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

TL;DR

This work tackles the challenge of multi-image reasoning in multimodal LLMs by introducing CINEMA, a Cognition-Inspired Meta-Action framework that decomposes reasoning into five explicit actions (Global, Focus, Hint, Think, Answer). It pairs this with a Retrieval-Based Tree Sampling strategy to generate diverse cold-start trajectories and a two-stage reinforcement learning regimen (Diversity-Preserving Strategy followed by annealed DAPO) to maintain exploration while consolidating performance. A large training corpus (57k cold-start and 58k RL instances) across multi-image, multi-frame, and single-image tasks supports learning, and extensive benchmarks show state-of-the-art results on multiple multi-image and video datasets, including surpassing GPT-4o on MUIR and MVMath. The results demonstrate strong generalization to single-image tasks as well, highlighting the framework’s robustness, scalability, and potential to enhance visual reasoning in real-world applications.

Abstract

While Multimodal Large Language Models (MLLMs) excel at single-image understanding, they exhibit significantly degraded performance in multi-image reasoning scenarios. Multi-image reasoning presents fundamental challenges including complex inter-relationships between images and scattered critical information across image sets. Inspired by human cognitive processes, we propose the Cognition-Inspired Meta-Action Framework (CINEMA), a novel approach that decomposes multi-image reasoning into five structured meta-actions: Global, Focus, Hint, Think, and Answer which explicitly modeling the sequential cognitive steps humans naturally employ. For cold-start training, we introduce a Retrieval-Based Tree Sampling strategy that generates high-quality meta-action trajectories to bootstrap the model with reasoning patterns. During reinforcement learning, we adopt a two-stage paradigm: an exploration phase with Diversity-Preserving Strategy to avoid entropy collapse, followed by an annealed exploitation phase with DAPO to gradually strengthen exploitation. To train our model, we construct a dataset of 57k cold-start and 58k reinforcement learning instances spanning multi-image, multi-frame, and single-image tasks. We conduct extensive evaluations on multi-image reasoning benchmarks, video understanding benchmarks, and single-image benchmarks, achieving competitive state-of-the-art performance on several key benchmarks. Our model surpasses GPT-4o on the MUIR and MVMath benchmarks and notably outperforms specialized video reasoning models on video understanding benchmarks, demonstrating the effectiveness and generalizability of our human cognition-inspired reasoning framework.

Figures (6)

• Figure 1: Overview of CINEMA.
• Figure 2: Pass@K performance.
• Figure 3: Results about RQ2 and RQ3.
• Figure 4: Results about RQ5.
• Figure 5: Case study.