ICML 2026 论文列表

Deploying guardrails for custom policies remains challenging, as generic safety models fail to capture task-specific requirements, while prompting LLMs suffers from inconsistent boundary-case performance and high inference costs. Training custom classifiers achieves both accuracy and efficiency, yet demands substantial labeled data that is costly to obtain. We present BARRED (Boundary Alignment Refinement through REflection and Debate), a framework for generating faithful and diverse synthetic training data using only a task description and a small set of unlabeled examples. Our approach decomposes the domain space into dimensions to ensure comprehensive coverage, and employs multi-agent debate to verify label correctness, yielding a high-fidelity training corpus. Experiments across diverse custom policies demonstrate that small language models finetuned on our synthetic data consistently outperform state-of-the-art proprietary LLMs (including reasoning models) and dedicated guardrail models. Ablation studies confirm that both dimension decomposition and debate-based verification are critical for ensuring the diversity and label fidelity required for effective fine-tuning. The BARRED framework eliminates the reliance on extensive human annotation, offering a scalable solution for accurate custom guardrails.

Vision-Language-Action (VLA) models benefit from Chain-of-Thought (CoT) reasoning, but existing approaches incur high inference overhead and rely on discrete reasoning representations that mismatch continuous perception and control. We propose Latent Reasoning VLA (LaRA-VLA), a unified VLA framework that internalizes multi-modal CoT reasoning into continuous latent representations for embodied action. LaRA-VLA performs unified reasoning and prediction in latent space, eliminating explicit CoT generation at inference time and enabling efficient, action-oriented control. To realize latent embodied reasoning, we introduce a curriculum-based training paradigm that progressively transitions from explicit textual and visual CoT supervision to latent reasoning, and finally adapts latent reasoning dynamics to condition action generation. We construct two structured CoT datasets, LIBERO-LaRA and Bridge-LaRA, and evaluate LaRA-VLA across simulation benchmarks and long-horizon real-robot manipulation tasks. Experimental results show that LaRA-VLA outperforms existing state-of-the-art VLA methods while achieving up to a 90\% reduction in inference latency compared to explicit CoT-based VLA approaches, highlighting latent reasoning as an effective and efficient paradigm for real-time embodied control.

Recent advances in large language models (LLMs) have demonstrated that reinforcement fine-tuning of pretrained base models can lead to significant gains in reasoning performance at inference time. In this work, we theoretically analyze why reinforcement fine-tuning induces better reasoning ability than purely supervised fine-tuning (SFT) methods. We model chain-of-thought (CoT) reasoning as a pathfinding problem on graphs and compare the popular method of reinforcement learning with verifiable rewards (RLVR) against traditional SFT. We prove that SFT, when trained on golden shortest paths without negative examples, fails to learn how to efficiently backtrack. In contrast, an RLVR-trained model can learn how to efficiently backtrack from dead ends using only outcome reward. This leads to a provable inference-time compute separation between the two methods, and demonstrates that RLVR allows the model to learn the location of difficult decisions in a reasoning chain, ultimately allowing for better allocation of inference-time compute. Finally, we show that the reasoning traces of an RLVR model can be distilled to train a base model to backtrack efficiently as well.

Out-of-distribution (OOD) detection requires accurately classifying in-distribution (ID) samples while effectively distinguishing anomalous OOD data. However, existing methodologies predominantly rely on real-valued magnitude features, neglecting the semantic richness embedded in phase information, and often lack a systematic theoretical framework for quantitively modeling uncertainty. To address this dual limitation of incomplete feature representation and insufficient uncertainty modeling, the trustworthy quantum evidence neural network (TrustworthyQENN) is proposed, a novel quantum-inspired framework bridging complex-valued representation learning with generalized quantum evidence theory (GQET). Specifically, supervised complex-valued contrastive learning (SCVCL) is proposed to synchronize amplitude distributions with phase correlations, thereby enforcing high intra-class compactness and inter-class separability for ID data. A quantum evidence generation mechanism based on GQET is subsequently devised, where the OOD state is formally grounded as the quantum empty set within a Hilbert space. Furthermore, the generalized quantum evidential combination rule (GQECR) is leveraged to fuse multi-view evidence, thereby achieving trustworthy inference. Extensive experiments on the MSTAR, EuroSAT, and FUSAR-Ship benchmarks substantiate the superiority of TrustworthyQENN, which achieves a peak AUROC of 95.94\% on the MSTAR dataset while consistently outperforming state-of-the-art methods across all evaluated scenarios.

