ICML 2026 论文列表

Large Language Models exhibit paradoxical fragility in fundamental arithmetic, implying a disconnect between internal computation and discrete output. By analyzing the residual stream geometry during multi-operand addition, we identify the **Iso-Raw-Sum Trajectory (IRST)**, a topological manifold where representations are anchored by semantic digits and modulated by continuous carry fibers. We propose the **Noisy Quantization Model**, which frames arithmetic errors as *topological slippages* caused by internal neural noise pushing a continuous, latent *carry potential* across quantization thresholds. This geometric framework further elucidates *probe versatility*, explaining how lightweight probes can disentangle conflicting latent signals (such as ground truth versus hallucination) from a single activation vector. Finally, we validate these insights through a geometric consistency check method that effectively detects and corrects these quantization failures during inference.

Large language models (LLMs) are often hypothesized to perform implicit Bayesian inference, yet a key coherence condition—the martingale property of predictive beliefs—has been shown to fail in controlled synthetic in-context learning settings. We revisit this question in a more typical usage regime: generic multiple-choice question answering. Exploiting the discrete answer space, we compute exact predictive distributions and study belief dynamics induced by autoregressive answer resampling. We introduce prompted predictive resampling (PPR), where an LLM generates a sequence of answers to the same question. Empirically, PPR reveals early-stage belief drift, indicating martingale violations. However, after sufficient resampling steps, the belief process self-stabilizes and converges to a coherent predictive distribution. Based on this observation, we further propose (i) a seed-answer prompting strategy to accelerate stabilization, and (ii) a self-consistency loss that amortizes early-stage drift into the model via fine-tuning. Experiments on multiple-choice QA benchmarks show that our methods substantially reduce belief drift and improve predictive coherence without sacrificing accuracy.

In this work, we propose One-shot Entropy Minimization (EM), a simple and fully unsupervised post-training approach that significantly improves reasoning and generation performance using only a single unlabeled data and approximately ten gradient steps. To avoid data contamination, we pretrain a 7-billion-parameter language model from scratch with strictly decontaminated data. Despite its extreme simplicity, one-shot EM yields substantial performance gains and improves reasoning abilities across a broad range of domains, including mathematical reasoning, logical reasoning, and coding. We further show that entropy minimization induces a characteristic right-skewed logit shift, amplifying high-probability tokens while suppressing low-probability tails, in contrast to reinforcement learning. Our findings suggest that entropy minimization primarily acts as a distribution shaping mechanism rather than a conventional learning process, offering an efficient and practical algorithm for post-training large language models.

Safe reinforcement learning (RL) is a standard paradigm for safety-critical decision making. However, real-world safety constraints can be complex, subjective, and even hard to explicitly specify. Existing works on constraint inference rely on restrictive assumptions or extensive expert demonstrations, which are not realistic in many real-world applications. How to cheaply and reliably learn these constraints is the major challenge we focus on in this study. While inferring constraints from human preferences offers a data-efficient alternative, we identify popular Bradley-Terry (BT) models fail to capture the asymmetric, heavy-tailed nature of safety costs, resulting in risk underestimation. It is still rare in the literature to understand the impacts of BT models on the downstream policy learning. To address the above knowledge gaps, we propose a novel approach namely Preference-based Constrained Reinforcement Learning (PbCRL). We introduce a novel dead zone mechanism into preference modeling and theoretically prove that it encourages heavy-tailed cost distributions, thereby achieving better constraint alignment. Additionally, we incorporate a Signal-to-Noise Ratio (SNR) loss to encourage exploration by cost variances, which is found to benefit policy learning. Further, two-stage training strategy is deployed to lower online labeling burdens while adaptively enhancing constraint satisfaction. Empirical results demonstrate that PbCRL achieves superior alignment with true safety requirements and outperforms state-of-the-art baselines in terms of safety and reward. Our work explores a promising and effective way for constraint inference in Safe RL, with great potential in various safety-critical applications.

Diffusion models achieve state-of-the-art performance in generative modeling but are limited by their reliance on Gaussian noise and the high computational cost of iterative sampling. Star-shaped diffusion addresses the former by introducing a non-Markovian forward process, yet this comes at the expense of temporal coherence in the reverse process. We propose a novel framework that resolves this trade-off by learning a Markovian projection of a star-shaped forward process, and its reversal. This design enables learning over a broad class of exponential models and recovers DDPM as a special case. It is particularly well-suited for knowledge distillation, allowing few-step or even single-step generation. Experiments demonstrate the effectiveness and flexibility of our approach across multiple generative tasks. Code and demo are available at: https://oublalkhalid.github.io/MStar-Diffusion/.

