Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
RAPO trains vision–language models to keep image information influential deeper into long reasoning chains.
The method targets reflection points where the model has several plausible next tokens and visual input could shape later decisions. Its training objective rewards downstream dependence on that input, guided by an information-theoretic analysis. The authors report substantial gains over strong baselines across multiple model backbones and reasoning-intensive and general-domain benchmarks. Their analyses also indicate stronger visual-dependence signals along generated reasoning chains, though the abstract provides no numerical results.
HF Daily Papers' note
The method targets reflection points where the model has several plausible next tokens and visual input could shape later decisions. Its training objective rewards downstream dependence on that input, guided by an information-theoretic analysis. The authors report substantial gains over strong baselines across multiple model backbones and reasoning-intensive and general-domain benchmarks. Their analyses also indicate stronger visual-dependence signals along generated reasoning chains, though the abstract provides no numerical results.
HF Daily Papers' note
score 4