Low-Bitrate Video Compression through Semantic-Conditioned Diffusion

By: Lingdong Wang · Guan-Ming Su · Divya Kothandaraman · Tsung-Wei Huang · Mohammad Hajiesmaili · Ramesh K. Sitaraman

7 April 2026 at 12:00

arXiv:2512.00408v2 Announce Type: replace-cross Abstract: Traditional video codecs optimized for pixel fidelity collapse at ultra-low bitrates and produce severe artifacts. This failure arises from a fundamental misalignment between pixel accuracy and human perception. We propose a semantic video compression framework named DiSCo that transmits only the most meaningful information while relying on generative priors for detail synthesis. The source video is decomposed into three compact modalities: a textual description, a spatiotemporally degraded video, and optional sketches or poses that respectively capture semantic, appearance, and motion cues. A conditional video diffusion model then reconstructs high-quality, temporally coherent videos from these compact representations. Temporal forward filling, token interleaving, and modality-specific codecs are proposed to improve multimodal generation and modality compactness. Experiments show that our method outperforms baseline semantic and traditional codecs by 2-10X on perceptual metrics at low bitrates.

Multi-ancestry transcriptome prediction with functionally informed variants in TOPMed MESA improves performance of transcriptome-wide association studies

3 April 2026 at 18:00

Am J Hum Genet. 2026 Apr 2;113(4):828-841. doi: 10.1016/j.ajhg.2026.03.008.

ABSTRACT

Reliable reference transcriptome prediction models are key to accurate multi-ancestry transcriptome-wide association studies (TWASs). We propose three methods leveraging functionally informed variants (FIVs) for transcriptome prediction models to improve multi-ancestry TWASs. We trained models on 1,287 multi-ancestry participants from the Trans-Omics for Precision Medicine (TOPMed) program Multi-Ethnic Study of Atherosclerosis (MESA) with RNA sequencing (RNA-seq) data from peripheral blood mononuclear cells (PBMCs). We validated models' prediction accuracy on two external independent datasets, Geuvadis and Jackson Heart Study. To test robustness of our methods for TWASs, we integrated models with three multi-ancestry GWASs from blood cell, lipid, and pulmonary function traits, respectively. Our methods presented similar prediction accuracy while using a smaller and functionally informed set of variants compared to the benchmark method, elastic net (EN). Overall, our methods achieved higher power and accuracy (with average improved accuracy of 24% over EN) for TWASs. However, no single proposed method outperformed all GWAS traits. To further improve TWAS performance, we propose an omnibus approach that aggregates TWAS summary statistics from our methods. The omnibus approach yielded the highest number of Bonferroni-significant TWAS genes for all GWAS traits, and it further improved TWAS power and accuracy for blood cell traits. Additionally, the omnibus approach detected some trait-relevant important genes that the EN missed. Our study demonstrates the value of including FIVs in multi-ancestry transcriptome prediction models for improving TWAS performance. Further, the observed TWAS improvement depends on the GWAS trait's relevance to the PBMCs used to build our transcriptome prediction models.

PMID:41932314 | DOI:10.1016/j.ajhg.2026.03.008

Improving Ensemble Forecasts of Abnormally Deflecting Tropical Cyclones with Fused Atmosphere-Ocean-Terrain Data

cs.AI, q-bio.NC updates on arXiv.org

By: Qixiang Li · Yuan Zhou · Shuwei Huo · Chong Wang · Xiaofeng Li

3 April 2026 at 12:00

arXiv:2603.29200v2 Announce Type: replace-cross Abstract: Deep learning-based tropical cyclone (TC) forecasting methods have demonstrated significant potential and application advantages, as they feature much lower computational cost and faster operation speed than numerical weather prediction models. However, existing deep learning methods still have key limitations: they can only process a single type of sequential trajectory data or homogeneous meteorological variables, and fail to achieve accurate forecasting of abnormal deflected TCs. To address these challenges, we present two groundbreaking contributions. First, we have constructed a multimodal and multi-source dataset named AOT-TCs for TC forecasting in the Northwest Pacific basin. As the first dataset of its kind, it innovatively integrates heterogeneous variables from the atmosphere, ocean, and land, thus obtaining a comprehensive and information-rich meteorological dataset. Second, based on the AOT-TCs dataset, we propose a forecasting model that can handle both normal and abnormally deflected TCs. This is the first TC forecasting model to adopt an explicit atmosphere-ocean-terrain coupling architecture, enabling it to effectively capture complex interactions across physical domains. Extensive experiments on all TC cases in the Northwest Pacific from 2017 to 2024 show that our model achieves state-of-the-art performance in TC forecasting: it not only significantly improves the forecasting accuracy of normal TCs but also breaks through the technical bottleneck in forecasting abnormally deflected TCs.

Improving Ensemble Forecasts of Abnormally Deflecting Tropical Cyclones with Fused Atmosphere-Ocean-Terrain Data

cs.AI, q-bio.NC updates on arXiv.org

By: Qixiang Li · Shuwei Huo · Chong Wang · Xiaofeng Li · Yuan Zhou

1 April 2026 at 12:00

arXiv:2603.29200v1 Announce Type: cross Abstract: Deep learning-based tropical cyclone (TC) forecasting methods have demonstrated significant potential and application advantages, as they feature much lower computational cost and faster operation speed than numerical weather prediction models. However, existing deep learning methods still have key limitations: they can only process a single type of sequential trajectory data or homogeneous meteorological variables, and fail to achieve accurate forecasting of abnormal deflected TCs. To address these challenges, we present two groundbreaking contributions. First, we have constructed a multimodal and multi-source dataset named AOT-TCs for TC forecasting in the Northwest Pacific basin. As the first dataset of its kind, it innovatively integrates heterogeneous variables from the atmosphere, ocean, and land, thus obtaining a comprehensive and information-rich meteorological dataset. Second, based on the AOT-TCs dataset, we propose a forecasting model that can handle both normal and abnormally deflected TCs. This is the first TC forecasting model to adopt an explicit atmosphere-ocean-terrain coupling architecture, enabling it to effectively capture complex interactions across physical domains. Extensive experiments on all TC cases in the Northwest Pacific from 2017 to 2024 show that our model achieves state-of-the-art performance in TC forecasting: it not only significantly improves the forecasting accuracy of normal TCs but also breaks through the technical bottleneck in forecasting abnormally deflected TCs.

Self-Improving Code Generation via Semantic Entropy and Behavioral Consensus

cs.AI, q-bio.NC updates on arXiv.org

By: Huan Zhang · Wei Cheng · Wei Hu

1 April 2026 at 12:00

arXiv:2603.29292v1 Announce Type: cross Abstract: Improving the code generation capabilities of large language models (LLMs) typically relies on supervised fine-tuning or preference optimization, both of which require costly external resources such as powerful teacher models or reliable test units. However, in real-world scenarios, it is much harder to obtain reference solutions and test oracles than problem descriptions and test inputs. In this paper, we tackle a challenging yet realistic question: Can a code language model improve itself without access to a superior teacher and a test oracle? To answer this, we propose ConSelf, a self-improving approach built upon two key ideas. First, we introduce code semantic entropy, a novel metric that measures problem-level uncertainty by assessing the functional diversity of program behaviors, enabling a curriculum construction with the most learnable problems. Second, we present consensus-driven direct preference optimization (Con-DPO), a preference-based fine-tuning method that weights each preference pair by its behavioral consensus, thereby mitigating the impact of noisy self-generated supervision. Experiments on various benchmarks and backbone LLMs demonstrate that ConSelf significantly outperforms baselines, validating the effectiveness of semantic entropy-based curriculum construction and consensus-driven optimization in improving code generation without external supervision.

KEditVis: A Visual Analytics System for Knowledge Editing of Large Language Models

cs.AI, q-bio.NC updates on arXiv.org

By: Zhenning Chen · Hanbei Zhan · Yanwei Huang · Xin Wu · Dazhen Deng · Di Weng · Yingcai Wu

1 April 2026 at 12:00

arXiv:2603.29689v1 Announce Type: cross Abstract: Large Language Models (LLMs) demonstrate exceptional capabilities in factual question answering, yet they sometimes provide incorrect responses. To address this issue, knowledge editing techniques have emerged as effective methods for correcting factual information in LLMs. However, typical knowledge editing workflows struggle with identifying the optimal set of model layers for editing and rely on summary indicators that provide insufficient guidance. This lack of transparency hinders effective comparison and identification of optimal editing strategies. In this paper, we present KEditVis, a novel visual analytics system designed to assist users in gaining a deeper understanding of knowledge editing through interactive visualizations, improving editing outcomes, and discovering valuable insights for the future development of knowledge editing algorithms. With KEditVis, users can select appropriate layers as the editing target, explore the reasons behind ineffective edits, and perform more targeted and effective edits. Our evaluation, including usage scenarios, expert interviews, and a user study, validates the effectiveness and usability of the system.

