Retrieving Thinking Traces Boosts AI Reasoning
A recent preprint on arXiv disputes the notion that retrieval-augmented generation (RAG) has limited utility for tasks requiring extensive reasoning, such as mathematics and code generation. The researchers contend that the issue is not with RAG itself, but rather with the selected corpus. They suggest that instead of retrieving documents, one should obtain thinking traces—intermediate reasoning paths created during problem-solving. They present T3, an offline technique that converts these traces into structured formats suitable for retrieval. Utilizing this new corpus, a straightforward retrieve-then-generate approach consistently enhances reasoning capabilities across robust models and benchmarks, including AIME 2025–2026, LiveCodeBench, and GPQA-Diamond, surpassing both non-RAG baselines and standard document retrieval.
Key facts
- RAG is widely believed to offer limited benefit for reasoning-intensive problems.
- The paper proposes retrieving thinking traces instead of documents.
- T3 transforms thinking traces into structured, retrieval-friendly representations.
- The method improves reasoning performance across AIME 2025–2026, LiveCodeBench, and GPQA-Diamond.
- The approach outperforms non-RAG baselines and retrieval over standard documents.
- The preprint is available on arXiv under ID 2605.03344.
- The method works with strong models and multiple benchmarks.
- The authors challenge the assumption that RAG is ineffective for reasoning tasks.
Entities
Institutions
- arXiv