跳到正文
arXiv cs.AI· Ziqi Miao, Haonan Jia, Lijun Li, Chen Qian, Yuan Xiong, Wenting Yan, Yongyao Wang, Zaoying Huang, Ken Deng, Jing Shao·· 2 小时前AI 评分42

PRCO:感知-推理协同进化的多模态推理 RLVR 框架

Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning

AI 导读

研究者提出 PRCO(Perception-Reasoning Coevolution),一种双角色 RLVR 框架,用共享策略让 Observer 生成面向问题的证据描述、Solver 基于该描述预测答案,并采用角色专属奖励信号解决传统 RLVR 共享奖励导致的视觉证据提取瓶颈。在 8 个多模态推理基准上,PRCO 相比基础模型平均准确率提升超 7 个点,优于此前开源 RL 调优基线。

正文 · 原文

View PDF HTML (experimental)

Abstract:Reinforcement learning with verifiable rewards (RLVR) has substantially enhanced the reasoning capabilities of multimodal large language models (MLLMs). However, existing RLVR approaches typically rely on outcome-driven optimization that updates both perception and reasoning using a shared reward based solely on the final answer. This shared reward blurs credit assignment, frequently improving reasoning patterns while failing to reliably enhance the accuracy of upstream visual evidence extraction. To address this perception bottleneck, we introduce PRCO (Perception-Reasoning Coevolution), a dual-role RLVR framework with a shared policy. PRCO consists of two cooperative roles: an Observer that generates an evidence caption tailored to the question and a Solver that predicts the final answer based on this caption. Crucially, PRCO employs role-specific reward signals: the Solver is optimized using verifiable outcome rewards on the final answer, while the Observer receives a utility reward derived from the Solver's downstream success. Extensive experiments across eight challenging multimodal reasoning benchmarks demonstrate that PRCO yields consistent improvements across model scales by over 7 points on average accuracy compared to the base model, outperforming prior open-source RL-tuned baselines.
Comments: 21 pages, 15 figures, 6 tables
Subjects: Artificial Intelligence (cs.AI)
Cite as: arXiv:2603.28618 [cs.AI]
  (or arXiv:2603.28618v3 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2603.28618

arXiv-issued DOI via DataCite

Journal reference: EMNLP 2026

Submission history

From: Miao Ziqi [view email]
[v1] Mon, 30 Mar 2026 16:03:56 UTC (2,178 KB)
[v2] Thu, 9 Apr 2026 14:15:05 UTC (2,177 KB)
[v3] Tue, 17 Nov 2026 12:57:34 UTC (2,175 KB)

来源:arXiv cs.AI · arxiv.org