LPNSR: Prior-Enhanced Diffusion Image Super-Resolution via LR-Guided Noise Prediction

cs.AI, q-bio.NC updates on arXiv.org

By: Shuwei Huang · Shizhuo Liu · Zijun Wei

1 April 2026 at 12:00

arXiv:2603.21045v3 Announce Type: replace-cross Abstract: Diffusion-based image super-resolution (SR), which aims to reconstruct high-resolution (HR) images from corresponding low-resolution (LR) observations, faces a fundamental trade-off between inference efficiency and reconstruction quality. The state-of-the-art residual-shifting diffusion framework achieves efficient 4-step inference, yet suffers from severe performance degradation in compact sampling trajectories. This is mainly attributed to two core limitations: the inherent suboptimality of unconstrained random Gaussian noise in intermediate steps, which leads to error accumulation and insufficient LR prior guidance, and the initialization bias caused by naive bicubic upsampling. In this paper, we propose LPNSR, a prior-enhanced efficient diffusion framework to address these issues. We first mathematically derive the closed-form analytical solution of the optimal intermediate noise for the residual-shifting diffusion paradigm, and accordingly design an LR-guided multi-input-aware noise predictor to replace random Gaussian noise, embedding LR structural priors into the reverse process while fully preserving the framework's core efficient residual-shifting mechanism. We further mitigate initial bias with a high-quality pre-upsampling network to optimize the diffusion starting point. With a compact 4-step trajectory, LPNSR can be optimized in an end-to-end manner. Extensive experiments demonstrate that LPNSR achieves state-of-the-art perceptual performance on both synthetic and real-world datasets, without relying on any large-scale text-to-image priors. The source code of our method can be found at https://github.com/Faze-Hsw/LPNSR.

Cellular Senescence in Gastric Cancer: Molecular Mechanisms, Microenvironment Remodeling and Therapeutic Implications

Omics in Gastric

By: Zhiyuan Shi · Zhao Sun · Yuan Liu · Yuping Ge · Ning Jia · Yuwei Hua · Lin Zhao · Xiang Wang · Yi Ba

30 March 2026 at 18:00

Aging Dis. 2026 Mar 19. doi: 10.14336/AD.2025.1571. Online ahead of print.

ABSTRACT

Gastric cancer (GC) remains a leading cause of cancer-related morbidity and mortality worldwide, with poor prognosis for advanced-stage patients. Therefore, in-depth exploration of the mechanisms underlying GC initiation and progression, as well as the development of novel therapeutic strategies, is of crucial importance. Cellular senescence is a stable cell cycle arrest program that plays a dual role in GC. It exerts tumor-suppressive effects via growth arrest but also promotes tumor progression and immune evasion by remodeling the tumor microenvironment (TME) through senescence-associated secretory phenotype (SASP). This review comprehensively elucidates the molecular mechanisms of cellular senescence in GC and the core regulatory networks involving gene regulation, epigenetic modifications, metabolic reprogramming, and cell cycle arrest. Additionally, the review highlights how senescent cells foster an immunosuppressive microenvironment via SASP, forming a self-reinforcing feed-forward loop. Regarding therapeutic strategies, we summarize potential approaches targeting cellular senescence, including senescence induction, senescent cell clearance, SASP modulation, and multi-target synergistic therapy by integrating epigenetic regulation, metabolic intervention, and immune microenvironment modulation. Despite progress, numerous challenges remain. Future studies should leverage multi-omics technologies, novel models' development, and large-scale clinical trials to advance the clinical translation of GC cellular senescence research, providing new insights for improving prognosis.

PMID:41910653 | DOI:10.14336/AD.2025.1571

PhySe-RPO: Physics and Semantics Guided Relative Policy Optimization for Diffusion-Based Surgical Smoke Removal

cs.AI, q-bio.NC updates on arXiv.org

By: Zining Fang · Chunhui Liu · Bin Xu · Ming Chen · Xiaowei Hu · Cheng Xue

25 March 2026 at 12:00

arXiv:2603.22844v2 Announce Type: new Abstract: Surgical smoke severely degrades intraoperative video quality, obscuring anatomical structures and limiting surgical perception. Existing learning-based desmoking approaches rely on scarce paired supervision and deterministic restoration pipelines, making it difficult to perform exploration or reinforcement-driven refinement under real surgical conditions. We propose PhySe-RPO, a diffusion restoration framework optimized through Physics- and Semantics-Guided Relative Policy Optimization. The core idea is to transform deterministic restoration into a stochastic policy, enabling trajectory-level exploration and critic-free updates via group-relative optimization. A physics-guided reward imposes illumination and color consistency, while a visual-concept semantic reward learned from CLIP-based surgical concepts promotes smoke-free and anatomically coherent restoration. Together with a reference-free perceptual constraint, PhySe-RPO produces results that are physically consistent, semantically faithful, and clinically interpretable across synthetic and real robotic surgical datasets, providing a principled route to robust diffusion-based restoration under limited paired supervision.

