ICLR 2026 论文列表

Multimodal Large Reasoning Models (MLRMs) have exhibited remarkable capabilities in complex multimodal tasks. However, our findings reveal a critical trade-off: reasoning-based models are more prone to generating harmful content, leading to degradation in safety performance. This paper presents a large-scale analysis of this safety–reasoning trade-off, identifying two main mechanisms of safety degradation: (i) visual attention drift, which reduces the model’s reliance on visual grounding and thereby exacerbates overlooked risks in cross-modal interactions; (ii) unsafe reasoning patterns, including flawed reasoning initiation and chain-of-thought safety attenuation, which compromise the model’s safety awareness. To mitigate these issues, we propose **P**olicy-guided **S**afety **T**uning (**PST**), a two-stage alignment framework. It first employs *Policy-Guided Supervised Fine-Tuning* to integrate explicit safety policies into the reasoning process, establishing a structured and interpretable foundation for safe decision-making. Then, PST applies *Safety Reasoning Preference Optimization* to encourage the model to generate safe, helpful, and informative responses while reducing oversensitive and homogeneous characteristics. Extensive experiments demonstrate that PST significantly reduces harmful outputs across multiple multimodal safety benchmarks, while maintaining competitive performance on general tasks.

Diffusion Policy has demonstrated strong visuomotor modeling capabilities, but its high computational cost renders it impractical for real-time robotic control. Despite huge redundancy across repetitive denoising steps, existing diffusion acceleration techniques fail to generalize to Diffusion Policy due to fundamental architectural and data divergences. In this paper, we propose **B**lock-wise **A**daptive **C**aching (**BAC**), a method to accelerate Diffusion Policy by caching intermediate action features. BAC achieves lossless action generation acceleration by adaptively updating and reusing cached features at the block level, based on a key observation that feature similarities exhibit non-uniform temporal dynamics and distinct block-specific patterns. To operationalize this insight, we first design an Adaptive Caching Scheduler to identify optimal update timesteps by maximizing the global feature similarities between cached and skipped features. However, applying this scheduler for each block leads to significant error surges due to the inter-block propagation of caching errors, particularly within Feed-Forward Network (FFN) blocks. To mitigate this issue, we develop the Bubbling Union Algorithm, which truncates these errors by updating the upstream blocks with significant caching errors before downstream FFNs. As a training-free plugin, BAC is readily integrable with existing transformer-based Diffusion Policy and vision-language-action models. Extensive experiments on multiple robotic benchmarks demonstrate that BAC achieves up to 3x inference speedup for free. Project page: https://block-wise-adaptive-caching.github.io.

We revisit policy-gradient optimization for Large Language Models (LLMs) from a single-stream perspective. Prevailing group-based methods like GRPO reduce variance with on-the-fly baselines but suffer from critical flaws: frequent degenerate groups erase learning signals, and synchronization barriers hinder scalability. We introduce Single-stream Policy Optimization (SPO), which eliminates these issues by design. SPO replaces per-group baselines with a persistent, KL-adaptive value tracker and normalizes advantages globally across the batch, providing a stable, low-variance learning signal for every sample. Being group-free, SPO enables higher throughput and scales effectively in long-horizon or tool-integrated settings where generation times vary. Furthermore, the persistent value tracker naturally enables an adaptive curriculum via prioritized sampling. Experiments using Qwen3-8B show that SPO converges more smoothly and attains higher accuracy than GRPO, while eliminating computation wasted on degenerate groups. Ablation studies confirm that SPO's gains stem from its principled approach to baseline estimation and advantage normalization, offering a more robust and efficient path for LLM reasoning. Across five hard math benchmarks with Qwen3-8B, SPO improves the average maj@32 by $+3.4\ \text{percentage points} (\mathrm{pp})$ over GRPO, driven by substantial absolute point gains on challenging datasets, including $+7.3\ \mathrm{pp}$ on BRUMO 25, $+4.4\ \mathrm{pp}$ on AIME 25, $+3.3\ \mathrm{pp}$ on HMMT 25, and achieves consistent relative gain in pass@$k$ across the evaluated $k$ values. SPO's success challenges the prevailing trend of adding incidental complexity to RL algorithms, highlighting a path where fundamental principles, not architectural workarounds, drive the next wave of progress in LLM reasoning.