Robust conformal prediction (RCP) extends conformal prediction (CP) to noisy inputs by producing prediction sets with guaranteed coverage, ensuring that the true label is contained in the set with a user-specified probability even under worst-case perturbations. Recent works use randomized smoothing, as it provides robustness for black-box models at larger radii. Currently, there exist two setups for smoothing-based RCP: one requires extensive Monte Carlo sampling at calibration and test time but results in smaller prediction sets; the other setup produces larger prediction sets but uses a single sample at both stages. In deployment, calibration—as a one-time pre-processing step—can accommodate substantially higher computational overhead than inference. Inspired by this observation, we introduce an RCP framework that strikes a balance between the two extremes of this trade-off: we increase the sample rate at calibration time while keeping it either one or very low during test time. This calibration-time sampling opens the possibility of reducing the size of the prediction sets. In production, where the number of test predictions typically far exceeds the size of the calibration set, our Front-Loaded RCP matches the computational complexity of the state of the art while producing considerably smaller prediction sets at larger radii.

Existing expressivity results for transformers typically rely on hardmax attention, high precision, and other architectural modifications that disconnect them from the models used in practice. We bridge this gap by analyzing standard transformer decoders with softmax attention and rounding of activations and attention weights, while allowing depth and width to grow logarithmically with the context length. As an intermediate step, we construct hardmax transformers with ternary activations and well-separated attention scores that simulate Turing machines using Chain-of-Thought (CoT). This lets us convert the constructions to equivalent softmax transformers without the unrealistic parameter magnitudes or activation precision that prior approaches would require. Using the same technique, we analyze a recently proposed summarized CoT paradigm and show that it simulates Turing machines more efficiently, with model size scaling logarithmically in a space bound rather than a time bound. We empirically test predictions made by our results on a Sudoku reasoning task and find better alignment with learnability than for prior high-precision results. Our code is available at https://github.com/moritzbroe/transformer-expressivity.

Tensor Product Representations provide the structural fidelity required for symbolic reasoning in models but suffer from *exponential* dimensionality growth when encoding deep recursive structures. Conversely, Vector Symbolic Architectures maintain *constant* dimensionality but sacrifice capacity and fidelity due to noisy compression via superposition. In this work, we propose **Orthogonal Subspace Carving (OSC)**, a memory architecture that binds *fillers* to *roles* by projecting onto the null space of the role basis before aggregating into a fixed order-$p$ tensor. OSC uses projections to enforce geometric orthogonality between bound structures within a {\em static} memory trace. We show that this mechanism decouples the tensor order from the structural depth, enabling deep recursive binding within a *constant* memory footprint. This construction allows for component vectors that are *orders of magnitude* smaller than the memory tensor, giving excellent memory efficiency in settings involving high superposition. We also show that TPR is a special case of binding in Clifford algebra, and give a Clifford formulation of OSC.

Medical audio data is difficult to collect due to privacy regulations and high annotation costs arising from domain expertise. Thus, existing benchmarks tend to underrepresent complex medical audio scenarios. To address this challenge, we present MedMosaic, a medical audio question–answering dataset designed to benchmark language and audio reasoning models under realistic clinical constraints. MedMosaic features a diverse range of medical audio types, including condition-related physiological sounds, carefully constructed synthetic voices to mimic speech with artifacts as well as real short and long length clinical conversations to model varying context lengths. The dataset also features a total of 46,701 question-answer pairs, spanning categories such as multiple-choice, sequential multi-turn, and open-ended question–answers, enabling systematic evaluation of multi-hop reasoning and answer generation capabilities. Benchmarking 13 audio and multimodal reasoning models reveals that reasoning remains challenging for all evaluated systems, with substantial performance variation across question types. In particular, even state-of-the-art model like Gemini-2.5-pro can only achieve 68.1\% accuracy approximately. These findings underscore persistent limitations in medical reasoning and highlight the need for more robust, domain-specific multimodal reasoning models. A sample of benchmark data is available here:https://shorturl.at/Lyp33