LPNSR: Prior-Enhanced Diffusion Image Super-Resolution via LR-Guided Noise Prediction

cs.AI, q-bio.NC updates on arXiv.org

By: Shuwei Huang · Shizhuo Liu · Zijun Wei

25 March 2026 at 12:00

arXiv:2603.21045v2 Announce Type: replace-cross Abstract: Diffusion-based image super-resolution (SR), which aims to reconstruct high-resolution (HR) images from corresponding low-resolution (LR) observations, faces a fundamental trade-off between inference efficiency and reconstruction quality. The state-of-the-art residual-shifting diffusion framework achieves efficient 4-step inference, yet suffers from severe performance degradation in compact sampling trajectories. This is mainly attributed to two core limitations: the inherent suboptimality of unconstrained random Gaussian noise in intermediate steps, which leads to error accumulation and insufficient LR prior guidance, and the initialization bias caused by naive bicubic upsampling. In this paper, we propose LPNSR, a prior-enhanced efficient diffusion framework to address these issues. We first mathematically derive the closed-form analytical solution of the optimal intermediate noise for the residual-shifting diffusion paradigm, and accordingly design an LR-guided multi-input-aware noise predictor to replace random Gaussian noise, embedding LR structural priors into the reverse process while fully preserving the framework's core efficient residual-shifting mechanism. We further mitigate initial bias with a high-quality pre-upsampling network to optimize the diffusion starting point. With a compact 4-step trajectory, LPNSR can be optimized in an end-to-end manner. Extensive experiments demonstrate that LPNSR achieves state-of-the-art perceptual performance on both synthetic and real-world datasets, without relying on any large-scale text-to-image priors. The source code of our method can be found at https://github.com/Faze-Hsw/LPNSR.

Multimodal electron microscopy of halide perovskite interfacial dynamics

Nature - Issue - nature.com science feeds

By: Xinjuan Li · Qichun Gu · Wei Huang · Simon M. Fairclough · Richard H. Friend · Samuel D. Stranks · Tianjun Liu · Caterina Ducati

11 March 2026 at 08:00

Nature, Published online: 11 March 2026; doi:10.1038/s41586-026-10238-8

A multimodal in situ electron microscopy approach enables direct visualization of structural and chemical evolution in a working halide perovskite light-emitting diode with nanometre precision.

Geodesic Gradient Descent: A Generic and Learning-rate-free Optimizer on Objective Function-induced Manifolds

cs.AI, q-bio.NC updates on arXiv.org

By: Liwei Hu · Guangyao Li · Wenyong Wang · Xiaoming Zhang · Yu Xiang

10 March 2026 at 12:00

arXiv:2603.06651v1 Announce Type: cross Abstract: Euclidean gradient descent algorithms barely capture the geometry of objective function-induced hypersurfaces and risk driving update trajectories off the hypersurfaces. Riemannian gradient descent algorithms address these issues but fail to represent complex hypersurfaces via a single classic manifold. We propose geodesic gradient descent (GGD), a generic and learning-rate-free Riemannian gradient descent algorithm. At each iteration, GGD uses an n-dimensional sphere to approximate a local neighborhood on the objective function-induced hypersurface, adapting to arbitrarily complex geometries. A tangent vector derived from the Euclidean gradient is projected onto the sphere to form a geodesic, ensuring the update trajectory stays on the hypersurface. Parameter updates are performed using the endpoint of the geodesic. The maximum step size of the gradient in GGD is equal to a quarter of the arc length on the n-dimensional sphere, thus eliminating the need for a learning rate. Experimental results show that compared with the classic Adam algorithm, GGD achieves test MSE reductions ranging from 35.79% to 48.76% for fully connected networks on the Burgers' dataset, and cross-entropy loss reductions ranging from 3.14% to 11.59% for convolutional neural networks on the MNIST dataset.

