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.
More from Multimodal
- Single continuous-shot hyperbike chase video prompt shared — umesh_ai · 2026-11-17
- Character design shown in both stylized 2D illustration and matching 3D render — azed_ai · 2026-11-17
- Solo dev expands RTX 4090 text-to-image benchmarks with first GenEval for FLUX.2 4B — hamshrew · 2026-11-17
- AI music video with Claude and Suno: 435 clips, 30 cuts, 3 days — full breakdown — MuffDivers2_ · 2026-11-17
- Claude Fable 5.2's artwork revealed: a striking abandoned-pool illustration series — repligate · 2026-11-17
- Ink Studio launches free on GitHub for 3D pixel-art Gaussian splat animation — ring_hyacinth · 2026-11-17