Research map / Vision-language and multimodal / Multimodal LLMs

Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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 MARS, which adds mono-source rollouts as anchors in GRPO/DAPO advantage normalization so multi-source visual reasoning gains information rather than suffering interference from conflicting sources.

Problem

Multi-source visual reasoning with RLVR often underperforms mono-source reasoning when one source (e.g., infrared or depth) carries the dominant signal, because standard advantage normalization treats multi-source inputs as simple accumulation.

Approach

Generate one mono-source rollout per image and include these rewards in the group-wise mean/std used for advantage normalization of multi-source rollouts, so the baseline reflects the information gain (or conflict) of multi-source fusion relative to single sources; no extra policy or architecture is needed.

Key results

• Averaged over four multi-source tasks on Qwen2.5-VL-3B, GRPO+MARS improves Multi accuracy from 74.7 to 77.9 (+3.2).
• DAPO+MARS improves Multi average from 75.8 to 80.7 (+4.9), with Union also rising from 83.0 to 85.2.
• Gains are 4.4% on the 3B model and 2.8% on the 7B model on average, and under visual degradation the method drops only 0.5% under severe Gaussian noise.

Surprising findings

• Standard multi-source reasoning (Multi) often underperforms mono-source reasoning (Union) by a large margin, and supervised and naive RLVR methods can perform worse than strong single-source reasoning.
• Increasing the number of actual rollouts in GRPO (M+N) yields only about 0.2% gain at roughly 30% extra overhead, versus the mono-anchor approach.

Limitations

• Evaluated mainly on Qwen2.5-VL-3B and 7B with four specific multi-source datasets; generality to other modalities is not shown beyond these.
• Union is defined by taking the best single-source result at inference, which requires knowing which source is correct, so it is an oracle upper bound rather than a deployable method.
• Theoretical analysis is largely expectation-level and the claimed convergence benefits are not rigorously proven.

Experimental setup

Datasets: SpatialQA (depth) LLVIP (infrared) nuScenes (multi-view) OCR-VQA (text-rich)

Benchmarks: SpatialQA LLVIP nuScenes OCR-VQA

Metrics: accuracy mIoU

Baselines: Qwen2.5-VL-3B LLaVA-OV-7B R1-Onevision-7B VisionReasoner-7B SFT Chain-of-Thought GRPO DAPO

Builds on: GRPO DAPO DeepSeek-R1

Closest papers by meaning

• Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning (NeurIPS 2026 · similarity 0.88)
• OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.87)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.87)
• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.87)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
• UniVR: Thinking in Visual Space for Unified Visual Reasoning (NeurIPS 2026 · similarity 0.87)