📝 SummarXiv

Abstract

Large Language Models (LLMs) have recently achieved impressive performance in math and reasoning benchmarks. However, they often struggle with logic problems and puzzles that are relatively easy for humans. To further investigate this, we introduce a new benchmark, SearchBench, which contains 11 unique search problems, each equipped with automated pipelines to generate an arbitrary number of instances and analyze the feasibility, correctness, and optimality of LLM-generated solutions. We show that using language-only reasoning, even the most advanced LLMs fail to solve SearchBench end-to-end, e.g., OpenAI's frontier models GPT4 and o1-preview solve only 1.4% and 18.6% of SearchBench problems, respectively. The reason is that SearchBench problems require considering multiple pathways to the solution and performing backtracking, posing a significant challenge to auto-regressive models. Instructing LLMs to generate code that solves the problem helps, but only slightly, e.g., GPT4's performance rises to 11.7%. Interestingly, we show that the current strongest baseline on SearchBench is obtained using in-context learning with A* algorithm implementations. We further show that this baseline can be further enhanced via a Multi-Stage-Multi-Try inference method, raising GPT4's performance above 57%.