ICLR 2026 论文列表

Large language models are becoming general knowledge engines for diverse applications. However, their computations are deeply entangled after training, resisting modularization which complicates interpretability, auditing, and long-term maintenance. We introduce Jet Expansions, a framework for expanding computational graphs using jet operators that generalize truncated Taylor series. Our method systematically decomposes language models into explicit input-to-output computational paths and complementary remainders. This functional decomposition provides a principled, knife-like operator for cutting through entanglement in LLMs, enabling scalable model inspection. We demonstrate how Jet Expansions ground and subsume the popular interpretability technique Logit Lens, reveal a (super-)exponential path structure with respect to recursive residual depth, and support several interpretability applications, including sketching a transformer language model with $n$-gram statistics extracted from its computations and indexing model toxicity levels *without* curated benchmarks.

Large language models (LLMs) acquire extensive prior knowledge through large-scale pretraining and can be further enhanced via supervised fine-tuning (SFT) or reinforcement learning (RL)-based post-training. A growing body of evidence has shown that RL fine-tuning improves the capability of LLMs beyond what SFT alone achieves. However, the underlying mechanisms why RL fine-tuning is able to enhance the capability of various LLMs with distinct intrinsic characteristics remain underexplored. In this study, we draw inspiration from prior work on edge attribution patching (EAP) to investigate the internal differences of LLMs before and after RL fine-tuning. Our analysis across multiple model families and mathematical datasets shows two robust effects of online RL post-training: (i) an overall increase in average activation intensity, indicating that more internal pathways are engaged and their signals become stronger, and (ii) greater diversity in activation patterns, reflected by higher entropy and less concentrated edge distributions. These changes suggest that RL reshapes information flow to be both more redundant and more flexible, which may explain its advantage in mathematical generalization. Notably, models fine-tuned with Direct Preference Optimization (DPO) deviate from these trends, exhibiting substantially weaker or inconsistent internal changes compared to PPO- and GRPO-based training. Together, our findings provide a unified view of how RL fine-tuning systematically alters the internal circuitry of LLMs and highlight the methodological distinctions between online RL and preference-based approaches. Our code is open source at https://anonymous.4open.science/r/llm_rl_probing_analysis-F673.

Despite the increasing prevalence of large language models (LLMs), we still have a limited understanding of how their representational spaces are structured. This limits our ability to interpret how and what they learn or relate them to learning in humans. We argue LLMs are best seen as an instance of lossy compression, where over training they learn by retaining only information in their training data relevant to their objective(s). We show pre-training results in models that are optimally compressed for next-sequence prediction, approaching the Information Bottleneck bound on compression. Across an array of open weights models, each compresses differently, likely due to differences in the data and training recipes used. However even across different families of LLMs the optimality of a model's compression, and the information present in it, can predict downstream performance on MMLU-Pro, letting us directly link representational structure to actionable insights about model performance. In the general case the work presented here offers a unified Information-Theoretic framing for how these models learn that is deployable at scale.

Language models serve as proxies for human preference judgements in alignment and evaluation, yet they exhibit systematic miscalibration, prioritizing superficial patterns over substantive qualities. This bias manifests as overreliance on features like length, structure, and style, leading to issues like reward hacking and unreliable evaluations. However, the connection between training data artifacts and the miscalibrated preferences exhibited by models remains poorly understood. In this work, we systematically investigate the relationship between training data biases and preference model miscalibration across five idiosyncratic features of language model generations: length, structure, jargon, sycophancy and vagueness. Using controlled counterfactual pairs, we first quantify the extent to which preference models favor responses with artificially magnified biases (\textit{skew}), finding this preference occurs in $>60$\% of instances, and model preferences show high \textit{miscalibration} ($\approx 40$\%) compared to human preferences. Notably, bias features only show mild negative correlations to human preference labels (mean $r_{\mathrm{human}} = -0.12$) but show moderately strong positive correlations with labels from a strong reward model (mean $r_{\mathrm{model}} = +0.36$), suggesting that models may overrely on spurious cues. To mitigate these issues, we propose a simple post-training method based on counterfactual data augmentation (CDA) using synthesized contrastive examples. Fine-tuning models with CDA reduces average miscalibration from 39.4\% to 32.5\% and average absolute skew difference from 20.5\% to 10.0\%, while maintaining overall RewardBench performance, indicating that targeted debiasing can strengthen the reliability of preference models within standard alignment pipelines.

