ICML 2026 论文列表

Reasoning in knowledge-intensive domains remains challenging as intermediate steps are often not locally verifiable: unlike math or code, evaluating step correctness may require synthesizing clues across large external knowledge sources. As a result, subtle errors can propagate through reasoning traces, potentially never to be detected. Prior work has proposed process reward models (PRMs), including retrieval-augmented variants, but these methods operate post hoc, scoring completed trajectories, which prevents their integration into dynamic inference procedures. Here, we introduce Process Reward Agents (PRA), an inference-time method for providing domain-grounded, online, step-wise rewards to a frozen policy. In contrast to prior retrieval-augmented PRMs, PRA enables search-based decoding to rank and prune candidate trajectories at every generation step. Experiments on multiple medical reasoning benchmarks demonstrate that PRA consistently outperforms strong baselines, achieving 81.9% accuracy on MedQA with Qwen3-4B, a new state of the art at the 4B scale. Importantly, PRA generalizes to unseen frozen policy models ranging from 0.5B to 8B parameters, improving their accuracy by up to 25.7% without any policy model updates. More broadly, PRA suggests a paradigm in which frozen reasoners are decoupled from domain-specific reward modules, allowing the deployment of new backbones in complex domains without retraining. All code and data are publicly available at https://process-reward-agents.github.io/.

Large language models have shown strong reasoning abilities and are increasingly explored as high-level coordinators for multi-agent systems. However, directly deploying LLMs for coordination remains challenging, as effective policies often fail to reliably emerge at the low-level control stage, and inference costs limit scalability. We propose SynCoord (Synthetic Coordination Distillation), a self-supervised pipeline that distills task-level decision-making for cooperation from high-capacity reasoning models into lightweight agent policies. Our approach does not rely on explicit supervision or handcrafted coordination rules. Instead, we define a set of task-level tool interfaces that constrain LLM interaction and enable the collection of interaction trajectories, which are then used to train compact coordinated policies. This distillation process transfers coordination behaviors that are difficult to elicit through prompting alone, while substantially reducing inference overhead at execution time. We evaluate our method on cooperative multi-agent benchmarks including Overcooked-AI and Level-Based Foraging (LBF), under varying team sizes and environment scales. Experimental results show that the distilled policies achieve success rates and execution efficiency comparable to reinforcement learning–based methods, while exhibiting fewer erroneous or redundant actions. Moreover, the learned task-level coordination policy generalizes effectively to unseen team compositions and larger layouts without retraining.

Instruction following in LLMs captures models' ability to change their visible behaviors as requested by users. Instead, we study models' ability to control their chain-of-thought (CoT). This capability -- CoT controllability -- is undesirable because it could allow models to suppress signs of misbehavior in their CoT, thereby undermining our ability to monitor them. To measure this, we introduce the \emph{CoT-Control} evaluation suite. We show that reasoning models are less able to follow instructions in their CoT than in their outputs: on instructions like reasoning about a genetics problem without mentioning the word ``chromosome", Claude-Sonnet-4.5 complies only 5\% of the time. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models that are told they're being monitored exhibit slightly higher controllability. Similarly, eliciting controllability by adversarially optimizing prompts doesn’t meaningfully increase controllability. Our results leave us cautiously optimistic: reasoning models generally seem characterized by low CoT controllability. However, the mechanism behind this phenomenon is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs keep tracking controllability for future models.

In this work, we show that it is possible to extract significant amounts of alignment training data from a post-trained model -- useful to steer the model to improve certain capabilities such as long-context reasoning, safety, instruction following, and maths. While the majority of related work on memorisation has focused on measuring success of training data extraction through string matching, we argue that embedding models are better suited for our specific goals. Distances measured through a high quality embedding model can identify semantic similarities between strings that a different metric such as edit distance will struggle to capture. In fact, in our investigation, approximate string matching would have severely undercounted (by a conservative estimate of $10\times$) the amount of data that can be extracted due to trivial artifacts that deflate the metric. Interestingly, we find that models readily regurgitate training data that was used in post-training phases such as SFT or RL. We show that this data can be then used to train a base model, recovering a meaningful amount of the original performance. We believe our work exposes a possibly overlooked risk towards extracting alignment data. Finally, our work opens up an interesting discussion on the downstream effects of distillation practices: since models seem to be regurgitating aspects of their training set, distillation can therefore be thought of as indirectly training on the model's original dataset.

