ICML 2026 论文列表

Despite the remarkable practical success of transformer-based language models, recent work has raised concerns about their ability to perform state tracking. In particular, a growing body of literature has shown this limitation primarily through failures in out-of-distribution (OOD) generalization, such as length extrapolation. In this work, we shift attention to the in-distribution implications of these limitations. We conduct a large-scale experimental study of the data efficiency of transformers and recurrent neural networks (RNNs) across multiple supervision regimes. We find that the amount of training data required by transformers grows much more rapidly with state-space size and sequence length than for RNNs. Furthermore, we analyze the extent to which learned state-tracking mechanisms are shared across different sequence lengths. We show that transformers exhibit negligible or even detrimental weight sharing across lengths, indicating that they learn length-specific solutions in isolation. In contrast, recurrent models exhibit effective amortized learning by sharing weights across lengths, allowing data from one sequence length to improve performance on others. Together, these results demonstrate that state tracking remains a fundamental challenge for transformers, even when training and evaluation distributions match.

While humans develop core visual skills long before acquiring language, contemporary Multimodal LLMs (MLLMs) still rely heavily on linguistic priors to compensate for their fragile visual understanding. We uncovered a crucial fact: state-of-the-art MLLMs consistently fail on basic visual tasks that humans, even 3-year-olds, can solve effortlessly. To systematically investigate this gap, we introduce BabyVision, a benchmark designed to assess core visual abilities independent of linguistic knowledge for MLLMs. BabyVision spans a wide range of tasks, with 388 items divided into 22 subclasses across four key categories. Empirical results and human evaluation reveal that leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores 49.7, falling well behind the average adult score of 94.1. These results show despite excelling in knowledge-heavy evaluations, current MLLMs still lack fundamental visual primitives. Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. We also explore solving visual reasoning with generation models by proposing BabyVision-Gen and automatic evaluation toolkit. Code and data are released at https://github.com/UniPat-AI/BabyVision.

Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are _sufficient_, i.e., if they contain enough information to explain the model’s output-generating process. We generalize classical sufficiency from feature attributions to arbitrary explanations and prove that explanation sufficiency can change depending on the input distribution, which must be explicitly defined for LLM explanations. We propose using the LLM itself to generate alternative inputs conditioned on an explanation, capturing its beliefs about possible inputs. We formalize _self-consistent sufficiency_ as a goal for free-text explanations and introduce an information-theoretic metric, SCSuff, that enables evaluation of free-text explanations without relying on predefined biases or shortcuts. Our experiments show that SCSuff agrees with targeted perturbation tests where applicable and demonstrate that explanation sufficiency can vary with the input distribution. We find LLM explanations are generally insufficient and weakly correlated with model size, accuracy, or output entropy. Analysis of final-token hidden states shows that top and bottom SCSuff scores can be predicted from internal representations, suggesting that SCSuff can guide detection and improvement of sufficient LLM explanations. The code for this paper is available at https://github.com/rajesh-lab/self-consistent-sufficiency.

Autonomous 3D indoor scene synthesis breaks down in non-convex rooms with tightly coupled spatial constraints. Data-driven generators lack topological priors for long-horizon planning, while iterative agents fragment semantics and become geometrically brittle. We present \textbf{ZoneMaestro}, a unified framework that shifts the paradigm from object-centric synthesis to Zone-Graph Orchestration. By internalizing a novel zone-based logic, ZoneMaestro translates high-level semantic intent into functional zones and topological constraints, enabling robust adaptation to diverse architectural forms. To support this, we construct \textit{Zone-Scene-10K}, a large-scale dataset enriched with explicit Zone-Graph annotations. We further introduce an \textit{Alternating Alignment Strategy} that cycles between reasoning internalization and Zone-Aware Group Relative Policy Optimization (\textit{Z-GRPO}), effectively reconciling the tension between semantic richness and geometric validity without relying on external physics engines. To rigorously evaluate spatial intelligence beyond convex primitives, we formally define the task of \textbf{Intricate Spatial Orchestration} and release SCALE, a stress-test benchmark for irregular indoor scenarios with complex, dense spatial relations. Extensive experiments demonstrate that ZoneMaestro resolves the density-safety dichotomy, significantly outperforming state-of-the-art baselines in both structural coherence and intent adherence.

Large language models (LLMs) often fail in systematic, model-specific ways under meaning-preserving question rewrites (paraphrases, format changes, benign distractors). In this work, we address this instability by identifying where the model's reasoning process diverges across semantically-equivalent inputs. For each target LLM, we sample multiple solution traces under rewrites and aggregate them into a graph of recurring intermediate steps, which pinpoints where incorrect traces diverge from correct ones. We then generate a small set of semantics-preserving examples that mirror the rewrite patterns most responsible for these divergences, and use them to steer the model (\emph{targeted alternation training}), either via fine-tuning or via in-context learning. Across MMLU-Pro, Big-MATH, and DROP, this yields consistent gains and cross-dataset generalization. On Humanity’s Last Exam, using 200 in-context examples, it improves GPT-5.2 (xhigh) from 35.4\% to 38.1\%, demonstrating that targeted alternation training can materially improve a frontier, API-accessible closed model under realistic access constraints.

