Research map / Large language models / Reasoning and chain of thought
Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Forest-of-Thought, an inference-time framework that ensembles multiple reasoning trees (ToT or MCTSr based) with sparse activation, dynamic self-correction, input augmentation, and consensus-guided expert decision, reporting gains on Game of 24, GSM8K, MATH, and AIME 2024.
Problem
Single-pass LLM reasoning (CoT, ToT, GoT) often fails to revisit flawed paths, and tree-based search becomes costly as problems grow complex.
Approach
Combine multiple independent reasoning trees into a forest, activate only the trees that yield valid outputs at every layer (sparse activation), apply logit-score-triggered self-correction with rule checks, and pick the final answer by majority consensus with an LLM expert tie-breaker.
Key results
- On Game of 24 with 8 activated subtrees, FoT reaches 96.84% success vs 85.40% for XoT and 74.00% for ToT (b=5), at 25.64 LLM calls.
- FoT (n=4) with QwQ-32B-preview achieves 97.33% on GSM8K, 91.20% on MATH-500, and 53.33% on AIME 2024, versus 95.30/90.60/50.00 for QwQ alone.
- Ablation on Llama3-8B-Instruct Game of 24: self-correction raises accuracy from 10.58% to 60.24%, and adding input enhancement reaches 77.98%; sparse activation cuts LLM calls from 32.32 to 26.99 with no accuracy loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add self-correction (ablation) | FoT · Game of 24 | accuracy: 10.58 -> 60.24 | Helped |
| Add input enhancement to self-correction (ablation) | FoT · Game of 24 | accuracy: 60.24 -> 77.98 | Helped |
| Add sparse activation to self-correction and input enhancement (ablation) | FoT · Game of 24 | LLM invocations: 32.32 to 26.99 | Helped |
| Increase activated subtrees from 2 to 8 (ablation) | FoT with Qwen2.5-Math-7B-Instruct · GSM8K | accuracy: 93.33 -> 96.89 | Helped |
| Increase activated subtrees from 2 to 8 (ablation) | FoT with Qwen2.5-Math-7B-Instruct · MATH-500 | accuracy: 83.02 -> 86.20 | Helped |
| Increase activated subtrees from 2 to 8 (ablation) | FoT with Qwen2.5-Math-7B-Instruct · AIME 2024 | accuracy: 26.67 -> 46.67 | Helped |
| Set the self-correction threshold to 0.5 (ablation) | Dynamic self-correction · GSM8K | accuracy: 87.34, 88.17, 90.14, 88.02 | Helped |
| Use 4-rollouts MCTSr with 2 trees instead of 8-rollouts MCTSr (ablation) | FoT · GSM8K | accuracy: 3.2% higher accuracy | Helped |
| Use 32 selectable leaf nodes per ToT layer (ablation) | ToT · Game of 24 | accuracy | No clear effect |
| Use CGED instead of Majority Vote (ablation) | FoT with 5 subtrees · Mathematical reasoning benchmark (not specified) | accuracy: 83.44 -> 85.44 | Helped |
| Use CGED instead of Math Expert (ablation) | FoT with 5 subtrees · Mathematical reasoning benchmark (not specified) | accuracy: 83.70 -> 85.44 | Helped |
| Use FoT instead of MCTSr | MATH · MATH levels 1–4 | accuracy: approximately 10% | Helped |
Surprising findings
- Simply widening ToT's leaf-node count plateaus at around 32 leaves, while the forest approach keeps improving.
- On GSM8K, the 4-rollout MCTSr with 2 trees outperforms the 8-rollout MCTSr, suggesting diversity matters more than depth.
Limitations
- Paper states code will be available but gives no release link or hardware details.
- Self-correction threshold (0.5) and CGED stopping rule are tuned on the reported benchmarks with limited analysis.
- Many results are on small or single-benchmark settings, and the GSM8K/MATH comparisons mix base models and baselines inconsistently.
Experimental setup
Datasets: Game of 24 GSM8K MATH
Benchmarks: Game of 24 GSM8K MATH MATH-500 AIME 2024
Metrics: accuracy success rate LLM invocations
Baselines: IO CoT CoT-SC GoT ToT BoT XoT MCTSr GPT-4o rStar-Math QwQ
Builds on: Tree-of-Thought MCTSr Chain-of-Thought Self-Refine XoT
Closest papers by meaning
- A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.91)
- Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs (NeurIPS 2024 · similarity 0.91)
- Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Bag of Tricks for Inference-time Computation of LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.90)
- RATT: A Thought Structure for Coherent and Correct LLM Reasoning (AAAI 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.