How to distill domain tasks when frontier labs stop sharing thinking traces
thoquz · reddit · 2026-10-16
A Reddit discussion on building distillation datasets now that frontier labs no longer return thinking traces. The author wants to fine-tune a small Qwen/Gemma model for photo-to-3D-coordinate extraction but can only get final answers from frontier APIs.
Their proposed workaround: send the image+task to a frontier API for an answer-only pair, then feed the question+answer to another frontier model and ask it to reconstruct the reasoning that would lead there — a reverse-CoT style synthetic trace. They suspect mismatched-flavor traces may hurt quality and ask for better approaches. The thread covers the practical options for domain distillation in the no-CoT era.
More from Models
- Opus 5.5 still leads a month later, but open-weight decision models are closing in — airesearch12 · 2026-10-16
- Google restricts Gemini Pro access for Google Plus subscribers, matching free tier — EmergencyLess8861 · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- Meta open-sources Muse Spark: runs on two Mac Studios, license bans passing output off as human-written — mawcopolow · 2026-10-16
- Golden Gate Claude explainer resurfaces: interpretability premise still holds — HarperSCarroll · 2026-10-16
- Hugging Face livestreams release of a new open-weight model — huggingface · 2026-10-16