AI research
The research paper "RAG over Thinking Traces Can Improve Reasoning Tasks" (arXiv:2605.06608), published on May 6, 2026, introduces a novel paradigm for enhancing Large Language Model (LLM) performance on reasoning--intensive tasks. Instead of retrieving standard text documents, the authors propose retrieving "thinking traces"—intermediate reasoning trajectories generated during previous problem--solving attempts. They introduce "T3", an offline method that transforms these traces into structured, retrieval--friendly representations. This approach demonstrates significant breakthroughs, achieving relative gains of up to 56.3% on the AIME 2025–2026 benchmark and improving performance on GPQA--Diamond, while simultaneously reducing inference costs by up to 15%.