Large-scale robot learning has made progress on complex manipulation tasks, yet long-horizon, contact-rich problems—especially those involving deformable objects—remain challenging due to inconsistent demonstration quality. We propose a stage-aware, video-based reward modeling framework that jointly predicts task stage and fine-grained progress, using natural-language subtask annotations to derive consistent labels across variable-length demonstrations. This avoids the brittleness of frame-index-based labeling and provides stable supervision even in tasks like T-shirt folding. Our reward model is robust to demonstration variability, generalizes to out-of-distribution scenarios, and improves downstream policy training. Building on it, we introduce Reward-Aligned Behavior Cloning (RA-BC), which filters and reweights demonstrations based on reward estimates. Experiments show that our method significantly outperforms baselines in both real-world rollouts and human validation. On T-shirt folding, we achieve 83\% success from the flattened state and 67\% from the crumpled state, compared to 8\% and 0\% with vanilla BC. Overall, our results highlight reward modeling as a scalable and annotation-efficient solution for long-horizon robotic manipulation. Project website: https://qianzhong-chen.github.io/sarm.github.io/.

Major progress on language models (LMs) in recent years has largely resulted from moving away from specialized models designed for specific tasks, to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data. Despite this trend, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechanism which automatically learns content- and context- dependent segmentation strategies learned jointly with the rest of the model. Incorporating this into an explicit hierarchical network (H-Net) allows replacing the (implicitly hierarchical) tokenization--LM--detokenization pipeline with a single model learned fully end-to-end. When compute- and data- matched, an H-Net with one stage of hierarchy operating at the byte level outperforms a strong Transformer language model operating over BPE tokens. Iterating the hierarchy to multiple stages further increases its performance by modeling multiple levels of abstraction, demonstrating significantly better scaling with data and matching the token-based Transformer of twice its size. H-Nets pretrained on English show significantly increased character-level robustness, and qualitatively learn meaningful data-dependent chunking strategies without any heuristics or explicit supervision. Finally, the H-Net's improvement over tokenized pipelines is further increased in languages and modalities with weaker tokenization heuristics, such as Chinese and code, or DNA sequences (nearly 4x improvement in data efficiency over baselines), showing the potential of true end-to-end models that learn and scale better from unprocessed data.

Healthcare insurance fraud detection presents unique machine learning challenges: labeled data are scarce due to delayed verification processes, and fraudulent behaviors evolve rapidly, often manifesting in complex, graph-structured interactions. Existing methods struggle in such settings. Pretraining routines typically overlook structural anomalies under limited supervision, while online models often fail to adapt to changing fraud patterns without labeled updates. To address these issues, we propose the Continual Fiedler Vector Graph model (ConFVG), a fraud detection framework designed for label-scarce and non-stationary environments. The framework comprises two key components. To mitigate label scarcity, we develop a Fiedler Vector-guided graph autoencoder that leverages spectral graph properties to learn structure-aware node representations. The Fiedler Vector, derived from the second smallest eigenvalue of the graph Laplacian, captures global topological signals such as community boundaries and connectivity bottlenecks, which are patterns frequently associated with collusive fraud. This enables the model to identify structurally anomalous nodes without relying on labels. To handle evolving graph streams, we propose a Subgraph Attention Fusion (SAF) module that constructs neighborhood subgraphs and applies attention-based reweighting to emphasize emerging high-risk structures. This design allows the model to adapt to new fraud patterns in real time. A Mean Teacher mechanism further stabilizes online updates and prevents forgetting of previously acquired knowledge. Experiments on real-world medical fraud datasets demonstrate that the Continual Fiedler Vector Graph model outperforms state-of-the-art baselines in both low-label and distribution-shift scenarios, offering a scalable and structure-sensitive solution for real-time fraud detection.

