Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning

Large Language Models (LLMs) have shown remarkable abilities across various language tasks, but solving complex reasoning problems remains a challenge. While existing methods like Chainof-Thought (CoT) and Tree-of-Thought (ToT) enhance reasoning by decomposing problems or structuring prompts, they typically perform a single pass of reasoning and may fail to revisit flawed paths, compromising accuracy. To address this, we propose a novel reasoning framework called Forest-of-Thought (FoT), which integrates multiple reasoning trees to leverage collective decision-making for solving complex logical problems. FoT utilizes sparse activation strategies to select the most relevant reasoning paths, improving both efficiency and accuracy. Additionally, we introduce a dynamic selfcorrection strategy that enables real-time error correction and learning from past mistakes, as well as consensus-guided decision making strategies to optimize correctness and computational resources. Experimental results demonstrate that the FoT framework, combined with these strategies, significantly enhances the reasoning capabilities of LLMs, enabling them to solve complex tasks with greater precision and efficiency.

Introduction. Large Language Models (LLMs) have revolutionized natural language processing by demonstrating remarkable abilities across a wide range of language tasks. Leveraging vast datasets and complex architectures, LLMs such as ChatGPT (Kojima et al., 2022; Achiam et al., 2023) and LLaMA (Touvron et al., 2023) can generate coherent essays, answer complex questions, and even engage in multiturn dialogues with human-like fluency. These models excel at tasks requiring not only linguistic understanding

Discussion / Conclusion. This paper introduces a novel method, the Forest of Thought (FoT), aimed at significantly enhancing the reasoning capabilities of large language models (LLMs). FoT leverages a structured framework that integrates multi-path exploration and dynamic activation of reasoning paths, addressing key limitations in existing LLM reasoning paradigms. This enables the model to achieve robust and efficient problem-solving across complex tasks while generating diverse reasoning outcomes without relying on backpropagation or fine-tuning.