Automatic loop invariant generation remains a challenging problem in program verification, particularly for memory-manipulating programs where shape invariants are required to characterize heap-allocated structures and memory layouts. While existing approaches succeed on numerical invariants, they achieve limited accuracy on shape invariants. We hypothesize that this stems from the need to reason about memory state evolution—information that remains implicit in source code. To address this, we ground LLM reasoning in symbolic execution traces that explicitly capture such transitions. We propose \textsc{Malice}, a two-stage framework incorporating these traces: (1) guided multi-step reasoning that predicts invariants via chain-of-thought analysis of traces, and (2) agentic iterative refinement that corrects candidates through verification tool feedback. Evaluated on LIG-MM+, a benchmark featuring common operations on typical memory structures, \textsc{Malice} substantially outperforms existing approaches.

Chain-of-thought (CoT) reasoning has proven effective for enhancing problem-solving in large language models. However, when applied to multimodal LLMs (MLLMs), existing CoT approaches suffer from a fundamental limitation: they perform reasoning entirely in text without accessing visual features during the reasoning process. After initial visual encoding, image information becomes inaccessible, forcing models to reason based solely on whatever was captured in the initial description, which forms a `vision-blind reasoning' paradigm that limits fine-grained visual extraction, error verification, and adaptive attention. We propose Text-Visual Interleaved Chain-of-Thought (TVI-CoT), a framework that enables explicit interleaving of textual reasoning and visual feature access through learnable control tokens ($\langle\text{Think}\rangle$, $\langle\text{Look}\rangle$, $\langle\text{Answer}\rangle$). These tokens allow dynamic switching between reasoning and visual grounding, attending to relevant image regions conditioned on the evolving reasoning state. Experiments on eight benchmarks demonstrate state-of-the-art results among MLLM-based CoT methods and notable performance boost compared to the baseline: +6.1\% on MMMU, +3.8\% on MathVerse, +3.4\% on MathVista, and +3.4\% on ScienceQA. Code is available at https://github.com/hulianyuyy/TVI-CoT.

Time series is a pervasive data type across various application domains, rendering the reasonable solving of diverse time series tasks a long-standing goal. Recent advances in large language models (LLMs), especially their reasoning abilities unlocked through reinforcement learning (RL), have opened new opportunities for tackling tasks with long Chain-of-Thought (CoT) reasoning. However, leveraging LLM reasoning for time series remains infancy, hindered by the absence of carefully curated time series CoT data for training, limited data efficiency caused by underexplored data scheduling, and the lack of RL algorithms tailored for exploiting such time series CoT data. In this paper, we introduce VeriTime, a framework that tailors LLMs for time series reasoning through data synthesis, data scheduling, and RL training. First, we propose a data synthesis pipeline that constructs a time series–text multimodal dataset with process-verifiable annotations. Second, we design a data scheduling mechanism that arranges training samples according to a principled hierarchy of difficulty and task taxonomy. Third, we develop a two-stage reinforcement finetuning featuring fine-grained, multi-objective rewards that leverages verifiable process-level CoT data. Extensive experiments show that VeriTime substantially boosts LLM performance across diverse time series reasoning tasks. Notably, it enables compact 3B–4B models to achieve reasoning capabilities on par with or exceeding those of larger proprietary LLMs.

