Introduction
Large Language Models (LLMs) have demonstrated impressive capabilities in natural language tasks, yet complex multi-step reasoning remains a significant challenge. Traditional prompting techniques like Chain-of-Thought (CoT) and Tree-of-Thought (ToT) provide structured paths for reasoning but often fail when a model makes early errors or lacks the ability to explore alternative perspectives. Forest-of-Thought (FoT) is proposed as a novel framework to overcome these limitations by scaling test-time compute through parallel reasoning trees.
Methodology: From Trees to Forests
The core innovation of FoT lies in its shift from a single reasoning path to a "forest" of independent trees. This approach mimics human problem-solving, where multiple mental frameworks are often tested before reaching a conclusion.
The FoT framework consists of five key components:
- Forest Structure: Initializes $n$ independent reasoning trees to explore diverse perspectives.
- Sparse Activation: Prunes inactive or invalid branches early to maintain computational efficiency.
- Input Data Augmentation: Uses "System 2" thinking by retrieving similar problems to prime the model.
- Dynamic Self-Correction: Monitors logit-based confidence scores to trigger refinement steps only when necessary.
- Consensus-Guided Expert Decision (CGED): Employs majority voting and an LLM-based "expert" judge to select the most rigorous solution from the forest.
Experimental Evaluation
FoT was evaluated on several high-difficulty benchmarks, including Game of 24, GSM8K, MATH, and the prestigious AIME 2024. The study utilized a range of base models such as Llama3-8B, Mistral-7B, GLM-4-9B, and the reasoning-specialized QwQ-32B-preview.
Key Insights and Scaling Laws
A pivotal finding of this research is that increasing the breadth of independent reasoning attempts (the forest) is more effective than increasing the depth of a single tree. Furthermore, the researchers identified a clear log-linear relationship between compute and performance.
As shown in the figure above, the error rate on datasets like GSM8K drops predictably as more subtrees are activated. This suggests that "thinking longer" via parallel exploration is a robust path toward achieving superhuman reasoning performance.
Results
FoT achieved state-of-the-art results across all tested benchmarks:
- Game of 24: Success rate increased from 74% (ToT) to 96.84%.
- AIME 2024: Accuracy jumped from 6.66% to 53.33% using the QwQ-32B base model, significantly outperforming previous state-of-the-art methods like rStar-Math.
- Efficiency: Sparse activation reduced LLM invocations from 32.32 to 26.99 while maintaining peak accuracy.
Conclusion
Forest-of-Thought demonstrates that scaling test-time compute through parallelized, self-correcting reasoning trees significantly enhances LLM performance on complex tasks. While computationally more intensive than standard methods, its ability to generalize across model families and follow predictable scaling laws makes it a promising framework for the future of AI reasoning.
Summary updated