TraceForge synthesizes reasoning traces from answers, beating Claude's own CoT by +3.7/+7.1
IndependentResearcher121 · hf · 2026-10-22
An independent researcher presents TraceForge, a two-stage supervised-plus-adversarial method that synthesizes reasoning traces matching a student model's style from teacher answers only — zero tokens from Claude or GPT in the entire pipeline.
- On HumanEval+/MBPP+, the synthetic CoT beats Claude's own traces by +3.7/+7.1 and Qwen3.8-Max's own traces and answers by +21/+35
- Pitch: "Bring an answer. Get back a chain of thought — and it teaches better than the real trace."
- Fully legal distillation path with no closed-model tokens
More from coding & agent
- How to keep a 156-second AI music video consistent: 7 parallel subagents — lmoroney · 2026-10-22
- How to keep a 156-second AI music video consistent: 7 parallel subagents — lmoroney · 2026-10-22
- Agentic coding will explode CI bills — Celesto claims it cut one from $1,000 to $50 — aniketmaurya · 2026-10-22
- OpenClaw opens its team environment to community contributors via Focus Weeks — RichardSocher · 2026-10-22
- ESA's cat keyboard incident is a warning for leaving AI agents unattended — vincent_koc · 2026-10-22
- Defense must run at machine speed: agents patch, engineers judge — saranormous · 2026-10-22