Improving LLM Symbolic Problem-Solving via Automated Heuristic Discovery

Hongyi Ling, Shubham Parashar, Sambhav Khurana, Blake Olson, Anwesha Basu, Gaurangi Sinha, Zhengzhong Tu, James Caverlee, Shuiwang Ji

Keywords: LLMs, reasoning, inference time techniques

Abstract: We consider enhancing large language models (LLMs) for symbolic problem-solving tasks. While existing inference-time techniques let LLMs explore intermediate steps for problem-solving, they rely on noisy self-verification or external verifiers, which demand significant data and computations. Here, we propose Automated Heuristics Discovery (AutoHD), a novel approach that enables LLMs to generate heuristic functions to guide inference-time search through accurate evaluation of intermediate steps. These heuristic functions are further refined through an evolution process, improving their robustness and effectiveness. Our method requires no additional model training or fine-tuning, and the explicit definition of heuristic functions provides interpretability and insights into the solving process. Extensive experiments on diverse tasks demonstrate significant gains over multiple baselines, including nearly twice the accuracy on some tasks, establishing AutoHD as a reliable and interpretable solution.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 16129