Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

2026-06-02

AI Summary

Existing multimodal large language models lack systematic modeling of geometric structure, viewpoint variation, fine-grained correspondence, and occlusion relationships in wide-baseline matching tasks. To address this limitation, this work introduces ReasonMatch-Bench, a new benchmark, and proposes Dynamic Correspondence Reinforcement Learning (DCRL), which trains models by automatically mining verifiable wide-baseline image pairs from video and 3D data. DCRL pioneers a reinforcement learning framework that jointly integrates image-level progressive viewpoint reasoning and point-level correspondence curricula, effectively enhancing spatial reasoning capabilities without explicit chain-of-thought supervision. Experiments demonstrate that DCRL boosts the F1 score on the challenging subset of ReasonMatch-Bench from 37.2 to 84.0—approaching human-level performance—and exhibits strong transferability across multiple spatial reasoning tasks.

Abstract

Wide-baseline matching (WBM) requires integrating geometric understanding, viewpoint changes, fine-grained perception, and occlusion reasoning, making it a challenging testbed for spatial reasoning in multimodal large language models (MLLMs) deployed in physical environments. However, current MLLMs lack systematic evaluation and training frameworks for these capabilities. We introduce ReasonMatch-Bench, a benchmark stratified by viewpoint displacement and matching granularity across indoor, outdoor, and object-centric scenarios, and show that current MLLMs still struggle with fine-grained wide-baseline correspondence: on a difficult 90-sample subset, human annotators achieve 84.0 F1, while the best existing baseline reaches 37.2. To bridge this gap, we build a scalable data-generation pipeline that automatically extracts wide-baseline view pairs from large-scale video-3D corpora, including RGB-D videos and SfM reconstructions, yielding diverse and verifiable supervision. We further propose Dynamic Correspondence Reinforcement Learning (DCRL), which combines Image-Level Viewpoint Progression and Point-Level Correspondence Curriculum to improve WBM training through verifiable rewards without explicit CoT supervision. Extensive experiments show that DCRL substantially improves ReasonMatch-Bench and transfers to related spatial benchmarks, while maintaining general visual understanding performance with modest gains on several benchmarks.

Problem

• Wide-Baseline Matching
• Spatial Reasoning
• Multimodal Large Language Models
• Fine-grained Correspondence
• Viewpoint Displacement

Innovation

• Wide-Baseline Matching
• ReasonMatch-Bench
• Dynamic Correspondence Reinforcement Learning
• Spatial Reasoning
• Multimodal Large Language Models