Irregular multivariate time series (IMTS) are prevalent in critical domains like healthcare and finance, where accurate forecasting is vital for proactive decision-making. However, the asynchronous sampling and irregular intervals inherent to IMTS pose two core challenges for existing methods: (1) how to accurately represent the raw information of irregular time series without introducing data distortion, and (2) how to effectively capture the complex dynamic dependencies between observation points. To address these challenges, we propose the Adaptive Spatio-Temporal Graph Interaction (ASTGI) framework. Specifically, the framework first employs a Spatio-Temporal Point Representation module to encode each discrete observation as a point within a learnable spatio-temporal embedding space. Second, a Neighborhood-Adaptive Graph Construction module adaptively builds a causal graph for each point in the embedding space via nearest neighbor search. Subsequently, a Spatio-Temporal Dynamic Propagation module iteratively updates information on these adaptive causal graphs by generating messages and computing interaction weights based on the relative spatio-temporal positions between points. Finally, a Query Point-based Prediction module generates the final forecast by aggregating neighborhood information for a new query point and performing regression. Extensive experiments on multiple benchmark datasets demonstrate that ASTGI outperforms various state-of-the-art methods.

Link sign prediction on a signed graph is a task to determine whether the relationship represented by an edge is positive or negative. Since the presence of negative edges violates the graph homophily assumption that adjacent nodes are similar, regular graph methods have not been applicable without auxiliary structures to handle them. We aim to directly model the latent statistical dependency among edges with the Gaussian copula and its corresponding correlation matrix, extending CopulaGNN (Ma et al., 2021). However, a naive modeling of edge-edge relations is computationally intractable even for a graph with moderate scale. To address this, we propose to 1) represent the correlation matrix as a Gramian of edge embeddings, significantly reducing the number of parameters, and 2) reformulate the conditional probability distribution to dramatically reduce the inference cost. We theoretically verify scalability of our method by proving its linear convergence. Also, our extensive experiments demonstrate that it achieves significantly faster convergence than baselines, maintaining competitive prediction performance to the state-of-the-art models.

Traditional on-policy Reinforcement Learning with Verifiable Rewards (RLVR) frameworks suffer from experience waste and reward homogeneity, which directly hinders learning efficiency on difficult samples during large language models post-training. In this paper, we introduce Batch Adaptation Policy Optimization (BAPO), an off-policy RLVR framework to improve the data efficiency in large language models post-training. It dynamically selects training batches by re-evaluating historically difficult samples and reusing high-quality ones, while holding a lower bound guarantee for policy improvement. Extensive experiments further demonstrate that BAPO achieves an average 12.5\% improvement over GRPO across mathematics, planning, and visual reasoning tasks. Crucially, BAPO successfully resolves 40.7\% of problems that base models consistently fail to solve.

Recent advancements in reinforcement learning from verifiable rewards (RLVR), particularly through Group Relative Policy Optimization (GRPO), have significantly improved the capabilities of large language models (LLMs) for interactive coding agents. However, these methods overlook process-verifiable environment feedback (e.g., code execution failures), leading to inaccurate advantage estimation at each reasoning step and insufficient learning. To address this issue, we propose Group Verification-based Policy Optimization (GVPO), a novel RL algorithm that introduces an advantage shaping framework integrating both outcome-verifiable and process-verifiable signals. While outcome-verifiable rewards ensure alignment with long-term task objectives, process-verifiable feedback derived from intermediate execution traces (e.g., syntax errors, runtime exceptions) serves as corrective shaping terms at the step level. By jointly leveraging these two forms of verifiability, GVPO achieves more accurate credit assignment, balancing short-term process guidance with long-term outcome alignment. This unified formulation yields more stable optimization, faster convergence, and stronger generalization in complex interactive environments. A 32B-parameter agent trained with GVPO in the AppWorld environment outperforms OpenAI’s o1 agent by 12.6\%s on the more challenging Test-C split and surpasses the strongest 32B RL-trained state-of-the-art baseline by 3.6\%.

Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce **T**hinking-**F**ree **P**olicy **I**nitialization (**TFPI**), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple *ThinkFree* operation, explicitly discarding the thinking content via a direct *</think>* append, to reduce token usage during inference. Training with *ThinkFree*-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that {\method} accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.

