Research map / Vision-language and multimodal / Multimodal LLMs

Mimic Human Cognition, Master Multi-Image Reasoning: A Meta-Action Framework for Enhanced Visual Understanding

CVPR 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 CINEMA, a five-meta-action (Global, Focus, Hint, Think, Answer) reasoning framework for multi-image MLLMs, trained with retrieval-based tree sampling for cold start and a two-stage RL (diversity-preserving strategy then annealed DAPO) on Qwen2.5-VL-7B.

Problem

MLLMs degrade on multi-image reasoning due to complex inter-image relationships and critical information scattered across images.

Approach

Decompose multi-image reasoning into five cognition-inspired meta-actions; generate cold-start trajectories via student-teacher retrieval-based tree sampling; and preserve meta-action-level diversity in an RL exploration stage before annealed DAPO exploitation.

Key results

• Overall multi-image/video average rises from 48.2 (Qwen2.5-VL-7B) to 54.3, with MUIR 71.6 vs GPT-4o 68.0 and MVMath 36.9 vs GPT-4o 32.1.
• On single-image benchmarks the overall average reaches 45.6 vs 40.8 for Qwen2.5-VL-7B, with MM-Math 43.8 as the best reported.
• Ablation removing the think action drops MUIR from 71.0 to 60.1; two-trajectory cold start improves RL MUIR from 65.1 (single trajectory) to 71.6.

Surprising findings

• Directly prompting Qwen2.5-VL with meta-actions without training degrades MUIR from 57.9 to 33.8, showing the format must be learned.
• The DPS-only variant and the two-stage variant reach nearly identical overall scores (54.2 vs 54.3), so the annealing stage adds little on multi-image benchmarks.

Limitations

• Relies on GPT-4o as teacher for trajectory refinement, so cold-start data quality depends on a proprietary model.
• Several comparisons are mixed: the DPS-only and two-stage variants are nearly tied, and the method underperforms some baselines on MMIU and MathVista-style tasks.
• Text and paper contain many unfilled placeholders and incomplete sections, making the reported evaluation and methodology details partly unverifiable.

Experimental setup

Datasets: MUIR-style multi-image sources (ChartVQA, SlideVQA, ALFRED, Nuscenes, RecipeQA, IconQA, NLVR2, Spot-the-Diff, LRV, RAVEN video sources (STAR, NextQA, CLEVRER, Perception Test) single-image sources (CLEVR, M3CoT, Share-GRPO, GeoQA-R1V, AI2D, MMK12, Geometry3k, ScienceQA, GQA, COCO)

Benchmarks: MUIR MMIU Mantis-Eval MVMath EMMA MIRB MVBench VideoMME VideoMMMU MMMU-Pro M3CoT MM-IQ

Metrics: accuracy pass@k

Baselines: GPT-4o GPT-4V Gemini-1.5-Pro Qwen2.5-VL-7B InternVL2.5 8B VISC MIA-DPO VideoRFT TW-GRPO VideoR1 VLAA-Thinker VisionR1

Builds on: DAPO GRPO Qwen2.5-VL BGE EasyR1

Closest papers by meaning

• OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.91)
• CMMCoT: Enhancing Complex Multi-Image Comprehension via Multi-Modal Chain-of-Thought and Memory Augmentation (AAAI 2026 · similarity 0.91)
• MiCo: Multi-image Contrast for Reinforcement Visual Reasoning (NeurIPS 2025 · similarity 0.90)
• Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.90)
• Toward Cognitive Supersensing in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)
• Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · similarity 0.90)