Jiafeng Liang, Xianhao Ou, Tao Ren, Rongchuan Mu, Runxuan Liu, Dandan Tu, Ming Liu, See-Kiong Ng 等 9 位作者
Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for enhancing the video reasoning capability of Large Multimodal Models (LMMs). Despite its empirical effectiveness, we show that outcome-level optimization can introduce an undesirable shortcut amplification effect, whereby models increasingly favor text-prior answers even when they are not supported by the visual evidence. To characterize this behavior, we investigate the training dynamics of video RLVR through the lens of perceptual uncertainty, analyzing how predictive uncertainty evolves under video-conditioned and text-only contexts. Our analysis identifies a perceptual uncertainty collapse phenomenon, in which RL training continuously sharpens the model's answer distribution while the uncertainty reduction attributable to visual evidence progressively diminishes. A further gradient-direction analysis reveals that such confidence enhancement becomes increasingly less aligned with visually grounded uncertainty reduction and is instead dominated by text-prior compression. Motivated by these findings, we propose Bifocal Self-distillation Policy Optimization (BSPO), a self-distillation framework that regulates policy updates with two complementary distributional signals. Specifically, BSPO derives empirical reasoning principles from intra-group correct and incorrect rollouts to identify reasoning-sensitive tokens, and introduces shortcut-aware visual perturbation to penalize predictions that remain invariant when task-relevant visual evidence is weakened. By adaptively reweighting the original RL advantage, BSPO preserves the task-solving benefits of RLVR while promoting updates that are both reasoning-sensitive and visually grounded. Extensive experiments validate the effectiveness of BSPO in improving video reasoning performance and strengthening visual faithfulness.
Rongchuan Mu, Zexin Wang Harbin Institute of Technology, Qianyu Wang, MingHua Ma, Zekun Wang Alibaba Group, Ming Liu Harbin Institute of Technology, Bing Qin Harbin Institute of Technology
Large Reasoning Models (LRMs) excel at complex problem-solving but frequently overlook specific instruction constraints. Existing alignment methods struggle to balance general reasoning with instruction-following (IF), hindered by dependency on teacher models, reward hacking, and reasoning-answer inconsistencies. We propose PARIF, a two-stage curriculum learning framework based on Reinforcement Learning from Verifiable Rewards (RLVR) to enhance both IF and general reasoning capabilities. The framework employs a correctness proxy across different stages to mitigate reward hacking. Stage I employs a dynamic weighting strategy simultaneously to optimize the model’s reasoning paradigm regarding constraints. Stage II introduces Decoupled-GRPO, which builds upon the first stage to enhance the logical consistency between the reasoning process and the final answer, enabling the model to better leverage its optimized reasoning paradigm. To support the framework, we curate 26,000 high-quality instructions featuring diverse constraints. Extensive experiments demonstrate PARIF’s effectiveness: our 7B model achieves a remarkable 21.25% relative average improvement to the original model across six representative IF tasks, while our 8B model outperforms leading models like DeepSeek-V3 on these IF tasks, effectively pushing the Pareto frontier of instruction following and reasoning for models of comparable scale. We open-source our code and models to facilitate future research.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Runxuan Liu, Xianhao Ou Harbin Institute of Technology, Xinyan Ma Harbin Institute of Technology, Jiyuan Wang, Jiafeng Liang Harbin Institute of Technology, Jiaqi Li Tianjin Normal University, Tao He Harbin Institute of Technology, Zheng Chu Harbin Institute of Technology 等 16 位作者
Long Chain-of-Thought (LCoT), achieved by Reinforcement Learning with Verifiable Rewards (RLVR), has proven effective in enhancing the reasoning capabilities of Large Language Models (LLMs). However, reasoning in current LLMs is primarily generated as plain text, where performing semantic evaluation on such unstructured data creates a computational bottleneck during training. Despite RLVR-based optimization, existing methods still suffer from coarse-grained supervision, reward hacking, high training costs, and poor generalization.To address these issues, we propose the Graph Reasoning Paradigm (GRP), which realizes structured and symbolic reasoning, implemented via graph-structured representations with step-level cognitive labels. Building upon GRP, we further design Process-Aware Stratified Clipping Group Relative Policy Optimization (PASC-GRPO), which leverages structured evaluation to replace semantic evaluation, achieves process-aware verification through graph-structured outcome rewards, and mitigates reward hacking via stratified clipping advantage estimation. Experiments demonstrate significant improvements across mathematical reasoning and code generation tasks. Data, models, and code will be released later.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表