Visuomotor policy learning has witnessed substantial progress in robotic manipulation, with recent approaches predominantly relying on generative models to model the action distribution. However, these methods often overlook the critical coupling between visual perception and action prediction. In this work, we introduce Triply-Hierarchical Diffusion Policy (H$^3$DP), a novel visuomotor learning framework that explicitly incorporates hierarchical structures to strengthen the integration between visual features and action generation. H$^3$DP contains $\mathbf{3}$ levels of hierarchy: (1) depth-aware input layering that organizes RGB-D observations based on depth information; (2) multi-scale visual representations that encode semantic features at varying levels of granularity; and (3) a hierarchically conditioned diffusion process that aligns the generation of coarse-to-fine actions with corresponding visual features. Extensive experiments demonstrate that H$^3$DP yields a $+ \mathbf{27.5}$% average relative improvement over baselines across $\mathbf{44}$ simulation tasks and achieves superior performance in $\mathbf{4}$ challenging bimanual real-world manipulation tasks. Project Page: https://h3-dp.github.io/.

Reinforcement learning (RL) policies represented in Reproducing Kernel Hilbert Spaces (RKHS) offer powerful representational capabilities. While second-order optimization methods like Newton's method demonstrate faster convergence than first-order approaches, current RKHS-based policy optimization remains constrained to first-order techniques. This limitation stems primarily from the intractability of explicitly computing and inverting the infinite-dimensional Hessian operator in RKHS. We introduce Policy Newton in RKHS, the first second-order optimization framework specifically designed for RL policies represented in RKHS. Our approach circumvents direct computation of the inverse Hessian operator by optimizing a cubic regularized auxiliary objective function. Crucially, we leverage the Representer Theorem to transform this infinite-dimensional optimization into an equivalent, computationally tractable finite-dimensional problem whose dimensionality scales with the trajectory data volume. We establish theoretical guarantees proving convergence to a local optimum with a local quadratic convergence rate. Empirical evaluations on a toy financial asset allocation problem validate these theoretical properties, while experiments on standard RL benchmarks demonstrate that Policy Newton in RKHS achieves superior convergence speed and higher episodic rewards compared to established first-order RKHS approaches and parametric second-order methods. Our work bridges a critical gap between non-parametric policy representations and second-order optimization methods in reinforcement learning.

Chain-of-Thought (CoT) reasoning often fails to faithfully reflect a language model's underlying decision process. We address this by introducing a \emph{Markovian} language model framework with an autoencoder-style \emph{reasoning bottleneck}: it creates a text-based bottleneck where CoT serves as an intermediate representation, forcing the model to compress essential reasoning into interpretable text before making predictions, in the sense of learning short intermediate descriptions that make answers easy to compute from questions. We train this system with a GRPO-style policy gradient algorithm using parallel sampling, a frozen baseline CoT$'$, within-batch standardized advantages, and actor-reward (chain-rule) gradients. On QA tasks, Markovian training recovers most of the gains of a non-Markovian GRPO variant while forcing the model to answer from the CoT alone (e.g., GSM8K: 19.6\% $\to$ 57.1\%; ARC-Challenge: 36.1\% $\to$ 79.9\%; on average only $\approx$3-4 pp below a non-Markovian upper bound). Perturbation analyses across types and severities show that Markovian models incur systematically larger log-probability drops under CoT corruption than matched Non-Markovian baselines, indicating stronger causal reliance on the CoT. Cross-model evaluation confirms that learned CoTs generalize across architectures, suggesting they capture transferable reasoning patterns rather than model-specific artifacts.

