Fork-Think Boosts LLM Reasoning Efficiency with Confidence-Based Branching

This research introduces "Fork-think with confidence," a novel approach to parallel thinking for Large Language Models (LLMs) aimed at improving reasoning task performance. Unlike existing methods that first generate multiple reasoning paths and then prune them, Fork-think adopts a "decide-first-then-think" paradigm. It begins by identifying critical "forking points" within a single initial reasoning path, using the model's confidence scores.Once a forking point is identified, the system triggers the generation of multiple continuations from that specific point. These continuations are then aggregated to form the final response. This targeted branching strategy significantly reduces computational overhead. Experiments across three different LLMs and three reasoning benchmarks demonstrate that Fork-think can cut token consumption by up to 30% and reduce runtime by as much as 57%.Crucially, these efficiency gains are achieved while maintaining or even surpassing the performance of traditional parallel thinking methods. Analysis reveals that Fork-think effectively identifies meaningful forking points, and sampling at later positions can lead to better generations. The method can also be combined with techniques like early stopping and weighted voting to further boost performance, making it a promising direction for efficient LLM reasoning without requiring warm-up or offline training.