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%.
More from Models
- Developer feedback: Instinct is interesting but ultimately uninteresting — signulll · 2026-08-22
- Ling-3.0 releases checkpoints across three training stages — Designer_Mouse_6109 · 2026-08-22
- Mysterious 'Ox Alpha' model offers 1M context, huge compute claims — udmrzn · 2026-08-22
- Surprisingly hard to finetune models to avoid unnecessary hyphens — ivan_bezdomny · 2026-08-22
- Hands-on: Qwen3.8-27B Q6 Delivers Strong Agentic Coding Performance — Ok_Ninja7526 · 2026-08-22
- User test finds AI stereotypes about Jews are relatively tame — doodlestein · 2026-08-22