Large language models have achieved remarkable success on final-answer mathematical problems, largely due to the ease of applying reinforcement learning with verifiable rewards. However, the reasoning underlying these solutions is often flawed. Advancing to rigorous proof-based mathematics requires reliable proof verification capabilities. We begin by analyzing multiple evaluation setups and show that focusing on a single benchmark can lead to brittle or misleading conclusions. To address this, we evaluate both proof-based and final-answer reasoning to obtain a more reliable measure of model performance. We then scale two major generative verification methods (GenSelect and LLM-as-a-Judge) to millions of tokens and identify their combination as the most effective framework for solution verification and selection. We further show that the choice of prompt for LLM-as-a-Judge significantly affects the model's performance, but reinforcement learning can reduce this sensitivity. However, despite improving proof-level metrics, reinforcement learning does not enhance final-answer precision, indicating that current models often reward stylistic or procedural correctness rather than mathematical validity. Our results establish practical guidelines for designing and evaluating scalable proof-verification and selection systems.

Large language models (LLMs) have demonstrated significant potential in formal theorem proving, yet state-of-the-art performance often necessitates prohibitive test-time compute via massive roll-outs or extended context windows. In this work, we address this scalability bottleneck by exploiting an informative structure in formal verification: the observation that compilers map a vast space of diverse proof attempts to a compact set of structured failure modes. We introduce a learning-to-refine framework that leverages this compression to perform efficient learning and proof exploration. We perform tree search that corrects errors locally conditioned on explicit verifier feedback, thereby circumventing the costs associated with accumulating a long history of proof attempts. Extensive evaluations show that our method consistently amplifies the reasoning capabilities of base provers across varying scales. Notably, our approach achieves state-of-the-art performance on PutnamBench among publicly reported $\sim$8B and $\sim$32B parameter models under comparable test-time budgets, offering a scalable paradigm for next-generation verifier-guided reasoning.

Recent studies indicate that when faced with explicit biases in prompts, models often omit mentioning these biases in their Chain-of-Thought (CoT) output, revealing that verbalized reasoning can give an incorrect picture of how models arrive at conclusions (unfaithfulness). In this work, we show that unfaithful CoT also occurs on naturally worded, non-adversarial prompts without adding artificial biases or editing model outputs. We find that when separately presented with the questions "Is X bigger than Y?" and "Is Y bigger than X?", models sometimes produce superficially coherent arguments to justify systematically answering Yes to both or No to both, despite the contradiction. We present preliminary evidence that this is due to models' implicit biases towards Yes or No, labeling this *Implicit Post-Hoc Rationalization*. Our results reveal rates up to 13% for production models, and while frontier models are more faithful, none are entirely so, including thinking models like DeepSeek R1 (0.37%) and Sonnet 3.7 with thinking (0.04%). We also investigate *Unfaithful Illogical Shortcuts*, where models use subtly illogical reasoning to make speculative answers to hard math problems seem rigorously proven. Our findings indicate that while CoT can be useful for assessing outputs, it is not a complete account of the internal process that produced the model's answer and should be used with caution in agentic or safety-critical settings.

Neural networks adapt through first-order parameter updates, yet it remains unclear whether such updates preserve logical coherence. We investigate the geometric limits of the Linear Propagation Assumption (LPA), the premise that local updates coherently propagate to logical consequences. To formalize this, we adopt relation algebra and study three core operations on relations: negation flips truth values, converse swaps argument order, and composition chains relations. For negation and converse, we prove that guaranteeing direction-agnostic first-order propagation necessitates a tensor factorization separating entity-pair context from relation content. However, for composition, we identify a fundamental obstruction. We show that composition reduces to conjunction, and prove that any conjunction well-defined on linear features must be bilinear. Since bilinearity is incompatible with negation, this forces the feature map to collapse. These results suggest that failures in knowledge editing, the reversal curse, and multi-hop reasoning may stem from common structural limitations inherent to the LPA.

