Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning

Fanhu Zeng, Zhicong Luo, Zefan Wang, You Li, Chi Chen, Maosong Sun

arXiv:2605.25437 · 2026-05-27 공개 · arXiv · PDF

reinforcement-learning grpo rlvr visual-reasoning dapo gradient-estimation information-gain modalities

Abstract

Visual reasoning through reinforcement learning with verifiable rewards (RLVR) has achieved remarkable progress. However, when dealing with multi-source inputs, existing approaches tend to treat them as a mere accumulation of information, lacking explicit mechanisms to distinguish whether integrating additional sources yields information gain or introduces interference. Therefore, they struggle to effectively model dynamic interaction when integrating multiple sources, particularly when they differ significantly in physical properties and semantics, e.g., infrared and depth, leading to inferior performance to mono-source reasoning when a certain source holds the dominant signal. To address this issue, we propose MARS, a novel mono-anchored multi-source reasoning framework that models each visual modality as an independent information source. Specifically, by treating mono-source rewards as dynamic anchors, our method explicitly incorporates the information gain introduced by multi-source fusion into advantage normalization and adaptively emphasizes mutual promotion between sources while suppressing potential noise or conflicts during RLVR. From theoretical analysis, our method effectively quantifies information gain introduced by multi-source integration in gradient estimation, enabling consistent modality regulation. Empirical results also show impressive 3.2% and 4.9% performance gains on GRPO and DAPO across diverse datasets, confirming effectiveness of our method.

한국어 요약

📋 한 줄 요약

**[Multi-Source Visual Reasoning / RLVR]** MARS가 mono-source reward를 dynamic anchor로 활용해 multi-source fusion의 information gain을 advantage normalization에 명시 통합, GRPO·DAPO에서 3.2%·4.9% 성능 향상.

🎯 핵심 기여도

💡 핵심 아이디어

Multi-source visual reasoning에서 단순 정보 누적은 dominant source의 성능을 약화시킬 수 있으며, mono-source reward를 dynamic anchor로 사용해 multi-source의 information gain을 advantage normalization에 명시 통합해야 source 간 mutual promotion과 conflict 억제를 동시 달성한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Multi-source RLVR의 dominant source 약화 문제를 명확 진단·해결, advantage normalization을 information gain 기반으로 재정의하는 일반 패턴, GRPO·DAPO 두 SOTA RL 알고리즘에 일관 적용. **한계**: Mono-source reward 평가의 계산 추가 비용, 매우 많은 source(>3) 확장은 추가 검증, vision modality 중심 — 다른 modality(음성·텍스트) 일반화 후속.

🚀 실용적 활용