DreamSAC: Learning Hamiltonian World Models via Symmetry Exploration

cs.AI, q-bio.NC updates on arXiv.org

By: Jinzhou Tang · Fan Feng · Minghao Fu · Wenjun Lin · Biwei Huang · Keze Wang

10 March 2026 at 12:00

arXiv:2603.07545v1 Announce Type: cross Abstract: Learned world models excel at interpolative generalization but fail at extrapolative generalization to novel physical properties. This limitation arises because they learn statistical correlations rather than the environment's underlying generative rules, such as physical invariances and conservation laws. We argue that learning these invariances is key to robust extrapolation. To achieve this, we first introduce \textbf{Symmetry Exploration}, an unsupervised exploration strategy where an agent is intrinsically motivated by a Hamiltonian-based curiosity bonus to actively probe and challenge its understanding of conservation laws, thereby collecting physically informative data. Second, we design a Hamiltonian-based world model that learns from the collected data, using a novel self-supervised contrastive objective to identify the invariant physical state from raw, view-dependent pixel observations. Our framework, \textbf{DreamSAC}, trained on this actively curated data, significantly outperforms state-of-the-art baselines in 3D physics simulations on tasks requiring extrapolation.

Solution to the 10th ABAW Expression Recognition Challenge: A Robust Multimodal Framework with Safe Cross-Attention and Modality Dropout

cs.AI, q-bio.NC updates on arXiv.org

By: Jun Yu · Naixiang Zheng · Guoyuan Wang · Yunxiang Zhang · Lingsi Zhu · Jiaen Liang · Wei Huang · Shengping Liu

10 March 2026 at 12:00

arXiv:2603.08034v1 Announce Type: cross Abstract: Emotion recognition in real-world environments is hindered by partial occlusions, missing modalities, and severe class imbalance. To address these issues, particularly for the Affective Behavior Analysis in-the-wild (ABAW) Expression challenge, we propose a multimodal framework that dynamically fuses visual and audio representations. Our approach uses a dual-branch Transformer architecture featuring a safe cross-attention mechanism and a modality dropout strategy. This design allows the network to rely on audio-based predictions when visual cues are absent. To mitigate the long-tail distribution of the Aff-Wild2 dataset, we apply focal loss optimization, combined with a sliding-window soft voting strategy to capture dynamic emotional transitions and reduce frame-level classification jitter. Experiments demonstrate that our framework effectively handles missing modalities and complex spatiotemporal dependencies, achieving an accuracy of 60.79% and an F1-score of 0.5029 on the Aff-Wild2 validation set.

RADAR: Learning to Route with Asymmetry-aware DistAnce Representations

cs.AI, q-bio.NC updates on arXiv.org

By: Hang Yi · Ziwei Huang · Yining Ma · Zhiguang Cao

5 March 2026 at 13:00

arXiv:2603.03388v1 Announce Type: cross Abstract: Recent neural solvers have achieved strong performance on vehicle routing problems (VRPs), yet they mainly assume symmetric Euclidean distances, restricting applicability to real-world scenarios. A core challenge is encoding the relational features in asymmetric distance matrices of VRPs. Early attempts directly encoded these matrices but often failed to produce compact embeddings and generalized poorly at scale. In this paper, we propose RADAR, a scalable neural framework that augments existing neural VRP solvers with the ability to handle asymmetric inputs. RADAR addresses asymmetry from both static and dynamic perspectives. It leverages Singular Value Decomposition (SVD) on the asymmetric distance matrix to initialize compact and generalizable embeddings that inherently encode the static asymmetry in the inbound and outbound costs of each node. To further model dynamic asymmetry in embedding interactions during encoding, it replaces the standard softmax with Sinkhorn normalization that imposes joint row and column distance awareness in attention weights. Extensive experiments on synthetic and real-world benchmarks across various VRPs show that RADAR outperforms strong baselines on both in-distribution and out-of-distribution instances, demonstrating robust generalization and superior performance in solving asymmetric VRPs.

