Search papers, labs, and topics across Lattice.
The paper introduces MARS-GPS, a novel approach to geometric problem solving with LLMs that emphasizes logical inference through multiple parallel chains-of-thought (CoT) augmented with Python code execution. MARS-GPS ranks these reasoning rollouts using token-level entropy, aggregates answers via a multi-stage voting process, and incorporates self-verification. Experiments on Geometry3K demonstrate that MARS-GPS with 8 rollouts achieves 88.8% accuracy, a significant +11% improvement over the previous state-of-the-art, highlighting the benefits of multi-faceted reasoning.
LLMs can leapfrog current geometric reasoning benchmarks by generating multiple, diverse reasoning paths and voting on the best answer.
Geometric Problem Solving (GPS) remains at the heart of enhancing mathematical reasoning in large language models because it requires the combination of diagrammatic understanding, symbolic manipulation and logical inference. In existing literature, researchers have chiefly focused on synchronising the diagram descriptions with text literals and solving the problem. In this vein, they have either taken a neural, symbolic or neuro-symbolic approach. But this solves only the first two of the requirements, namely diagrammatic understanding and symbolic manipulation, while leaving logical inference underdeveloped. The logical inference is often limited to one chain-of-thought (CoT). To address this weakness in hitherto existing models, this paper proposes MARS-GPS, that generates multiple parallel reasoning rollouts augmented with Python code execution for numerical verification, ranks them using token-level entropy as a confidence signal, and aggregates answers through a multi-stage voting and self-verification pipeline. Empirical results show that MARS-GPS with 8 parallel rollouts achieves 88.8% on Geometry3K, a nearly +11% improvement over the prior state-of-the-art, with accuracy scaling consistently as the number of rollouts increases from 1 to 16 (+6.0% on ablation subset). We provide our code and data in an anonymous repository: https://anonymous.4open.science/r/MARS-GPS-DE55.