Compiler auto-tuning faces a dichotomy between traditional black-box search methods, which lack semantic guidance, and recent Large Language Model (LLM) approaches, which often suffer from superficial pattern matching and causal opacity. In this paper, we introduce ECCO, a framework that bridges interpretable reasoning with combinatorial search. We first propose a reverse engineering methodology to construct a Chain-of-Thought dataset, explicitly mapping static code features to verifiable performance evidence. This enables the model to learn the causal logic governing optimization decisions rather than merely imitating sequences. Leveraging this interpretable prior, we design a collaborative inference mechanism where the LLM functions as a strategist, defining optimization intents that dynamically guide the mutation operations of a genetic algorithm. Experimental results on seven datasets demonstrate that ECCO outperforms the LLVM opt -O3 baseline, achieving an average 24.44% reduction in cycles.

As Large Language Models (LLMs) increasingly assist secure software development, their ability to meet the rigorous demands of Rust program verification remains unclear. Existing evaluations treat Rust verification as a black box, assessing models only by binary pass or fail outcomes for proof hints. This obscures whether models can systematically reconstruct the explicit deductive steps required for verifying nontrivial Rust code. To bridge this gap, we introduce VCoT-Lift, a framework that lifts low-level solver reasoning into high-level, human-readable verification steps. By exposing solver-level reasoning as an explicit Verification Chain-of-Thought, VCoT-Lift provides a concrete ground truth for fine-grained evaluation. Leveraging VCoT-Lift, we introduce VCoT-Bench, a comprehensive benchmark of 1,988 VCoT completion tasks for rigorously evaluating LLMs' ability to reconstruct the entire verification process. VCoT-Bench measures performance along three orthogonal dimensions: robustness to varying degrees of missing proofs, competence across different proof types, and sensitivity to proof locations. Evaluation of ten state-of-the-art models reveals severe fragility, indicating that current LLMs fall well short of the reasoning capabilities exhibited by automated theorem provers.

We present Layered Ray Intersections (LaRI), a fully supervised method for occluded geometry reasoning from a single image. Unlike conventional depth estimation, which is limited to visible surfaces, LaRI predicts multiple surfaces intersected by the camera rays using layered point maps. Compared to the existing approaches that leverage neural implicit representations or iterative refinement, LaRI achieves complete scene reconstruction in one feed-forward pass, enabling efficient and view-aligned geometric reasoning to underpin both object-level and scene-level tasks. We further propose to predict the ray stopping index, which identifies valid intersecting pixels and layers from LaRI's output. To better underpin and evaluate this task, we build an annotation pipeline using rendering engines, construct annotations for five public datasets, including synthetic and real-world data covering 3D objects and scenes. As a generic method, LaRI's performance is validated in object-level and scene-level reconstruction tasks.

Token-level credit assignment remains a key obstacle for reinforcement learning (RL) in large language models (LLMs), where RL recipes typically treat all tokens equally, failing to distinguish decisive reasoning steps from routine formatting or fluent filler. Recent attempts leverage model-internal signals to assign finer-grained credit, but these are often point-wise heuristics that ignore the global structure of information propagation. We propose FlowTracer, an RL framework that traces answer-targeted reasoning flow on an attention-induced directed acyclic graph in which nodes correspond to tokens and edge capacities come from aggregated attention weights and derives token credit from this global structure. The edge capacities are reweighted to retain only the influence that can reach the answer region, while enforcing local flow conservation so intermediate tokens neither lose nor gain effective mass due to path length or irrelevant branches. On this graph, FlowTracer extracts an information-flow backbone connecting the question to the answer and scores tokens by flow throughput, revealing high-impact hubs and aggregation checkpoints that mediate long-range dependencies. These derived importances are used to shape token-level rewards, enabling learning signals to focus precisely on the tokens that route information toward (or away from) correct answers and delivering consistent performance gains across a range of reasoning tasks.