Many fields collect large-scale temporal data through repeated measurements (`trials’), where each trial is labeled with a set of metadata variables spanning several categories. For example, a trial in a neuroscience study may be linked to a value from category (a): task difficulty, and category (b): animal choice. A critical challenge in time-series analysis is to understand how these labels are encoded within the multi-trial observations, and disentangle the distinct effect of each label entry across categories. Here, we present MILCCI, a novel data-driven method that i) identifies the interpretable components underlying the data, ii) captures cross-trial variability, and iii) integrates label information to understand each category's representation within the data. MILCCI extends a sparse per-trial decomposition that leverages label similarities within each category to enable subtle, label-driven cross-trial adjustments in component compositions and to distinguish the contribution of each category. MILCCI also learns each component’s corresponding temporal trace, which evolves over time within each trial and varies flexibly across trials. We demonstrate MILCCI’s performance through both synthetic and real-world examples, including voting patterns, online page view trends, and neuronal recordings.

Symbolic expressions can effectively characterize and predict circuit behavior, but deriving them directly from circuit schematics is challenging. This process requires accurate visual-to-symbolic construction of circuit structure from images and correct multi-step symbolic derivation, both of which impose strict correctness requirements. This work proposes AutoVSR, an automated framework for visual-to-symbolic generation of circuit expressions using Vision Language Models (VLMs). By reconstructing circuit diagrams into an executable intermediate representation (Executable IR) and leveraging a symbolic solver for reasoning, AutoVSR significantly improves the accuracy of symbolic expression generation. AutoVSR introduces two key innovations: an IR construction method guided by component rule retrieval and verification-based feedback, and a symbolic solver implemented as a planning agent equipped with a symbolic tool library for reliable multi-step derivation. Compared with end-to-end VLM approaches and specialized methods on the main symbolic expression generation task, AutoVSR achieves accuracy improvements of 30.01--59.45% and 41.96--51.84%, respectively. Moreover, AutoVSR surpasses closed-source state-of-the-art VLMs in inference cost and computational efficiency. Code is available at https://github.com/LongfeiLi1/AutoVSR.

We introduce TFRBench, the first benchmark designed to evaluate the reasoning capabilities of forecasting systems. Traditionally, time-series forecasting has been evaluated solely on numerical accuracy, treating foundation models as "black boxes." Unlike existing benchmarks, TFRBench provides a protocol for evaluating the reasoning generated by forecasting systems--specifically their analysis of cross-channel dependencies, trends, and external events. To enable this, we propose a systematic multi-agent framework that utilizes an iterative verification loop to synthesize numerically grounded reasoning traces. Spanning ten datasets across five domains, our evaluation confirms that this reasoning is causally effective; useful for evaluation; and prompting LLMs with our generated traces significantly improves forecasting accuracy compared to direct numerical prediction (e.g., avg. $\sim40.2$% $\rightarrow$ $\sim56.6$%), validating the quality of our reasoning. Conversely, benchmarking experiments reveal that off-the-shelf LLMs consistently struggle with both reasoning (lower LLM-as-a-Judge scores) and numerical forecasting, frequently failing to capture domain-specific dynamics. TFRBench thus establishes a new standard for interpretable, reasoning-based evaluation in time-series forecasting. Our benchmark is available at: https://tfrbench.github.io

We propose a system for marking sensitive or copyrighted texts to detect their use in fine-tuning large language models under black-box access with statistical guarantees. Our method builds digital “marks” using invisible Unicode characters organized into (“cue”, “reply”) pairs. During an audit, prompts containing only “cue” fragments are issued to trigger regurgitation of the corresponding “reply”, indicating document usage. To control false positives, we compare against held-out counterfactual marks and apply a ranking test, yielding a verifiable bound on the false positive rate. Empirically, we obtain a true positive rate of 96.7\% at 0\% false positive rate and reply regurgitation rates exceeding 28\% per document with only 40 (4\%) watermarked documents. The approach is minimally invasive, scalable across many sources, robust to standard processing pipelines, and achieves high detection power even when marked data is a small fraction of the fine-tuning corpus.