Visual causal reasoning is essential for understanding and intervening in the physical world, requiring identification of causal variables from visual inputs and reasoning over intervention effects. Despite recent progress, large vision-language models (VLMs) remain brittle at such tasks, especially for interventional and counterfactual queries over multi-image inputs. Most existing explorations inject causal knowledge via textual prompts, leaving causal mechanisms external to model execution and limiting reliable control during inference. To address this problem, we propose BridgeVLM, which internalizes visual causal reasoning by inducing a causal graph from multi-image inputs and converting it into structured Causal Tokens executed by RAMP layers injected into the LLM decoder for causal message passing. We further introduce a unified training interface M3S for fine-grained causal supervision from different granularities (local/global level). BridgeVLM achieves 54.4\% accuracy on intervention tasks on CausalVLBench (vs. 33.2\% with prompt-level supervision), improves results on Causal3D from 43.6\% to 49.0\%, and substantially improves causal structure learning on CausalVLBench ($F_1$: 33.4\% $\rightarrow$ 75.1\%).

Evaluating open-ended outputs from large language models (LLMs) remains challenging due to the absence of ground truth. Existing metrics rely on final-answer accuracy or surface-level statistics, leaving the reasoning process itself unexamined. We introduce TRACE (Toulmin-based Reasoning Assessment through Constructive Elements), a metric that analyzes Chain-of-Thought (CoT) reasoning processes. Rather than judging outcomes, TRACE inspects how arguments are constructed by integrating Toulmin's argumentation theory with Flavell's metacognitive framework to assess reasoning structure. Experiments on 26.3K QA samples across 7 reasoning models show strong correlation with benchmark accuracy (r=0.74). Furthermore, TRACE is effective as a reinforcement learning reward signal, outperforming accuracy-only baselines. Together, these results indicate that logically sound reasoning leads to higher-quality answers. TRACE thus serves as a complementary metric for evaluating open-ended outputs. Code is available at https://github.com/hyyangkisti/trace.

Self-reflection is a powerful mechanism for credit assignment in human learning, converting sparse outcome feedback into actionable guidance. However, its potential for post-training Large Language Models (LLMs) remains underexplored. We propose Self-Reflective Policy Optimization (SRPO), a framework that internalizes this capability. SRPO enables LLMs to analyze their own completed trajectories, synthesize errors into concise "reflection patches," and use these reflection-conditioned rollouts as high-quality, on-policy distillation targets. This process effectively transforms sparse terminal supervision into dense, token-level learning signals without requiring external critics, separate reward models, or larger teacher models. We demonstrate that SRPO achieves state-of-the-art performance across mathematical reasoning and long-horizon agentic benchmarks with exceptional data efficiency. Using a Qwen3-8B base model, SRPO attains 73.3\% on AIME’24 using only 8\% (0.08$\times$) of the training FLOPs required by scaled supervised fine-tuning, while significantly improving success rates on WebShop (64.7\%), ALFWorld (76.8\%), and SWE-Bench-Lite (31.2\%).

Large Reasoning Models (LRMs) solve complex tasks by generating long Chain-of-Thought (CoT) sequences; however, the emergent dynamics governing reasoning trajectories are not well understood and can lead to inconsistencies and reasoning pathologies. In this work, we propose to approximate LRM's emerging hierarchical reasoning dynamics as a trajectory within a Finite State Machine (FSM) transitioning among six abstract cognitive states. We demonstrate that these states and transitions can be captured in the latent state of the model. We believe that this representation can have different applications in the interpretability and optimization of LRM models. For example, by analyzing the topology of these transitions, we identify statistical shifts in reasoning strategies that help identify effective reasoning chains from those that fail. To illustrate these potential advantages, we propose $Q$-Value guided steering, a training-free inference-time control method that treats reasoning as a planning problem. We estimate the long-horizon utility of state transitions and apply sparse, orthogonal activation steering at sentence boundaries to align the CoT generation with optimal reasoning policies. Experiments across four benchmarks (AIME25, MATH-500, GSM8k, and GPQA Diamond) using three state-of-the-art open reasoning models demonstrate that $Q$-Value steering policy achieves significant performance gains with "surgical'' efficiency, often requiring $25\times$ fewer interventions than greedy and weighted baselines, which suggests that reasoning can be effectively controlled by guiding high-level cognitive dynamics rather than micro-managing token generation. Code is available at: https://github.com/shahariar-shibli/CoT-FSM.