We introduce Unsupervised Partner Design (UPD), a population-free multi-agent reinforcement learning method for robust ad-hoc teamwork. UPD generates training partners on-the-fly and selects them adaptively based on a learnability criterion, removing the need for pre-trained partner populations or manual parameter tuning. We show that this simple mechanism enables effective partner diversity and can be extended to joint partner-environment selection when a procedural level generator is available. Across Level-Based Foraging, Overcooked-AI, and the Overcooked Generalisation Challenge, UPD consistently achieves strong performance compared to both population-based and population-free baselines. In a human-AI user study, agents trained with UPD achieve higher returns and are rated as more adaptive, more human-like, and less frustrating than all evaluated baseline methods.

With the widespread deployment of large language models (LLMs), aligning model outputs with pluralistic human values has become an important research problem. Recent approaches that train task-specific experts and merge them through parameter aggregation have shown promise for pluralistic alignment. However, these methods often overlook the intrinsic complexity of real-world value data, where multiple correlated value dimensions coexist, resulting in highly similar and entangled expert representations. Consequently, modifying the contribution of one value expert may unintentionally influence other values, limiting fine-grained controllability. To address this issue, we propose DisAlign, a model-merging framework that explicitly decomposes value representations into consensus and value-specific components using an information-geometric perspective. DisAlign first extracts a consensus anchor and subspace to capture shared structure across values, and then applies spectral decomposition to the residual representations to construct disentangled value subspaces. This design enables more precise and independent modulation of multiple values. Experiments on three datasets covering different value frameworks demonstrate that DisAlign consistently improves value disentanglement and achieves more accurate pluralistic value control compared to existing baselines. Our code is available at \url{https://anonymous.4open.science/r/DisAlign-7F35}

Consider a market of competing model providers selling query access to models with varying costs and capabilities. Customers submit problem instances and are willing to pay up to a budget for a verifiable solution. An arbitrageur efficiently allocates inference budget across providers to undercut the market, thus creating a competitive offering with no model-development risk. In this work, we initiate the study of arbitrage in AI model markets, empirically demonstrating the viability of arbitrage and illustrating its economic consequences. We conduct an in-depth case study of SWE-bench GitHub issue resolution using two representative models, GPT-5 mini and DeepSeek v3.2. In this verifiable domain, simple arbitrage strategies generate net profit margins of up to 40\%. Robust arbitrage strategies that generalize across different domains remain profitable. Distillation further creates strong arbitrage opportunities, potentially at the expense of the teacher model’s revenue. Multiple competing arbitrageurs drive down consumer prices, reducing the marginal revenue of model providers. At the same time, arbitrage reduces market segmentation and facilitates market entry for smaller model providers by enabling earlier revenue capture. Our results suggest that arbitrage can be a powerful force in AI model markets with implications for model development, distillation, and deployment.

We introduce a formal active learning methodology for guiding the placement of Lagrangian observers to infer time-dependent vector fields -- a key task in oceanography, marine science, and ocean engineering -- using a physics-informed spatio-temporal Gaussian process surrogate model. The majority of existing placement campaigns either follow standard `space-filling' designs or relatively ad-hoc expert opinions. A key challenge to applying principled active learning in this setting is that Lagrangian observers are continuously advected through the vector field, so they make measurements at different locations and times. It is, therefore, important to consider the likely future trajectories of placed observers to account for the utility of candidate placement locations. To this end, we present BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories. We observe noticeable benefits of BALLAST-aided sequential observer placement strategies on both synthetic and high-fidelity ocean current models. In addition, we developed a novel GP inference method -- the Vanilla SPDE Exchange (VaSE) -- to boost the GP posterior sampling efficiency, which is also of independent interest.

Likelihood-based training is the dominant paradigm in survival prediction. Under independent censoring, we can factorize the likelihood and optimize only the terms related to event modeling, effectively treating the censoring mechanism as incidental. This is justified when censoring is *non-informative*, i.e., when the censoring process shares no parameters with the event-time model. However, this may not hold in practice, and ignoring censoring contributions may discard useful signals for learning representations that can help to effectively estimate event distributions. Motivated by this, we argue that explicitly modeling censoring can improve representation learning and time-to-event estimation, particularly when event and censoring processes are coupled. We introduce a latent decomposition view in which observed covariates are mapped to latent components corresponding to event-specific, censoring-specific, confounding, and irrelevant information. We then learn decomposed representations for the first three categories to guide a better estimation of the event distribution. We instantiate our method on 4 popular deep-learning survival models and evaluate on 10 datasets (2 semi-synthetic and 8 real-world), showing consistent gains over strong baselines and multiple SOTA methods.

