Research map / Vision-language and multimodal / Multimodal LLMs

Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RAPO, a GRPO-based RL method that selects high-entropy "reflection anchors" and adds a branching-room-weighted chain-masked finite-window KL term to encourage sustained visual dependence in long-CoT LVLM reasoning.

Problem

Visual information fades during long chain-of-thought generation in LVLMs ("visual forgetting"); existing fixes either re-inject vision at inference or implicitly train for grounding, without principled selection of where to intervene or how influence propagates.

Approach

Information-theoretic lower bound on downstream visual gain of a one-step intervention, yielding two factors (native token entropy as branching room, and suffix divergence from a vision-marginalized reference). Practically: select top-20% entropy anchors and add a branching-weighted chain-masked finite-window KL term to GRPO, with no inference-time change.

Key results

• On Qwen3-VL-8B-Instruct, RAPO_G reaches 55.11 average vs 53.38 for GRPO and 53.85 for VPPO_G (base 49.56).
• On Qwen2-VL-7B-Instruct, Base+RAPO_D reaches 37.09 average (+21.34 over base), above TVC (29.94) and MINT-CoT (27.32).
• On Qwen3-VL-2B-Instruct, RAPO_G averages 43.01 vs 40.79 for GRPO; replacing chain masking with image-level masking drops it to 40.43.

Surprising findings

• Replacing chain masking with blanket image-level masking lowers the average score from 43.01 to 40.43, showing structured masking matters.
• Large KL coefficient (gamma=0.1) causes training collapse, and visual-token attention fails to distinguish correct from incorrect trajectories.

Limitations

• Theoretical bound is not directly computable; entropy selection is a proxy and the suffix-divergence term is approximated by a finite window.
• Gains on strong backbones (Qwen3-VL) are modest and described as controlled improvements under reduced headroom.
• Large KL coefficient can collapse training, requiring careful tuning.

Experimental setup

Datasets: ViRL

Benchmarks: MathVision MathVerse EMMA LogicVista MMMU-Pro RealWorldQA

Metrics: mean@8 accuracy

Baselines: GRPO PAPO VPPO MM-Eureka VL-Rethinker R1-ShareVL TVC MINT-CoT

Builds on: GRPO DAPO PAPO VPPO

Closest papers by meaning

• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.91)
• Dual-Latent Memory Routing for Vision-Language Reasoning (ICML 2026 · similarity 0.90)
• From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning (ICLR 2026 · similarity 0.90)
• On the Faithfulness of Visual Thinking: Measurement and Enhancement (NeurIPS 2026 · similarity 0.90)
• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.90)