Memory-based test-time adaptation (TTA) assigns streaming test samples into class-specific memory slots based on pseudo-labels predicted by models like CLIP, and retrieves them to facilitate subsequent predictions under distribution shift. However, this process introduces two challenges: (1) Each sample is hard-assigned to a single class based on CLIP's prediction, where inaccurate CLIP prediction leads to memory contamination that biases subsequent prediction. (2) Samples are evicted under biased selection due to fixed memory capacity, which risks discarding informative samples and undermining the efficacy of the memory. To address these challenges, we propose A²Memory (Attributive-Associative Memory for Test-time Adaptation). For challenge (1), we propose Attribute-centric Memory Construction that builds prior textual representations from class-shared representative and diverse visual attributes, and applies soft assignment to generate surrogate visual representations. For challenge (2), we design Class-wise Associative Memory that dynamically compresses streaming samples into fixed-capacity memory through gradient-based optimization and data-dependent retention, then retrieves sample-adaptive class prototypes for reliable inference. Extensive experiments demonstrate consistent improvements over state-of-the-art methods across 15 benchmarks.

Understanding how and why large language models (LLMs) fail is becoming a central challenge as models rapidly evolve and static evaluations fall behind. While automated probing has been enabled by dynamic test generation, existing approaches often discover isolated failure cases, lack principled control over exploration, and provide limited insight into the underlying structure of model weaknesses. We propose ProbeLLM, a benchmark-agnostic automated probing framework that elevates weakness discovery from individual failures to structured failure modes. ProbeLLM formulates probing as a hierarchical Monte Carlo Tree Search, explicitly allocating limited probing budgets between global exploration of new failure regions and local refinement of recurring error patterns. By restricting probing to verifiable test cases and leveraging tool-augmented generation and verification, ProbeLLM grounds failure discovery in reliable evidence. Discovered failures are further consolidated into interpretable failure modes via failure-aware embeddings and boundary-aware induction. Across diverse benchmarks and LLMs, ProbeLLM reveals substantially broader, cleaner, and more fine-grained failure landscapes than static benchmarks and prior automated methods, supporting a shift from case-centric evaluation toward principled weakness discovery

Vision-language models (VLMs) generate fluent causal explanations, but current evaluations cannot distinguish linguistic plausibility from faithful causal reasoning. We introduce a dual-probe methodology that isolates these properties. The Text-Only Probe measures linguistic quality. The Chain-Text Probe requires models to first generate explicit causal chains. The Abstraction Gap (AG) metric quantifies the normalized performance difference. Evaluating eight VLMs on CAGE (Causal Abstraction Gap Evaluation), a benchmark of 49,500 questions across 5,500 images spanning Pearl's causal hierarchy, we find seven models exhibit AG exceeding 0.50 with text scores of 6--8 but chain scores below 2.5. Fine-tuning on 45,000 chain-annotated examples fails to close the gap. However, one model achieves near-zero AG. The capability exists within current VLM architectures and depends on pretraining and architectural choices. CAGE provides a diagnostic tool for assessing faithful causal reasoning in VLMs.

Large Reasoning Models (LRMs) have recently achieved strong mathematical and code reasoning performance through Reinforcement Learning (RL) post-training. However, we show that modern reasoning post-training induces an unintended exploration collapse: temperature-based sampling no longer increases pass@$n$ accuracy. Empirically, the final-layer posterior of post-trained LRMs exhibit sharply reduced entropy, while the entropy of intermediate layers remains relatively high. Motivated by this entropy asymmetry, we propose Latent Exploration Decoding (LED), a depth-conditioned decoding strategy. LED aggregates intermediate posteriors via cumulative sum and selects depth configurations with maximal entropy as exploration candidates. Without additional training or parameters, LED consistently improves pass@1 and pass@16 accuracy by 0.61 and 1.03 percentage points across multiple reasoning benchmarks and models. Relevant code is included in the supplementary material and will made be fully public after this paper is accepted.

State-of-the-art reasoning models utilize long chain-of-thought (CoT) to solve increasingly complex problems using more test-time computation. In this work, we explore a long CoT setting where the model makes up to K successive attempts at solving a problem, in which each attempt is allowed to build on earlier ones after the model receives a hard verifier feedback. This motivates RL methods that can harness per-attempt rewards by carefully weighting individual attempts. We study optimizing the Verification@K reward (the model succeeds by the K-th attempt) and show that naively weighing the attempts by their pass/fail results in biased gradients. We introduce Calibrated Attempt-Level (CAL) GRPO by devising a weighing strategy to obtain unbiased gradients while maintaining small variance. Our theory reveals how incorporating per-attempt rewards influences the training and the eventual Verification@K performance. Experiments, baselines, and ablations on synthetic and real data corroborate our theory and the benefits of CAL-GRPO over vanilla GRPO as well as naive weighting.

While the ecosystem of Lean and Mathlib has enjoyed celebrated success in formal mathematical reasoning with the help of large language models (LLMs), the absence of many folklore lemmas in Mathlib remains a persistent barrier that limits Lean's usability as an everyday tool for mathematicians like LaTeX or Maple. To address this, we introduce MathlibLemma, a modular LLM-based pipeline for automated folklore-lemma mining: the discovery, formalization, and proving of reusable intermediate facts that mathematicians often take for granted but that are not always present in formal libraries. At its core, MathlibLemma proactively mines the missing connective tissue of mathematics. The pipeline produces a verified library of folklore-style lemmas, including 1,506 Lean-checked proofs that pass a proof-bypass screen; a small curated pilot subset has also been merged into Mathlib, providing external evidence that selected outputs can meet expert library standards. Leveraging this pipeline, we further construct the MathlibLemma benchmark, a suite of 4,028 non-trivial type-checked Lean statements spanning a broad range of mathematical domains. By transforming the role of LLMs from passive consumers to active contributors, this work takes a step toward AI-assisted expansion of formal mathematical libraries.

Recent research has made growing efforts to leverage large language models (LLMs) for computer-aided design (CAD), a domain that demands advanced geometric and spatial reasoning across long operation sequence. However, existing studies remain limited in addressing complex modeling tasks that necessitate step-by-step reasoning, primarily due to the scarcity of high-quality CAD datasets and the absence of fine-grained evaluation frameworks. In response to these challenges, we introduce Op-CAD, the first large-scale, multi-modal dataset for operation-oriented CAD generation, encompassing four operation types and five modalities. Furthermore, we introduce a novel CAD parsing module together with a geometry-guided hierarchical annotation pipeline, which decomposes modeling sequences into discrete operations and substantially improves the annotation accuracy of Vision-Language Models (VLMs). Based on our dataset, we redefine the CAD modeling task by decoupling geometric and spatial perspectives and introduce a novel metric, Chamfer/Fillet Intersection over Union (CF-IoU), to fill the void in assessing chamfer and fillet operations. By comprehensively evaluating eight LLMs on Op-CAD, we establish a benchmark for current models on operation-oriented tasks. Finally, we investigate performance enhancement strategies through fine-tuning on Op-CAD and propose Chain-of-Operation (COOP), a novel prompting strategy that emulates human-engineer reasoning.

