Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models
This work addresses a foundational methodological flaw in RLVR for LVLMs, offering a novel fine-grained credit assignment framework to improve perception-grounded multimodal reasoning, though it is incremental in refining existing RLVR approaches.
The paper tackled the problem of Reinforcement Learning from Verifiable Rewards (RLVR) in Large Vision-Language Models (LVLMs) by addressing the flaw of distributing identical advantages across all tokens, which dilutes learning signals for visually-grounded reasoning. The result was the introduction of Perception-Grounded Policy Optimization (PGPO), which boosted models by 18.7% on average across seven multimodal reasoning benchmarks.
While Reinforcement Learning from Verifiable Rewards (RLVR) has advanced reasoning in Large Vision-Language Models (LVLMs), prevailing frameworks suffer from a foundational methodological flaw: by distributing identical advantages across all generated tokens, these methods inherently dilute the learning signals essential for optimizing the critical, visually-grounded steps of multimodal reasoning. To bridge this gap, we formulate \textit{Token Visual Dependency}, quantifying the causal information gain of visual inputs via the Kullback-Leibler (KL) divergence between visual-conditioned and text-only predictive distributions. Revealing that this dependency is highly sparse and semantically pivotal, we introduce Perception-Grounded Policy Optimization (PGPO), which is a novel fine-grained credit assignment framework that dynamically reshapes advantages at the token level. Through a threshold-gated, mass-conserving mechanism, PGPO actively amplifies learning signals for visually-dependent tokens while suppressing gradient noise from linguistic priors. Extensive experiments based on the Qwen2.5-VL series across seven challenging multimodal reasoning benchmarks demonstrate that PGPO boosts models by 18.7% on average. Both theoretical and empirical analyses confirm that PGPO effectively reduces gradient variance, prevents training collapse, and acts as a potent regularizer for robust, perception-grounded multimodal reasoning. Code will be published on https://github.com/Yzk1114/PGPO.