EMNLP 2026 Paper: RAG over Thinking Traces Boosts Reasoning by 43%

matei_zaharia · x · 2026-08-22

Accepted to EMNLP 2026, the paper "RAG over Thinking Traces Can Improve Reasoning Tasks" challenges the belief that RAG is ineffective for reasoning. The authors propose retrieving "thinking traces" (intermediate reasoning steps) instead of documents, introducing the T³ method to transform these traces into structured formats. Experiments on benchmarks like AIME 2025-2026 and GPQA-Diamond show consistent performance gains for top models like GPT-5, with relative improvements up to 56.3%.

Original post →

More from Models

Models channel →