Research map / Large language models / Reasoning and chain of thought
Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models
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 Buffer of Thoughts (BoT), which stores reusable high-level thought-templates in a meta-buffer, retrieves and instantiates them per task, and updates the buffer via a buffer-manager, to improve LLM reasoning accuracy, efficiency and robustness.
Problem
LLM reasoning methods either need task-specific exemplars or manually designed prompts (single-query) or expensive recursive multi-query search (ToT, GoT), and they do not reuse high-level guidelines distilled from previously solved tasks.
Approach
A meta-buffer of task-agnostic thought-templates distilled from past problem-solving runs; each new problem is distilled, matched to a template by embedding similarity, and instantiated into a single reasoning query, with a buffer-manager adding new templates over time.
Key results
- BoT achieves 82.4 on Game of 24 vs 74.0 for ToT (second best), and 86.4 on Checkmate-in-One vs 57.2 for Meta Prompting.
- BoT uses about 12% of the cost of multi-query methods such as ToT and meta-prompting on average, per the paper's claim.
- BoT with Llama3-8B is reported to have the potential to surpass Llama3-70B on Game of 24, word sorting and Checkmate-in-One.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BoT instead of ToT | GPT-4 · Game of 24 | accuracy: 74.0 -> 82.4 | Helped |
| Use BoT instead of GoT | GPT-4 · MGSM (avg) | accuracy: 87.0 -> 89.2 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · Multi-Step Arithmetic | accuracy: 90.0 -> 99.8 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · WordSorting | accuracy: 99.6 -> 100.0 | Helped |
| Use BoT instead of PAL | GPT-4 · Python Puzzles | accuracy: 47.3 -> 52.4 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · Geometric Shapes | accuracy: 78.2 -> 93.6 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · Checkmate-in-One | accuracy: 57.2 -> 86.4 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · Date Understanding | accuracy: 79.2 -> 88.2 | Helped |
| Use BoT instead of PAL | GPT-4 · Penguins | accuracy: 93.3 -> 94.7 | Helped |
| Use BoT instead of Meta Prompting | GPT-4 · Sonnet Writing | accuracy: 79.6 -> 80.0 | Helped |
| Add the problem-distiller (ablation) | BoT with GPT-4 or Llama3-70B · Game of 24 and Checkmate-in-One | accuracy | Helped |
| Add the problem-distiller (ablation) | BoT with GPT-4 or Llama3-70B · Word list sorting and MGSM | accuracy | Helped |
Surprising findings
- Single-query Expert prompting performs below standard GPT-4 on Game of 24 (3.0 vs 3.0 tie) and PAL scores only 10.8 on Checkmate-in-One, showing method-task mismatch.
Limitations
- Gains are limited on tasks requiring human-like creativity that do not rely on a specific template.
- Quality of distilled thought-templates depends on the strength of the model used to build the meta-buffer.
- Template distillation is not optimizable (fixed prompt-based).
Experimental setup
Datasets: Game of 24 BIG-Bench Hard Geometric Shapes Multi-Step Arithmetic Two Word Sorting Checkmate-in-One Penguins Date Understanding Python Programming Puzzles MGSM Shakespearean Sonnet Writing
Benchmarks: Game of 24 MGSM Multi-Step Arithmetic Word Sorting Python Puzzles Geometric Shapes Checkmate-in-One Date Understanding Penguins Sonnet Writing
Metrics: accuracy inference time success rate
Baselines: Standard prompting Zero-shot CoT Expert Prompting PAL ToT GoT Meta Prompting
Builds on: Chain-of-Thought Tree of Thoughts Graph of Thoughts Meta Prompting Program-aided Language Models
Closest papers by meaning
- THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS (ICLR 2024 · similarity 0.90)
- Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models (ICLR 2024 · similarity 0.89)
- Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models (ICML 2024 · similarity 0.89)
- Toward Adaptive Reasoning in Large Language Models with Thought Rollback (ICML 2024 · similarity 0.89)
- StreamingThinker: Large Language Models Can Think While Reading (ICLR 2026 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.88)
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.