DeepFusion: Accelerating MoE Training via Federated Knowledge Distillation from Heterogeneous Edge Devices

cs.AI, q-bio.NC updates on arXiv.org

By: Songyuan Li · Jia Hu · Ahmed M. Abdelmoniem · Geyong Min · Haojun Huang · Jiwei Huang

17 February 2026 at 13:00

arXiv:2602.14301v1 Announce Type: cross Abstract: Recent Mixture-of-Experts (MoE)-based large language models (LLMs) such as Qwen-MoE and DeepSeek-MoE are transforming generative AI in natural language processing. However, these models require vast and diverse training data. Federated learning (FL) addresses this challenge by leveraging private data from heterogeneous edge devices for privacy-preserving MoE training. Nonetheless, traditional FL approaches require devices to host local MoE models, which is impractical for resource-constrained devices due to large model sizes. To address this, we propose DeepFusion, the first scalable federated MoE training framework that enables the fusion of heterogeneous on-device LLM knowledge via federated knowledge distillation, yielding a knowledge-abundant global MoE model. Specifically, DeepFusion features each device to independently configure and train an on-device LLM tailored to its own needs and hardware limitations. Furthermore, we propose a novel View-Aligned Attention (VAA) module that integrates multi-stage feature representations from the global MoE model to construct a predictive perspective aligned with on-device LLMs, thereby enabling effective cross-architecture knowledge distillation. By explicitly aligning predictive perspectives, VAA resolves the view-mismatch problem in traditional federated knowledge distillation, which arises from heterogeneity in model architectures and prediction behaviors between on-device LLMs and the global MoE model. Experiments with industry-level MoE models (Qwen-MoE and DeepSeek-MoE) and real-world datasets (medical and finance) demonstrate that DeepFusion achieves performance close to centralized MoE training. Compared with key federated MoE baselines, DeepFusion reduces communication costs by up to 71% and improves token perplexity by up to 5.28%.

Governing AI Forgetting: Auditing for Machine Unlearning Compliance

cs.AI, q-bio.NC updates on arXiv.org

By: Qinqi Lin · Ningning Ding · Lingjie Duan · Jianwei Huang

17 February 2026 at 13:00

arXiv:2602.14553v1 Announce Type: cross Abstract: Despite legal mandates for the right to be forgotten, AI operators routinely fail to comply with data deletion requests. While machine unlearning (MU) provides a technical solution to remove personal data's influence from trained models, ensuring compliance remains challenging due to the fundamental gap between MU's technical feasibility and regulatory implementation. In this paper, we introduce the first economic framework for auditing MU compliance, by integrating certified unlearning theory with regulatory enforcement. We first characterize MU's inherent verification uncertainty using a hypothesis-testing interpretation of certified unlearning to derive the auditor's detection capability, and then propose a game-theoretic model to capture the strategic interactions between the auditor and the operator. A key technical challenge arises from MU-specific nonlinearities inherent in the model utility and the detection probability, which create complex strategic couplings that traditional auditing frameworks do not address and that also preclude closed-form solutions. We address this by transforming the complex bivariate nonlinear fixed-point problem into a tractable univariate auxiliary problem, enabling us to decouple the system and establish the equilibrium existence, uniqueness, and structural properties without relying on explicit solutions. Counterintuitively, our analysis reveals that the auditor can optimally reduce the inspection intensity as deletion requests increase, since the operator's weakened unlearning makes non-compliance easier to detect. This is consistent with recent auditing reductions in China despite growing deletion requests. Moreover, we prove that although undisclosed auditing offers informational advantages for the auditor, it paradoxically reduces the regulatory cost-effectiveness relative to disclosed auditing.

ESPO: Entropy Importance Sampling Policy Optimization

cs.AI, q-bio.NC updates on arXiv.org

By: Yuepeng Sheng · Yuwei Huang · Shuman Liu · Anxiang Zeng · Haibo Zhang

17 February 2026 at 13:00

arXiv:2512.00499v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has become a central component of post-training for large language models (LLMs), particularly for complex reasoning tasks that require stable optimization over long generation horizons. However, achieving performance at scale often introduces a fundamental trade-off between training stability and training efficiency. Token-level optimization applies fine-grained updates at the individual units, but is prone to high variance in gradient estimation, which can result in unstable training dynamics. In contrast, Sequence-level optimization often relies on aggressive clipping mechanisms to ensure stable updates. However, such design may discard a large fraction of valid training samples, leading to inefficient gradient utilization and reduced training efficiency. We refer to this phenomenon as gradient underutilization. In this work, we propose Entropy Importance Sampling Policy Optimization (ESPO), a novel framework that aims to combine fine-grained updates with stable training. ESPO decomposes sequences into groups based on predictive entropy, enabling (1) Entropy Grouping Importance Sampling to capture intra-sequence heterogeneity, and (2) Entropy Adaptive Clipping to dynamically allocate trust regions based on model uncertainty. Extensive experiments on mathematical reasoning benchmarks demonstrate that ESPO not only accelerates convergence but also achieves state-of-the-art performance, notably improving accuracy on the challenging mathematical benchmarks.

Reading view