YangLing0818/buffer-of-thought-llm

Implements a meta-buffer system that stores and dynamically retrieves high-level thought-templates distilled from diverse reasoning tasks, using adaptive instantiation and a buffer-manager for scalable updates. Evaluated on 10 reasoning-intensive benchmarks (Game of 24, geometric reasoning, chess), achieving 11-51% improvements over prior approaches while requiring ~88% less computational cost than multi-path methods like tree-of-thought. Integrates with multiple LLM backends via OpenAI-compatible APIs and supports RAG-based meta-buffer management for efficient template selection.

675 stars. No commits in the last 6 months.