T3

Transformation of Thinking Traces

RAG is widely believed to offer limited benefit for reasoning-intensive tasks like math and code. We challenge this assumption: the limitation is the corpus, not the approach. We show that retrieving thinking traces — intermediate reasoning trajectories from strong models — consistently improves performance across frontier models and benchmarks. We further introduce T3, an offline method that transforms raw traces into structured, retrieval-friendly representations, unlocking even stronger gains.

---