Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
Abstract
Long chain-of-thought (CoT) reasoning improves large vision--language models, but visual information often fades during generation, limiting long-horizon multimodal reasoning. Existing methods either re-inject vision at inference or train policies for stronger grounding, but where to intervene relies on perception heuristics rather than principled gain analysis, and how local visual influence propagates remains implicit. We study this problem from an information-theoretic standpoint and derive a lower bound on the downstream visual gain of a one-step intervention, which suggests two factors: local branching room (token entropy) and downstream visual propagation potential (suffix divergence from a vision-marginalized reference). Guided by this analysis, we propose reflection-anchor policy optimization (RAPO), a GRPO-based policy optimization method that selects high-entropy reflection anchors and optimizes a chain-masked finite-window KL surrogate for downstream visual dependence. Experiments on reasoning-intensive and general-domain benchmarks show that RAPO delivers substantial gains over strong baselines across multiple LVLM backbones. Mechanism analyses further indicate that reflection anchors are enriched for visually sensitive decision points and that RAPO increases contrastive visual-dependence signals along generated trajectories.
Community
RAPO: Keeping Visual Evidence in the Loop during Long-Chain Reasoning
Excited to share our NeurIPS 2026 paper, Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning!
As multimodal reasoning chains grow longer, models can gradually lose their dependence on visual evidence and rely increasingly on generated text.
We introduce Reflection-Anchor Policy Optimization (RAPO), which identifies high-entropy reasoning positions as sparse reflection anchors and trains the policy to preserve visual influence in downstream reasoning through a chain-masked KL objective.
Key highlights:
- Principled motivation: An information-theoretic analysis of local branching and downstream visual propagation.
- Propagation-aware optimization: Strengthening visual dependence beyond individual reasoning steps.
- Stronger performance: On Qwen3-VL-2B-Instruct, RAPO improves the six-benchmark average from 40.79 (GRPO) to 43.01 (mean@8 accuracy).
- Standard inference: No additional visual re-injection or anchor selection during decoding.
Paper: https://arxiv.org/abs/2605.09614
Project: https://coder-gx.github.io/RAPO/
GitHub: https://github.com/coder-gx/RAPO
We welcome feedback and discussion!
Get this paper in your agent:
hf papers read 2605.09614 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper