Lukas Kaiser explains the recipe: distilled reasoning traces plus RL
lukaszkaiser · x · 2026-09-26
Responding to Yoav Goldberg's confusion, Lukas Kaiser explains how high-quality reasoning is trained: a huge amount of synthetic data comes from taking tons of model reasoning traces, having another model clean up and summarize them (with heavy human prompting, domain knowledge and filtering), training on the result, then running RL again.
Related event: Researchers Debate Where Frontier Models' Reasoning Comes From(12 posts)→
More from Models
- DeepMind launches AlphaGenome 2, spotting nearly 2x more gene-regulating mutations — GoogleDeepMind · 2026-11-11
- Gemini 4 Carbon matches predecessor's score but explores less, gets distracted — patience_cave · 2026-11-11
- Fable 5.2 takes MazeBench frontier, matches GPT-6 Astra puzzle-solving — patience_cave · 2026-11-11
- MazeBench: Fable 5.2 scores 15%, near-frontier at half price; Gemini 4 Carbon 11% — patience_cave · 2026-11-11
- ERNIE 5.5 Flash Preview enters LMArena Battle Mode for crowdsourced ranking — arena · 2026-11-11
- Blind tests: DeepSeek matches Opus on 95% of agentic loops at 40x lower cost — bindureddy · 2026-11-11