cs.CVMar 28, 2026

Bridging Visual Representation and Reinforcement Learning from Verifiable Rewards in Large Vision-Language Models

Authors: Yuhang Han, Yuyang Wu, Zhengbo Jiao, Yiyu Wang, Xuyang Liu, Shaobo Wang, Hanlin Xu, Xuming Hu, 

Organizations: EPIC Lab, SJTU · HKUST (GZ) · Huawei

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Reinforcement Learning from Verifiable Rewards (RLVR) has substantially enhanced the reasoning capabilities of large language models in abstract reasoning tasks. However, its application to Large Vision-Language Models (LVLMs) remains constrained by a structural representational bottleneck. Existing approaches generally lack explicit modeling and effective utilization of visual information, preventing visual representations from being tightly coupled with the reinforcement learning optimization process and thereby limiting further improvements in multimodal reasoning performance. To address this limitation, we propose KAWHI (Key-Region Aligned Weighted Harmonic Incentive), a plug-and-play reward reweighting mechanism that explicitly incorporates structured visual information into uniform reward policy optimization methods (e.g., GRPO and GSPO). The method adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps. Extensive empirical evaluations on diverse reasoning benchmarks substantiate KAWHI as a general-purpose enhancement module, consistently improving the performance of various uniform reward optimization methods. Project page: KAWHI (https://kawhiiiileo.github.io/KAWHI_PAGE/)

Figures & tables

Figure 1: (a) Error distribution across five categories: VP (visual perception error), AO (answer only), CE (calculation error), RAE (rule application error), and Others. (b) Two VP cases from MathVerse. Red highlights indicate visual misinterpretation.
Figure 2: Overview of the KAWHI mechanism. Critical regions are selected using the SGUF algorithm (see section 3.2 for details). The decoded sequence denotes the textual output generated from response tokens, segmented by the paragraph delimiter \n\n .
Figure 3: Vision-Critical head identification via global ablation on the MME benchmark. Here, b denotes the performance score of the baseline models (Qwen2.5-VL-7B-Instruct [ 5 ] and Qwen3-VL-4B-Instruct [ 4 ] ).
MethodMathVistaMathVerseMathVisionWeMathAvg
Qwen2.5-VL-7B-Instruct
Base67.4745.5623.3654.3247.68
fine-grained reward
Step-GRPO66.87 −- 0.6044.32 −- 1.2424.76 ++ 1.4055.32 ++ 1.0047.82 ++ 0.14
FT-RL67.68 ++ 0.2146.88 ++ 1.3226.12 ++ 2.7656.78 ++ 2.4649.37 ++ 1.69
VPPO68.42 ++ 0.9547.21 ++ 1.6528.42 ++ 5.0657.12 ++ 2.8050.29 ++ 2.61
Table 1: Main results on Qwen2.5-VL-7B-Instruct and Qwen3-VL-4B-Instruct . +Δ+\Delta / −Δ-\Delta denote changes vs. baselines; (+X.XX) shows gains over base RL methods. Bold = best per model size; gray rows = ours.
MethodCharXiv DescCharXiv ReaChart QAChart QAProChart Mimic
Qwen2.5-VL-7B-Instruct67.942.187.441.940.2
+ GRPO70.1 ++ 2.244.3 ++ 2.288.3 ++ 0.943.7 ++ 1.842.8 ++ 2.6
+ GRPO + KAWHI (Ours)72.2 ++ 4.346.4 ++ 4.389.1 ++ 1.745.3 ++ 3.445.1 ++ 4.9
Table 2: Main results on chart understanding benchmarks . +Δ+\Delta / −Δ-\Delta denote changes vs. baselines; Bold = best per benchmark; gray rows = ours.
AblationExperimentMathVerse (%) ↑\uparrowMathVision (%) ↑\uparrowAverage
–GRPO45.94 ↓1.6228.29 ↓0.9937.11
–GRPO+ KAWHI (Ours)47.5629.2838.42
① Region SelectionAll vision tokens46.58 ↓0.9828.45 ↓0.8337.52
Random46.34 ↓1.2228.98 ↓0.3037.66
Inverse selection45.74 ↓1.8227.82 ↓1.4636.78
② Reward MetricKey–Key47.08 ↓0.4828.95 ↓0.3338.02
Table 3: Ablation on Qwen2.5-VL-7B-Instruct. Results on MathVerse and MathVision. ↓\downarrow X.XX indicates drop vs. GRPO+ KAWHI .
Figure 4: Validation of VisionZip as an effective alternative to SGUF under 50% visual tokens.
Figure 5: Computational Efficiency Analysis of KAWHI . Component-wise isolation analysis using GRPO to assess the additional computational complexity introduced by our method.
Figure 6: Qualitative Analysis of Weight Allocation in KAWHI . Results are illustrated using cases from Geo3K [ 33 ] . Color intensity represents weight magnitude, where darker colors indicate larger weights, illustrating the allocation behavior of KAWHI .
Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Sensitivity Analysis of the SGUF Hyperparameter skip_ratio (Qwen2.5-VL-7B-Instruct with GRPO+ KAWHI on MathVerse and MathVision)
More in this topic

VLM Reasoning

Vision-language model reasoning studies how multimodal systems use visual evidence and language to answer questions, plan, and explain.

See all 853 papers

Explore similar work

CardsList

  1. MIRL: Mutual Information-Guided Reinforcement Learning for Vision-Language Models

    May 2, 2026Yin Zhang, Jiaxuan Zhao, Zonghan Wu +5VLM ReasoningVision-Language Models

    School of Mathematics, Tianjin University, Tianjin, China · Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · Shanghai Advanced Institute of Finance (SAIFS), East China Normal University, Shanghai, China +4

  2. Anchoring What Matters: A Dual-Level Learning Framework for Visually-Grounded Multimodal Reasoning

    Sep 16, 2026Xinxin Song, Siyuan Li, Tingxiong Xiao +1Credit Assignment in RLExperience Replay

    Department of Automation, Tsinghua University · Institute for Brain and Cognitive Science, Tsinghua University

  3. Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models

    Apr 2, 2026Zekai Ye, Qiming Li, Xiaocheng Feng +6RL for VLMsLarge Vision-Language Models

    Harbin Institute of Technology · Peng Cheng Laboratory · Huawei Technologies Co., Ltd