Cross-task generalization is a core challenge in open-world robotic manipulation, and the key lies in extracting transferable manipulation knowledge from seen tasks. Recent in-context learning approaches leverage seen task demonstrations to generate actions for unseen tasks without parameter updates. However, existing methods provide only low-level continuous action sequences as context, failing to capture composable skill knowledge and causing models to degenerate into superficial trajectory imitation. We propose Decompose and Recompose, a skill reasoning framework using atomic skill-action pairs as intermediate representations. Our approach decomposes seen demonstrations into interpretable skill--action alignments, enabling the model to recompose these skills for unseen tasks through compositional reasoning. Specifically, we construct a task-adaptive dynamic demonstration library via visual-semantic retrieval combined with skill sequences from a planning agent, complemented by a coverage-aware static library to fill missing skill patterns. Together, these yield skill-comprehensive demonstrations that explicitly elicit compositional reasoning for skill composition and execution ordering. Experiments on the AGNOSTOS benchmark and real-world environments validate our method's zero-shot cross-task generalization capability.

Document visual question answering requires models not only to answer questions correctly, but also to precisely localize answers within complex document layouts. While large vision-language models (VLMs) achieve strong spatial grounding, their inference cost and latency limit real-world deployment; on the other hand, compact VLMs are efficient but suffer substantial localization degradation under standard fine-tuning or distillation. To address this gap, we propose **DocVAL**, a validated chain-of-thought (CoT) distillation framework that transfers explicit spatial reasoning from large teacher models to compact, deployable student VLMs. DocVAL combines **(1)** teacher-generated spatial CoT supervision, **(2)** a rule-based dual-mode validator that filters low-quality training signals and provides fine-grained, pixel-level corrective feedback, and **(3)** a validation-driven two-stage training procedure with iterative refinement. Text detection is used only as training-time scaffolding for supervision and validation, enabling the final student to operate as a pure VLM without OCR or detection at inference. Across multiple document understanding benchmarks, the proposed **DocVAL** yields consistent improvements of up to **6-7 ANLS** points over comparable compact VLMs. We further introduce mean Average Precision (mAP) as a localization metric for document question answering and report strong spatial grounding performance under this new evaluation. We release **95K validator-verified CoT traces** and show that high-quality, validated supervision is more effective than scaling unfiltered data, enabling efficient and trustworthy document grounding. Code/Data: [GitHub](https://github.com/ahmad-shirazi/DocVAL).

While reinforcement learning has achieved impressive progress in language model reasoning, it is constrained by the requirement for verifiable rewards. Recent verifier-free RL methods address this limitation by utilizing the probabilities that LLMs generate reference answers as reward signals. However, these approaches typically sample reasoning traces conditioned only on the question. This design decouples reasoning-trace sampling from answer information, leading to inefficient exploration and incoherence between traces and final answers. In this paper, we propose \textit{\b{Co}upled \b{V}ariational \b{R}einforcement \b{L}earning} (CoVRL), which bridges variational inference and reinforcement learning by coupling prior and posterior distributions through a hybrid sampling strategy. By constructing and optimizing a composite distribution that integrates these two distributions, CoVRL enables efficient exploration while preserving strong thought-answer coherence. Extensive experiments on mathematical and general reasoning benchmarks show that CoVRL improves performance by 12.4\% over the base model and achieves an additional 2.3\% improvement over state-of-the-art verifier-free RL baselines, providing a principled framework for enhancing the general reasoning capabilities of language models.

Understanding what graph neural networks can learn, especially their ability to learn to execute algorithms, remains a central theoretical challenge. In this work, we prove exact learnability results for graph algorithms under bounded-degree and finite-precision constraints. Our approach follows a two-step process. First, we train an ensemble of multi-layer perceptrons (MLPs) to execute the local instructions of a single node. Second, during inference, we use the trained MLP ensemble as the update function within a graph neural network (GNN). Leveraging Neural Tangent Kernel (NTK) theory, we show that local instructions can be learned from a small training set, enabling the complete graph algorithm to be executed during inference without error and with high probability. To illustrate the learning power of our setting, we establish a rigorous learnability result for the LOCAL model of distributed computation. We further demonstrate positive learnability results for widely studied algorithms such as message flooding, breadth-first and depth-first search, and Bellman-Ford.

We study one-sided and $\alpha$-correct sequential hypothesis testing for data generated by an ergodic, finite-state Markov chain. The *null* hypothesis is that the unknown transition matrix belongs to a prescribed set $\cal P$ of stochastic matrices, and the *alternative* corresponds to a disjoint set $\cal Q$. We establish *a non-asymptotic instance-dependent lower bound* on the expected stopping time of any valid sequential test under the alternative, which is asymptotically tight. Our novel analysis improves the existing lower bounds, which are either asymptotic or provably sub-optimal in this setting. Our lower bound incorporates both the stationary distribution and the transition structure induced by the unknown Markov chain. We further propose an optimal test whose expected stopping time matches this lower bound asymptotically as $\alpha \to 0$. We illustrate the usefulness of our framework through applications to sequential detection of model misspecification in Markov Chain Monte Carlo and to testing structural properties, such as the linearity of transition dynamics, in Markov decision processes. Our findings yield a sharp and general characterization of optimal sequential testing procedures under Markovian dependence.

