Agent edits Qwen3.5-0.8B weights directly, lifting MATH-500 from 7.9% to 48.1%
giffmana · x · 2026-11-03
giffmana shared an example of an agent (Astra) improving Qwen3.5-0.8B's math ability by directly editing its weights with no training — only 50 forward-pass tool calls. One run made 17 edits, including 64 MLP output columns each in layers 15 and 19 (131,072 weights), taking MATH-500 from 7.9% (base) to 48.1% over 80 runs; the combined intervention gained 49.8 points (SE 1.5), mostly by fixing a repetitive-thinking failure that exhausted the token budget. The prompt demands mechanism-level evidence and bans global shortcuts like editing the final norm or embeddings.
Related event: MIT demos LLM directly editing model weights, boosting math scores sixfold(21 posts)→
More from coding & agent
- Benchmarking System 1 decision models: 40 yes/no questions in one pass, 159ms agent routing — ThatOneSnare · 2026-11-03
- Browser agents rarely fail at reasoning — modals, focus rendering and silent save failures break them — yantra-lucy02 · 2026-11-03
- Zed CEO shares weekly agent changelog automation summarizing 250 commits — zeeg · 2026-11-03
- 'Forget PRs, push direct to main': the emerging human+agent engineering workflow debate — hugobowne · 2026-11-03
- OpenAI Agents API ships automatic review, bringing Codex-style "approve for me" to agents — craigsdennis · 2026-11-03
- FastAPI Cloud goes GA after beta with thousands of apps deployed — honnibal · 2026-11-03