Research map / Vision-language and multimodal / Multimodal LLMs

MiCo: Multi-image Contrast for Reinforcement Visual Reasoning

NeurIPS 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 2 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 MiCo, a self-supervised RL method that trains VLMs to compare image triplets (two augmented views of one image plus a similar distinct image) with rule-based GRPO-style rewards and no human QA pairs, improving multi-image reasoning.

Problem

VLMs struggle to link fine-grained visual cues across multiple images, and rule-based RL for them normally requires manually curated question-answer pairs.

Approach

Use inherent image constraints as supervision: build triplets from two augmented views of one image and a similar but different image (from video frames or image-editing pairs), train the model to reason about same/different with rule-based RL, and use Augmented GRPO that samples rollouts under weak augmentation and optimizes under strong augmentation.

Key results

• On VLM2-Bench, MiCo-7B-CoT reaches 61.06 average accuracy, above GPT-4o (60.36) and the Qwen2.5-VL-7B-CoT base (48.91).
• MiCo-7B improves over Qwen2.5VL-7B on MuirBench (60.53 vs 58.43) and BLINK (57.23 vs 55.54), with single-image gains on MMStar, MMMU and MathVista of about +0.7 to +1.5 points.
• Ablations show CoT RL (62.13 general / 65.53 object / 57.18 person) beats SFT and no-CoT RL, and strong-augmentation optimization with weak-augmentation rollouts beats (strong, strong) and (weak, weak).

Measured effects reported in the paper

Surprising findings

• CoT reasoning gives limited or negative gains on person/face sub-tasks of VLM2-Bench, which the authors attribute to face identity cues being hard to verbalize.
• Importance-sampling reweighting and confidence-based reward reweighting did not improve performance.

Limitations

• Limited effectiveness on specialized tasks such as face verification, visual math and spatial understanding that need structured priors.
• Lags behind models trained with curated reasoning data on scene understanding and forensic detection, and on relative depth.
• T/F answer format leaves a 12.5% chance of correct guess for triplets, which the authors could not mitigate with confidence reweighting.

Experimental setup

Datasets: OmniEdit Vidgen-1M VLM2-Bench MuirBench BLINK MMStar MMMU HallusionBench MathVista

Benchmarks: VLM2-Bench MuirBench BLINK MMStar MMMU HallusionBench MathVista

Metrics: accuracy average accuracy across VLM2-Bench tracks normalized error (Cnt) GPT-4o-scored open-ended (VID)

Baselines: Qwen2.5-VL-7B GPT-4o MM-Eureka-7B NoisyRollout-7B ThinkLite-VL-7B VLAA-Thinker-7B Qwen2.5-VL-7B-CoT LLaVA-OneVision LLaVA-Video-7B InternVL2.5-8B

Builds on: GRPO DeepSeek-R1 self-supervised contrastive learning NoisyRollout

Closest papers by meaning

• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.91)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.90)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.90)
• HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (NeurIPS 2026 · similarity 0.90)
• Latent Visual Reasoning (ICLR 2026 · similarity 0.90)