Polysemanticity is pervasive in language models and remains a major challenge for interpretation and model behavioral control. Leveraging sparse autoencoders (SAEs), we map the polysemantic topology of two small models (Pythia-70M and GPT-2-Small) to identify SAE feature pairs that are semantically unrelated yet exhibit interference within models. We intervene at four loci (prompt, token, feature, neuron) and measure induced shifts in the next-token prediction distribution, uncovering polysemantic structures that expose a systematic vulnerability in these models. Critically, interventions distilled from counterintuitive interference patterns shared by two small models transfer reliably to larger instruction-tuned models (Llama-3.1-8B/70B-Instruct and Gemma-2-9B-Instruct), yielding predictable behavioral shifts without access to model internals. These findings challenge the view that polysemanticity is purely stochastic, demonstrating instead that interference structures generalize across scale and family. Such generalization suggests a convergent, higher-order organization of internal representations, which is only weakly aligned with intuition and structured by latent regularities, offering new possibilities for both black-box control and theoretical insight into human and artificial cognition.

Representation learning is critical to the empirical and theoretical success of reinforcement learning. However, many existing methods are induced from model-learning aspects, misaligning them with the RL task in hand. This work introduces the Spectral Bellman Method, a novel framework derived from the Inherent Bellman Error (IBE) condition. It aligns representation learning with the fundamental structure of Bellman updates across a space of possible value functions, making it directly suited for value-based RL. Our key insight is a fundamental spectral relationship: under the zero-IBE condition, the transformation of a distribution of value functions by the Bellman operator is intrinsically linked to the feature covariance structure. This connection yields a new, theoretically-grounded objective for learning state-action features that capture this Bellman-aligned covariance, requiring only a simple modification to existing algorithms. We demonstrate that our learned representations enable structured exploration by aligning feature covariance with Bellman dynamics, improving performance in hard-exploration and long-horizon tasks. Our framework naturally extends to multi-step Bellman operators, offering a principled path toward learning more powerful and structurally sound representations for value-based RL.

A simple yet effective method for inference-time alignment of generative models is Best-of-$N$ (BoN), where $N$ outcomes are sampled from a reference policy, evaluated using a proxy-reward model, and the highest-scoring one is selected. While prior work argues that BoN is almost optimal in reward vs KL tradeoffs, the effectiveness of BoN depends critically on the quality of the proxy-reward model used for selection. For this purpose, we study BoN through a smooth version known as Soft Best-of-N (SBoN) and develop a theoretical framework to address this gap. We analyze the scaling behaviour of BoN by providing bounds on the KL divergence between the SBoN policy and the reference policy, offering insights into how performance varies with the number of samples. We also study the regret gap, i.e., the gap between the expected true-reward under the optimal policy and the SBoN policy. Our theoretical and empirical findings show that smoothing helps SBoN mitigate reward overoptimization, especially when the quality of the proxy-reward is low.

State Space Models (SSMs) have become the leading alternative to Transformers for sequence modeling tasks. Their primary advantage is efficiency in long-context and long-form generation, enabled by fixed-size memory and linear scaling of computational complexity. We begin this work by showing a simple theoretical result stating that SSMs cannot accurately solve any long-form generation problem, undermining their main competitive advantage. However, we show that this limitation can be mitigated by allowing SSMs interactive access to external tools. In fact, we show that given the right choice of tool access and problem-dependent training data, SSMs can learn to solve any tractable problem and generalize to arbitrary problem length/complexity (i.e., achieve length generalization). Following our theoretical finding, we demonstrate that tool-augmented SSMs achieve remarkable length generalization on a variety of arithmetic, reasoning, and coding tasks. These findings highlight SSMs as a potential efficient alternative to Transformers in interactive tool-based and agentic settings.