Policy gradient methods are known to be highly sensitive to the choice of policy parameterization. In particular, the widely used softmax parameterization can induce ill-conditioned optimization landscapes and lead to exponentially slow convergence. Although this can be mitigated by preconditioning, this solution is often computationally expensive. Instead, we propose replacing the softmax with an alternative family of policy parameterizations based on the generalized $f$-$\textit{softargmax}$. We further advocate coupling this parameterization with a regularizer induced by the same $f$-divergence, which improves the optimization landscape and ensures that the resulting regularized objective satisfies a Polyak--Łojasiewicz inequality. Leveraging this structure, we establish the $\textit{first explicit non-asymptotic last-iterate convergence guarantees}$ for stochastic policy gradient methods for finite MDPs $\textit{without any form of preconditioning}$. We also derive sample-complexity bounds for the unregularized problem and show that $f$-PG, with Tsallis divergences achieves $\textit{polynomial sample complexity}$ in contrast to the exponential complexity incurred by the standard softmax parameterization.

We propose ImitSAT, a branching policy for conflict-driven clause learning (CDCL) solvers based on imitation learning for the Boolean satisfiability problem (SAT). Unlike previous methods that predict instance-level signals to improve CDCL branching indirectly, or rely on reinforcement learning and insufficient CDCL information to enhance branching, ImitSAT learns from expert KeyTrace that collapses a full run into the sequence of surviving decisions. Replaying a KeyTrace on the same instance is nearly conflict-free, providing dense decision- level supervision and directly reducing propagations—the dominant contributor to wall-clock time. This prefix-conditioned supervision enables ImitSAT to reproduce high-quality branches without exploration, yielding faster convergence, stable training, and seamless integration into CDCL. Extensive experiments demonstrate that ImitSAT reduces propagation counts and runtime, outperforming state-of-the-art learned approaches. We released the source code and trained model at https://github.com/zewei-Zhang/ImitSAT.

Safe exploration is a key requirement for reinforcement learning agents to learn and adapt online, beyond controlled (e.g. simulated) environments. In this work, we tackle this challenge by utilizing suboptimal yet conservative policies (e.g., obtained from offline data or simulators) as priors. Our approach, SOOPER, uses probabilistic dynamics models to optimistically explore, yet pessimistically fall back to the conservative policy prior if needed. We prove that SOOPER guarantees safety throughout learning, and establish convergence to an optimal policy by bounding its cumulative regret. Extensive experiments on key safe RL benchmarks and real-world hardware demonstrate that SOOPER is scalable, outperforms the state-of-the-art and validate our theoretical guarantees in practice.

Reinforcement learning has been central to recent advances in large language model reasoning, but most algorithms rely on on-policy training that demands fresh rollouts at every update, limiting efficiency and scalability. Asynchronous RL systems alleviate this by decoupling rollout generation from training, yet their effectiveness hinges on tolerating large staleness in rollout data, a setting where existing methods either degrade in performance or collapse. We revisit this challenge and uncover a \emph{prosperity-before-collapse} phenomenon: stale data can be as informative as on-policy data if exploited properly. Building on this insight, we introduce M2PO (Second-Moment Trust Proxy Optimization), which constrains the second moment of importance weights to suppress only extreme outliers while preserving informative updates. Notably, M2PO sharply reduces the fraction of clipped tokens under high staleness (from 1.22% to 0.006% over training), precisely masking high-variance tokens while maintaining stable optimization. Extensive evaluation across six model scales (1.7B–32B) and eight reasoning benchmarks shows that M2PO delivers stable off-policy training even with data stale by *at least 256 model updates* and matches on-policy performance.

Self-improvement is a significant techniques within the realm of large language model (LLM), aiming to enhance the LLM performance without relying on external data. Despite its significance, generally how LLM performances evolve during the self-improvement process remains underexplored. In this paper, we theoretically model the training dynamics of self-improvement via the concept of solver-verifier gap. This is inspired by the conjecture that the performance enhancement of self-improvement stems from the gap between LLM's solver capability and verifier capability. Based on the theoretical framework, we further show how to model the entire training trajectory. This framework allows quantifying the capability limit of self-improvement by fitting the theoretical model to the experiment results. We validate the effectiveness of the theoretical framework on various LLMs and datasets. Beyond self-improvement, we extend our analysis to investigate how external data influences these dynamics within the framework. Notably, we find that under limited external data regimes, such external data can be utilized at any stage without significantly affecting final performances, which accords with the empirical observations.