Repository-level code editing requires models to understand complex dependencies and execute precise multi-file modifications across a large codebase. While recent gains on SWE-bench rely heavily on complex agent scaffolding, it remains unclear how much of this capability can be internalised via high-quality training signals. To address this, we propose Clean Pull Request (Clean-PR), a mid-training paradigm that leverages real-world GitHub pull requests as a training signal for repository-level editing. We introduce a scalable pipeline that converts noisy pull request diffs into Search/Replace edit blocks through reconstruction and validation, resulting in the largest publicly available corpus of 2 million pull requests spanning 12 programming languages. Using this training signal, we perform a mid-training stage followed by an agentless-aligned supervised fine-tuning process with error-driven data augmentation. On SWE-bench, our model significantly outperforms the instruction-tuned baseline, achieving absolute improvements of 13.6% on SWE-bench Lite and 12.3% on SWE-bench Verified. These results demonstrate that repository-level code understanding and editing capabilities can be effectively internalised into model weights under a simplified, agentless protocol, without relying on heavy inference-time scaffolding.

Low-Rank Adaptation (LoRA) enables efficient federated fine-tuning of segmentation foundation models for medical imaging. However, most federated LoRA methods adopt a uniform aggregation rule, which breaks under the encoder–decoder asymmetry in medical segmentation: the encoder is dominated by appearance shifts, while the decoder is dominated by supervision variations. This mismatch entangles shared anatomy with site-specific biases and harms generalization. To address this, we propose Inverse Asymmetric Tuning (IAT). IAT aligns adaptation with heterogeneity sources by personalizing module-specific components in the encoder to absorb appearance shifts and in the decoder to accommodate site-dependent supervision, while retaining a shared pathway for transferable consensus. However, structural separation alone is insufficient under LoRA’s bilinear parameterization, where multiplicative coupling can still cause site-specific updates to leak into the shared direction. We therefore introduce a Subspace Orthogonality Regularizer that penalizes shared–local collinearity in the effective update space, mitigating leakage without extra communication. Experiments show consistent improvements over strong federated LoRA and parameter-efficient FL baselines.

Foundation models must handle multiple generative processes, yet mechanistic interpretability largely studies capabilities in isolation; it remains unclear how a single transformer organizes multiple, potentially conflicting "world models". Previous experiments on Othello-playing neural networks test world-model learning, but focus on a single game with a single set of rules. We introduce *MetaOthello*, a controlled suite of Othello-like games with shared syntax but different rules or tokenizations, and train small GPTs on mixed-variant data. We show that transformers trained on multiple Othello variants learn **shared world-state representations**: linear probes trained on one game intervene on another's board state nearly as well as matched probes. When the games conflict, the model resolves the resulting *ambiguity* through a localized mechanism we identify and steer. For isomorphic games with token remapping, representations are equivalent up to a single orthogonal rotation that generalizes across layers, showing the shared structure is abstract rather than tied to surface form. Together, these results show that transformers reconcile conflicting world models by sharing structure and localizing conflict. *MetaOthello* thus offers a path toward understanding how transformers organize many world models at once.

Causal machine learning aims to answer "what if" questions using machine learning algorithms, making it a promising tool for high-stakes decision-making. Yet, empirical evaluation practices remain limited. Existing benchmarks often rely on a handful of hand-crafted or semi-synthetic datasets, leading to brittle, non-generalizable conclusions. To bridge this gap, we introduce CausalProfiler, a synthetic benchmark generator for causal machine learning methods. Based on a set of explicit design choices about the class of causal models, queries, and data considered, CausalProfiler randomly samples causal models, data, queries, and ground truths constituting the synthetic causal benchmarks. In this way, causal machine learning can be rigorously and transparently evaluated under a variety of conditions. This work offers the first random generator of synthetic causal benchmarks with coverage guarantees and transparent assumptions operating on the three levels of causal reasoning: observation, intervention, and counterfactual. We demonstrate its utility by evaluating several state-of-the-art methods under diverse conditions and assumptions, both in and out of the identification regime, illustrating the types of analyses and insights CausalProfiler enables.

Reasoning Video Object Segmentation (RVOS) demands a sophisticated integration of temporal dynamics, spatial details, and linguistic reasoning to achieve precise pixel-level localization. Existing methods are limited to reasoning over fixed initial inputs and lack the capacity to actively acquire further visual evidence, which is often essential for resolving complex references in long or intricate videos. To address this, we propose $\textbf{VideoSEG-O3}$, the first multi-turn reinforcement learning framework for RVOS that emulates the human $\textit{``coarse-to-fine''}$ cognitive process. It employs a $\textit{multi-turn temporal-spatial chain-of-thought}$ to capture fine-grained details by iteratively pinpointing critical intervals and keyframes. Additionally, to enable the policy to perceive segmentation quality beyond mere text probability of $\texttt{[SEG]}$ during the RL stage, we introduce $\textit{SEG-aware logit calibration}$, which integrates pixel-wise segmentation feedback directly into the token-level logits. Furthermore, we design a $\textit{decoupled thinking trace}$ to hierarchically decompose the reasoning process into temporal, spatial, and linguistic dimensions, and construct $\textbf{VTS-CoT}$, a specialized cold-start dataset featuring comprehensive reasoning trajectories. Extensive experiments demonstrate that VideoSEG-O3 achieves advanced performance across 8 mainstream RVOS benchmarks, particularly excelling in long-horizon and complex reasoning tasks.