Distribution shifts introduce uncertainty that undermines the robustness and generalization capabilities of machine learning models. While conventional wisdom suggests that learning causal-invariant representations enhances robustness to such shifts, recent empirical studies present a counterintuitive finding: (i) empirical risk minimization (ERM) can rival or even outperform state-of-the-art out-of-distribution (OOD) generalization methods, and (ii) OOD generalization performance improves when all available covariates, including non-causal ones, are utilized. We present theoretical and empirical explanations that attribute this phenomenon to hidden confounding. Shifts in hidden confounding induce changes in data distributions that violate assumptions commonly made by existing approaches. Under such conditions, we prove that generalization requires learning environment-specific relationships, rather than relying solely on invariant ones. Furthermore, we explain why models augmented with non-causal but informative covariates can mitigate the challenges posed by hidden confounding shifts. These findings offer new theoretical insights and practical guidance, serving as a roadmap for future research on OOD generalization and principled covariate-selection strategies.

Converging evidence suggests that human systems of semantic categories achieve near-optimal compression via the Information Bottleneck (IB) complexity-accuracy tradeoff. Large language models (LLMs) are not trained for this objective, which raises the question: are LLMs capable of evolving efficient human-aligned semantic systems? To address this question, we focus on color categorization --- a key testbed of cognitive theories of categorization with uniquely rich human data --- and replicate with LLMs two influential human studies. First, we conduct an English color-naming study, showing that LLMs vary widely in their complexity and English-alignment, with larger instruction-tuned models achieving better alignment and IB-efficiency. Second, to test whether these LLMs simply mimic patterns in their training data or actually exhibit a human-like inductive bias toward IB-efficiency, we simulate cultural evolution of pseudo color-naming systems in LLMs via a method we refer to as Iterated in-Context Language Learning (IICLL). We find that akin to humans, LLMs iteratively restructure initially random systems towards greater IB-efficiency. However, only a model with strongest in-context capabilities (Gemini 2.0) is able to recapitulate the wide range of near-optimal IB-tradeoffs observed in humans, while other state-of-the-art models converge to low-complexity solutions. These findings demonstrate how human-aligned semantic categories can emerge in LLMs via the same fundamental principle that underlies semantic efficiency in humans.

Epilepsy affects over 50 million people worldwide, and one-third of patients suffer drug-resistant seizures where surgery offers the best chance of seizure freedom. Accurate localization of the epileptogenic zone (EZ) relies on intracranial EEG (iEEG). Clinical workflows, however, remain constrained by labor-intensive manual review. At the same time, existing data-driven approaches are typically developed on single-center datasets that are inconsistent in format and metadata, lack standardized benchmarks, and rarely release pathological event annotations, creating barriers to reproducibility, cross-center validation, and clinical relevance. With extensive efforts to reconcile heterogeneous iEEG formats, metadata, and recordings across publicly available sources, we present $\textbf{Omni-iEEG}$, a large-scale, pre-surgical iEEG resource comprising $\textbf{302 patients}$ and $\textbf{178 hours}$ of high-resolution recordings. The dataset includes harmonized clinical metadata such as seizure onset zones, resections, and surgical outcomes, all validated by board-certified epileptologists. In addition, Omni-iEEG provides over 36K expert-validated annotations of pathological events, enabling robust biomarker studies. Omni-iEEG serves as a bridge between machine learning and epilepsy research. It defines clinically meaningful tasks with unified evaluation metrics grounded in clinical priors, enabling systematic evaluation of models in clinically relevant settings. Beyond benchmarking, we demonstrate the potential of end-to-end modeling on long iEEG segments and highlight the transferability of representations pretrained on non-neurophysiological domains. Together, these contributions establish Omni-iEEG as a foundation for reproducible, generalizable, and clinically translatable epilepsy research.

Multiple agents are increasingly combined to make decisions with the expectation of achieving complementary performance, where the decisions they make together outperform those made individually. However, knowing how to improve the performance of collaborating agents requires knowing what information and strategies each agent employs. With a focus on human-AI pairings, we contribute a decision-theoretic framework for characterizing the value of information. By defining complementary information, our approach identifies opportunities for agents to better exploit available information–in AI-assisted decision workflows. We present a novel explanation technique (ILIV-SHAP) that adapts SHAP explanations to highlight human-complementing information. We validate the effectiveness of the framework on examples from chest X-ray diagnosis and deepfake detection and ILIV-SHAP through a study of human-AI decision-making. We also find that presenting ILIV-SHAP with AI predictions leads to reliably greater reductions in error over non-AI assisted decisions more than vanilla SHAP.

Humans organize knowledge into compact categories that balance compression with semantic meaning preservation. Large Language Models (LLMs) demonstrate striking linguistic abilities, yet whether they achieve this same balance remains unclear. We apply the Information Bottleneck principle to quantitatively compare how LLMs and humans navigate this compression-meaning trade-off. Analyzing embeddings from 40+ LLMs against classic human categorization benchmarks, we uncover three key findings. First, LLMs broadly align with human categories but miss fine-grained semantic distinctions crucial for human understanding. Second, LLMs demonstrate aggressive statistical compression, achieving ``optimal'' information-theoretic efficiency, while humans prioritize contextual richness and adaptive flexibility. Third, encoder models surprisingly outperform decoder models in human alignment, suggesting that generation and understanding rely on distinct mechanisms in current architectures. In addition, training dynamics analysis reveals that conceptual structure develops in distinct phases: rapid initial formation followed by architectural reorganization, with semantic processing migrating from deeper to mid-network layers as models discover more efficient encoding. These divergent strategies, where LLMs optimize for compression and humans for adaptive utility, reveal fundamental differences between artificial and biological intelligence, guiding development toward more human-aligned AI.

Large language models (LLMs) exhibit strong abilities in multi-document reasoning, yet their evidence identification is highly sensitive to input order. We trace this limitation to attention mechanisms, where many heads overemphasize sequence boundaries and neglect central content. We systematically analyze attention distributions under document permutations and discover a small subset of heads that consistently prioritize task-relevant documents regardless of position. We formalize these as Expert Heads, identified via activation frequency and stability across permutations. Experiments on LLaMA, Mistral, and Qwen reveal architecture-specific patterns: mid-layer heads in LLaMA and Mistral dominate semantic integration, while deeper-layer heads in Qwen specialize in evidence selection. Moreover, Expert Heads exhibit concentrated focus during understanding and more distributed engagement during generation. Their activation strongly correlates with answer correctness, providing diagnostic signals for hallucination detection. Leveraging Expert Heads for document voting significantly improves retrieval and ranking on HotpotQA, 2WikiMultiHopQA, and MuSiQue, outperforming dense retrievers and LLM-based ranking with minimal overhead. Ablations confirm that even a small subset achieves robust gains. Our findings establish Expert Heads as a stable and interpretable mechanism for evidence integration, offering new directions for context pruning, hallucination mitigation, and head-guided training of LLMs

Vision foundation models are typically trained as static feature extractors, forcing the burden of task adaptation onto large downstream models. We propose a different paradigm: instead of solely feeding visual features into language, we use language itself to dynamically guide the vision encoder. Our method, Language-Instructed Vision Embeddings (LIVE), leverages language as high-level guidance to produce task-centric embeddings at inference time—without requiring task-specific retraining. This enables the encoder to focus attention on contextually relevant aspects of the input, yielding more controllable and generalizable representations. Empirically, LIVE reduces visual hallucinations (+34 points on MMVP), outperforms vision–language models with orders of magnitude more parameters on visual question answering, and generalizes to unseen instructions and tasks---offering a direct path toward adaptive, instruction-driven visual intelligence.

