RAPO: reflection anchors curb visual information decay in long-chain multimodal reasoning

Xuan Gong · hf · 2026-11-17

Long CoT reasoning boosts large vision-language models, but visual information fades during generation, limiting long-horizon multimodal reasoning. The paper derives an information-theoretic lower bound on the downstream visual gain of a one-step intervention, hinging on local branching room (token entropy) and downstream propagation potential (suffix divergence from a vision-marginalized reference). Guided by this, Reflection-Anchor Policy Optimization (RAPO), a GRPO-based method, selects high-entropy reflection anchors and optimizes a chain-masked finite-window KL surrogate. RAPO substantially outperforms strong baselines across reasoning-intensive and general benchmarks on multiple LVLM backbones, with anchors enriched at visually sensitive decision points.

Original post →

More from Multimodal

Multimodal channel →