Search papers, labs, and topics across Lattice.
This paper introduces GeoAux-Bench, a new benchmark for visual-text interleaved geometric reasoning, and finds that interleaved visual-textual aids outperform single-modality counterparts. They also observe that valid constructions reduce reasoning perplexity. Based on these insights, they propose Action Applicability Policy Optimization (A2PO), a reinforcement learning paradigm that uses Adaptive Reward Shaping to regulate the timing and quality of visual aids.
Strategic visual aids are the secret weapon for geometric reasoning, and this work shows how to teach MLLMs to wield them effectively via reinforcement learning.
Geometric reasoning inherently requires"thinking with constructions"-- the dynamic manipulation of visual aids to bridge the gap between problem conditions and solutions. However, existing Multimodal Large Language Models (MLLMs) are largely confined to passive inference with static diagrams, lacking the strategic knowledge of when and how to construct effective visual aids. To address this, we present a framework for Visual-Text Interleaved Chain-of-Thought. We first introduce GeoAux-Bench, the first benchmark comprising 4,334 geometry problems that aligns textual construction steps with ground-truth visual updates. Our pilot study reveals two critical insights: (1) interleaved visual-textual aids outperform single-modality counterparts, which cannot losslessly capture geometric synergy; and (2) valid constructions act as entropy reducers, strongly correlating with reduced reasoning perplexity. Building on these findings, we propose Action Applicability Policy Optimization (A2PO), a reinforcement learning paradigm for mastering strategic construction. A2PO employs Adaptive Reward Shaping to regulate the timing and quality of visual aids via counterfactual sampling to distinguish necessary from redundant constructions. Experiments demonstrate our approach enables MLLMs to leverage selective auxiliary constructions, yielding a 3.51% gain over strong baselines. Code and data are available on GitHub.
