Three-Body Alignment: Aligning Chess Agents with Human Reasoning by Reranking Reasoning Bases
Three-Body Alignment: Aligning Chess Agent with Human Reasoning through Reranked Rationale
AI summary
A study proposes "Three-Body Alignment," which aligns chess agents with human reasoning by reranking reasoning bases. The study uses an agent data engineering pipeline to construct a multi-source reasoning base dataset; t-SNE visualization shows that grandmasters, human commentators explaining NNUE output, and LLMs form distinct clusters. Experiments show that the reranking mechanism can improve alignment with humans, but there is a clear trade-off with tactical performance; the related dataset has been open-sourced on Hugging Face.