Aligning large language models (LLMs) to serve users with heterogeneous and potentially conflicting preferences is a central challenge for personalized and trustworthy AI. We formalize an ideal notion of *universal alignment* through *test-time scaling*: for each prompt, the model produces $k\ge 1$ candidate responses and a user selects their preferred one. We introduce *$(k,f(k))$-robust alignment*, which requires the $k$-output model to have win rate $f(k)$ against any other single-output model, and *asymptotic universal alignment (U-alignment)*, which requires $f(k)\to 1$ as $k\to\infty$. Our main result characterizes the optimal convergence rate: there exists a family of *single-output* policies whose $k$-sample product policies achieve U-alignment at rate $f(k)=\frac{k}{k+1}$, and no method can achieve a faster rate in general. We show that popular post-training methods, including Nash learning from human feedback (NLHF), can fundamentally underutilize the benefits of test-time scaling. Even though NLHF is optimal for $k=1$, sampling from the resulting (often deterministic) policy cannot guarantee win rates above $\tfrac{1}{2}$ except for an arbitrarily small slack. This stems from a lack of output diversity: existing alignment methods can collapse to a single majority-preferred response, making additional samples redundant. In contrast, our approach preserves output diversity and achieves the optimal test-time scaling rate. In particular, we propose a family of symmetric *multi-player alignment games* and prove that any symmetric Nash equilibrium policy of the $(k+1)$-player alignment game achieves the optimal $(k,\frac{k}{k+1})$-robust alignment. Finally, we provide theoretical convergence guarantees for self-play learning dynamics in these games and extend the framework to opponents that also generate multiple responses.

Test-time adaptation (TTA) has emerged as a popular paradigm for improving the performance of vision–language models (*e.g.*, CLIP) on downstream tasks. Among existing CLIP-based TTA methods, Test-Time Prompt Tuning (TPT) is a pioneering work that optimizes textual prompts using multiple test-time augmentations and remains a strong baseline to date. In this work, we revisit TPT and reveal that its optimization can be interpreted as implicitly learning from self-generated pseudo labels. Building on this perspective, we propose a unified self-ensembling framework (**USE**) that ensures consistency between the optimization and inference stages. During optimization, we introduce a simple yet effective self-ensembling (**SE**) strategy that emphasizes the test image itself over its augmented views adaptively to obtain more reliable pseudo labels. To fully exploit the potential of augmentations, we further apply the same strategy at inference time, unifying the objectives of both stages. Notably, **SE** can also act as a lightweight optimization-free TTA method. Extensive experiments across multiple datasets demonstrate that **SE** and **USE** outperform their counterparts, respectively. Furthermore, **SE** yields consistent performance gains when integrated with existing TTA methods. The code is available at *https://github.com/sirujiang/USE*.

What do different *thinking* language models learn during training? We introduce *constructive model diffing*, a framework for understanding fine-tuned models by constructing the base-to-fine-tuned difference from interpretable components, producing hybrid models whose performance recovery measures how well the components capture the diff. For thinking models, we decompose the diff into *reasoning mechanisms* (category-specific vectors that induce reasoning behaviors in the base model) and *reasoning heuristics* (a classifier determining when each mechanism fires). To ground this decomposition, we use Sparse Autoencoders to discover interpretable taxonomies of reasoning behaviors. Across nine model configurations (four RL-trained, four SFT-distilled from a larger thinking model, and one mixed), we find a striking difference: hybrid models recover much more performance for RL-trained models than for SFT-trained or mixed ones. This indicates that RL primarily teaches models to apply sophisticated heuristics to pre-existing base capabilities, while SFT-based distillation teaches the base model new mechanisms. These results offer a new lens on what different training paradigms teach, with implications for efficient reasoning model development.

Understanding physical transformations is fundamental for reasoning in dynamic environments. While Vision Language Models (VLMs) show promise in embodied applications, whether they genuinely understand physical transformations remains unclear. We introduce ***ConservationBench*** evaluating ***conservation***—whether physical quantities remain invariant under transformations. Spanning four properties with paired conserving/non-conserving scenarios, we generate 23,040 questions across 112 VLMs. Results reveal systematic failure: performance remains near chance with improvements on conservation tasks accompanied by drops on controls. Control experiments show strong textual priors favoring invariance, yet models perform worse with visual content. Neither temporal resolution, prompting, nor curated sampling helps. These findings show that current VLMs fail to maintain transformation-invariant representations of physical properties across dynamic scenes.