Berkeley's T3 corpus shows RAG over thinking traces boosts AIME reasoning by up to 43.9% at EMNLP 2026
berkeley_ai · x · 2026-10-27
At EMNLP 2026 in Budapest, a UC Berkeley team (Negar Empr et al.) presents "RAG over Thinking Traces Can Improve Reasoning Tasks." Counterintuitively, RAG can improve reasoning—if you use the right corpus. They introduce T3 (Transformed Thinking Traces), a corpus built from model thinking traces for RAG over reasoning tasks, yielding gains of up to 43.9% on AIME 2025–2026. Paper and dataset links are in the thread; poster at Poster Session A, Hall H.
More from Research
- Anthropic folder with possible second ETH refutation proof spotted in shared TCS repository — NoahSD · 2026-10-27
- Why an AI-disputed ETH matters: subexponential n/√log₂n scaling vs 2ⁿ, huge savings at large n — robleclerc · 2026-10-27
- Why OpenAI's claimed 3-SAT breakthrough would be a big deal, explained — micoolcho · 2026-10-27
- Google's StreamVibe: proactive streaming video agents that track evolving user intent — google · 2026-10-27
- Qwen-Image-PE: two-stage training lifts text-to-image prompts by 7.81 points — Qwen · 2026-10-27
- OPTIC compresses per-layer embeddings 1000x with minimal performance loss — Jiacheng Liu · 2026-10-27