Existing code reasoning methods primarily supervise final code outputs, ignoring intermediate states, often leading to reward hacking where correct answers are obtained through inconsistent reasoning. We propose StepCodeReasoner, a framework that introduces explicit intermediate execution-state supervision. By automatically inserting structured print-based execution-trace anchors into code, the model is trained to predict runtime states at each step, transforming code reasoning into a verifiable, stepwise execution modeling problem. Building on this execution-aware method, we introduce Dual-Granularity GRPO (DG-GRPO), a reinforcement learning algorithm for structured credit assignment at two complementary granularities: inter-trajectory comparison across sampled execution paths and intra-trajectory shaping based on downstream correctness. Extensive experiments demonstrate that StepCodeReasoner achieves state-of-the-art performance in code reasoning. In particular, our 7B model achieves 91.1\% on CRUXEval and 86.5\% on LiveCodeBench, outperforming the CodeReasoner-7B baseline (86.0\% and 77.7\%) and GPT-4o (85.6\% and 75.1\%). Furthermore, on the execution-trace benchmark REval, our model scores 82.9\%, outperforming baseline CodeReasoner-7B (72.3\%), its 14B counterpart (81.1\%), and GPT-4o (77.3\%). In addition, StepCodeReasoner improves code generation, achieving 90.1 on HumanEval, 85.0 on MBPP, and 19.4 on LiveCodeBench generation, with an average score of 64.8 versus 62.6 for CodeReasoner-7B.

The forest matrix of a signed graph plays an important role in network science and social opinion dynamics, yet existing algorithms are mainly designed for unsigned graphs and are difficult to extend to signed graphs. In this paper, we study the problem of efficiently estimating the forest matrix of signed graphs with \(n\) nodes and introduce the signed forest matrix theorem, which establishes the relationship between generalized spanning converging forests and the forest matrix. Based on this result, we propose a novel algorithm \textsc{GSCF}, built on a variant of loop-erased random walks, to generate generalized spanning converging forests in expected \(O(n)\) time. We further develop two sampling algorithms, \textsc{FMDE} and \textsc{FMDE+}, for estimating the diagonal of the forest matrix, both with time complexity \(O(ln)\), where \(l\) is the number of samples. Extensive experiments on various signed graphs show that our methods achieve high estimation accuracy, significantly improve computational efficiency, and scale to graphs with over twenty million nodes. Our source code is publicly available on \url{https://github.com/HaoxinSun98/SignedForestDiagonal}.

Continuous diffusion for categorical data is a framework belonging to the diffusion family and aiming at generating discrete data. The scientific interest to such models has been constantly increasing these days because researchers try to achieve a challenging goal of finding reasonable alternatives to autoregressive large language models. In this paper, we study the properties of the structure of the latent space corresponding to discrete tokens expressed in terms of Kullback-Leibler divergence on diffusion path measures and accuracy of the correct token prediction by the optimally trained diffusion model. We find that FSQ tokenization scheme has the latent space structure with the properties that make it best suited for continuous diffusion for categorical data as verified through rigorous theoretical analysis and numerical experiments. To validate our findings in real-life scenario, we train several text-to-speech diffusion models having speech tokens as intermediate acoustic features, and show that the one based on FSQ tokens indeed performs the best, and, moreover, it outperforms its strong LLM-based counterpart, at the same time being significantly smaller and faster.

Extended reasoning in large language models (LLMs) requires long and accurate decoding and creates severe KV cache memory bottlenecks. Leading KV cache compression methods estimate KV importance using attention scores from recent post-RoPE queries. However, queries rotate with position during RoPE, making representative queries very few, leading to poor top-key selection and unstable reasoning. To avoid this issue, we turn to the pre-RoPE space, where we observe that Q and K vectors are highly concentrated around fixed non-zero centers and remain stable across positions—*Q/K concentration*. We show that this concentration causes queries to preferentially attend to keys at specific distances (e.g., nearest keys), with the centers determining which distances are preferred via a trigonometric series. Based on this, we propose TriAttention to estimate key importance by leveraging these centers. Via the trigonometric series, we use the distance preference characterized by these centers to score keys according to their positions, and also leverage Q/K norms as an additional signal for importance estimation. On AIME25 with 32K-token generation, TriAttention matches Full Attention reasoning accuracy while achieving 2.5× higher throughput or 10.7× KV memory reduction, whereas leading baselines achieve only about half the accuracy at the same efficiency.