Competitive programming is increasingly being used to evaluate the algorithmic reasoning capabilities of large language models (LLMs). However, existing benchmarks primarily focus on full-information tasks where all problem inputs are provided upfront. This overlooks a critical dimension of algorithmic reasoning: the ability of generated programs to operate when key information is not revealed upfront. *Interactive* problems, a distinctive component of competitive programming, embody this challenge. These problems require programs to engage in multi-round interaction with an interactor (a judge program) under strict protocol constraints and limited query budgets, with new information revealed *only* in response to queries. To address this gap, we introduce *InteractBench*, a benchmark comprising 322 high-quality interactive problems curated from Codeforces, AtCoder, IOI, and ICPC. Each problem is packaged with executable local interactors, enabling fully offline evaluation. Unlike existing benchmarks, InteractBench assesses whether model-generated code can acquire information and track state dynamically. Our evaluation reveals a significant interaction gap: even the most advanced reasoning models achieve limited success on interactive problems. Beyond success rates, we propose a fine-grained failure taxonomy to diagnose the root causes of these deficiencies. Although algorithmic logic errors remain dominant, protocol violations and query-budget overruns are frequent. Code is available at https://github.com/kmsgk0/InteractBench.

Personalizing large language models (LLMs) to individual user preferences is a critical step beyond generating generically helpful responses. However, current personalization methods are ill-suited for new users, as they typically require either slow, resource-intensive fine-tuning or a substantial amount of pre-existing user data, creating a significant cold-start problem. To address this challenge, we introduce a new paradigm for real-time personalization by learning from online pairwise preference feedback collected during text generation. We propose T-POP (Test-Time Personalization with Online Preference Feedback), a novel algorithm that synergistically combines test-time alignment with dueling bandits. Without updating the LLM parameters, T-POP steers the decoding process of a frozen LLM by learning a reward function online that captures user preferences. By leveraging dueling bandits, T-POP intelligently queries the user to efficiently balance between exploring their preferences and exploiting the learned knowledge to generate personalized text. Extensive experiments demonstrate that T-POP achieves rapid and data-efficient personalization, significantly outperforming existing baselines and showing consistent improvement with more user interactions.

As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.

Recent reinforcement-learning frameworks for visual perception policy usually incorporate intermediate reasoning chains expressed in natural language. Empirical observations indicate that such purely linguistic intermediate reasoning often reduces performance on perception tasks. We argue that the core issue lies not in reasoning per se but in the form of reasoning: while these chains perform semantic reasoning in an unstructured linguistic space, visual perception requires reasoning in a spatial and object-centric space. In response, we introduce Artemis, a perception-policy learning method that performs structured visual reasoning, where each intermediate step is represented as a (label, bounding-box) pair capturing a verifiable visual state. This design enables explicit tracking of intermediate states, direct supervision for proposal quality, and avoids ambiguity introduced by language-based reasoning. Building upon verifiable and spatially grounded reasoning chains, Artemis provides a unified architecture for diverse perceptual tasks, without requiring the task-specific designs relied upon by prior perceptual policy models. Trained using grounding and detection sampeles in natural image domains, Artemis generalizes to counting and geometric perception tasks. At its core, a spatially grounded, object-centric chain rule provides a principled foundation for scalable and general perceptual policies.

Large language models exhibit complementary reasoning errors: on the same instance, one model may succeed with a particular decomposition while another fails. We propose Collaborative Reasoning (CoRe), a training-time collaboration framework that converts peer success into a learning signal via a cross-teaching protocol. Each problem is solved in two stages: a cold round of independent sampling, followed by a contexted rescue round in which failed models receive a hint extracted from a successful peer. CoRe optimizes a combined reward that balances (i) correctness, (ii) a lightweight DPP-inspired diversity term to reduce error overlap, and (iii) an explicit rescue bonus for successful recovery. We evaluate CoRe in low-data regimes across GSM8K, MATH, AIME, and GPQA. With only 1,000 training examples, a pair of small open-source models (3B+4B) reaches 99.54% oracle Team Pass@2 on GSM8K and 92.08% on MATH, compared to 82.50% and 74.82% for single-model training. On harder datasets, the same pair reaches 77.34% oracle Team Pass@2 on GPQA (trained on 348 examples) and 79.65% on AIME (trained on 792 examples), using a training-time budget of at most 1536 context tokens and 3072 generated tokens. A three-model non-oracle AIME evaluation further closes most of the oracle-selection gap (81.20% majority vote vs. 82.60% oracle Team Pass@2). Overall, these results show that training-time collaboration can reliably convert model complementarity into large gains without scaling model size.