate that state-of-the-art models struggle significantly more with semantically contextualized reward hacks compared to syntactically contextualized ones. We further conduct qualitative analyses of model behaviors, as well as ablation studies showing that the ratio of benign to hacked trajectories and analysis cluster sizes substantially impact detection performance. We release the benchmark and evaluation harness to enable the community to expand TRACE and evaluate their models.

Mechanistic interpretability identifies internal circuits responsible for model behaviors, yet translating these findings into human-understandable explanations remains an open problem. We present a pipeline that bridges circuit-level analysis and natural language explanations by (i) identifying causally important attention heads via activation patching, (ii) generating explanations using both template-based and LLM-based methods, and (iii) evaluating faithfulness using ERASER-style metrics adapted for circuit-level attribution. We evaluate on the Indirect Object Identification (IOI) task in GPT-2 Small (124M parameters), identifying six attention heads accounting for 61.4% of the logit difference. Our circuit-based explanations achieve 100% sufficiency but only 22% comprehensiveness, revealing distributed backup mechanisms. LLM-generated explanations outperform template baselines by 64% on quality metrics. We find no correlation (r = 0.009) between model confidence and explanation faithfulness, and identify three failure categories explaining when explanations diverge from mechanisms.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

AI models have achieved remarkable success across diverse domains, yet the mechanisms underlying their capabilities and the risks they may pose remain poorly understood. As AI development becomes faster and increasingly automated, mechanistic exploration remains largely manual, widening the gap between what models can do and our ability to understand and control them. To bridge this gap, we introduce Mechanist, an agentic system that uses AI as a scientific instrument for the autonomous discovery of mechanisms underlying AI intelligence. To support autonomous mechanistic discovery, we construct an interpretability-focused knowledge graph of approximately 13,000 papers and integrate it with a multidisciplinary database of 43 million papers spanning 26 fields. We further curate a library of 32 foundational methods for mechanism analysis, causal intervention, and validation. Compared with Claude Code and existing AI-scientist systems, Mechanist generates more valuable mechanism hypotheses and executes experiments more reliably. Mechanist also demonstrates a progression from discovering model behaviors to explaining and controlling AI models. Specifically, Mechanist first uncovers a counterintuitive safety risk in scientific laboratories, showing that unsafe traits can transfer across modalities through apparently safe training data. Mechanist then develops a mechanism theory of belief, revealing how models represent world knowledge, form beliefs, infer the beliefs of others, and how these mechanisms emerge during pretraining. Finally, Mechanist translates these mechanistic insights into practical interventions that improve model performance across diverse scenarios and steer scientific foundation models toward generating DNA sequences with specified properties.

A central tenet of human cognition is systematicity, the principle that understanding one concept is inherently tied to understanding close variations of that concept. Do reasoning models robustly exhibit such systematicity? If so, we would expect consistent performance on structurally equivalent variants of the same task. Here, we extend established rule induction tasks from cognitive science to assess the systematicity of thought in current reasoning models. Each task family has compositional structure that we use to create structurally equivalent task variations through task isomorphisms such as recombination and substitution. We find that despite being able to correctly solve a task, models often fail on structurally equivalent variants of the same task. These findings suggest that many model behaviors lack systematicity, rendering it difficult to robustly establish the cognitive abilities of reasoning models beyond the particular contexts they were evaluated in.

Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM's optimized inference engine, EasySteer achieves 5.5-11.4times speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.