Despite multilingual pretraining, large language models often struggle with non-English tasks, particularly in language control--the ability to respond in the intended language. We identify and characterize two key failure modes: the *multilingual transfer bottleneck* (correct language, incorrect task response) and the *language consistency bottleneck* (correct task response, wrong language). To systematically surface these issues, we design a four-scenario evaluation protocol spanning MMLU, MGSM, and XQuAD benchmarks. To probe these issues with interpretability, we extend logit lens analysis to track language probabilities layer by layer and compute cross-lingual semantic similarity of hidden states. The results reveal a three-phase internal structure: early layers align inputs into shared semantic space, middle layers perform task reasoning, and late layers drive language-specific generation. Guided by these insights, we introduce *selective fine-tuning* of only the final layers responsible for language control. On Qwen-3-32B and Bloom-7.1B, this method achieves over 98% language consistency across six languages while fine-tuning only 3–5% of parameters, without sacrificing task accuracy. Importantly, this result is nearly identical to that of full-scope fine-tuning (e.g., $>98\%$ language consistency for both methods across all prompt scenarios) but uses a fraction of the computational resources. To the best of our knowledge, this is the first approach to leverage *layer-localization of language control* for efficient multilingual adaptation.

Knowledge editing aims to efficiently correct factual errors in Language Models (LMs). The popular locate-then-edit approach modifies an MLP layer by finding an optimal mapping between its input vector (key) and output vector (value) that leads to the expression of the edited knowledge. However, existing methods without any constraints on the key and value vectors cause significant perturbations to the edited model. To address this, we propose Subspace Knowledge Edit (SUIT), a method that identifies and modifies only the subspace of critical features relevant to the edit. Our empirical results on LLaMA-3-8B, GPT-J-6B, and Qwen2.5-7B models show that SUIT dramatically improves knowledge preservation over strong baselines while maintaining high edit efficacy. This effectiveness confirms that SUIT successfully identifies the critical subspace for the edit. Further analyses provide additional validation for our approach. The source code and data will be released to the public upon publication of the paper.

Finetuning on narrow domains has become an essential tool to adapt Large Language Models (LLMs) to specific tasks and to create models with known unusual properties that are useful for safety research. Model diffing--the study of differences between base and finetuned models--is a promising approach for understanding how finetuning modifies neural networks. In this paper, we show that narrow finetuning creates easily readable biases in LLM activations that can be detected using simple model diffing tools, suggesting that the finetuning data is overrepresented in the model's activations. In particular, analyzing activation differences between base and finetuned models on the first few tokens of random text and steering with this difference allows us to recover the format and general content of the finetuning data. We call this the Activation Difference Lens (ADL). We demonstrate that these analyses significantly enhance an LLM-based interpretability agent's ability to identify subtle finetuning objectives through interaction with base and finetuned models. Our analysis spans synthetic document finetuning for false facts, emergent misalignment, subliminal learning, and taboo guessing game models across different architectures (Gemma, LLaMA, Qwen) and scales (1B to 32B parameters). Our work: (1) demonstrates that researchers should be aware that narrow finetuned models will represent their training data and objective very saliently, (2) warns AI safety and mechanistic interpretability researchers that these models might not be a realistic proxy for studying broader finetuning, despite current literature widely using them. While we show that mixing pretraining data into the finetuning corpus is enough to remove this bias, a deeper investigation is needed to understand the side effects of narrow finetuning and develop truly realistic case studies for model-diffing, safety and interpretability research.

We show that a language model’s ability to predict text is tightly linked to the breadth of its embedding space: models that spread their contextual representations more widely tend to achieve lower perplexity. Concretely, we find that representation dispersion—the average pairwise cosine distance among hidden vectors—strongly and negatively correlates with perplexity across diverse model families (LLaMA, Qwen, and others) and domains (Wikipedia, news, scientific abstracts). Beyond illustrating this link, we show how dispersion can be leveraged for a range of practical tasks—without requiring labeled data. First, measuring dispersion on unlabeled text allows us to predict downstream accuracy in new domains, offering a data‐efficient tool for model selection. Next, we find that identifying layers with higher dispersion pinpoints the best representations for retrieval‐based methods such as kNN‐LM, bypassing exhaustive layer‐by‐layer searches. Finally, we integrate a simple “push‐away” objective into training, which increases dispersion in both single‐domain and cross‐domain scenarios and directly improves perplexity in each.