AI edits another AI's weights to lift Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 — mostly a bug fix
geoffwolfe · x · 2026-11-03
A viral post claims MIT researchers had GPT-6 Astra improve Qwen3.5-0.8B without training by editing just 0.016% of its weights, raising MATH-500 from 7.9% to 48.1%. Astra, given the weights and a 50-run test budget, traced the model's self-repetition to an internal signal written by two layers and subtracted it out.
But the reposter notes it mostly fixed a bug: Astra found and removed a repetition loop in Qwen's weights, and did almost nothing on a model without that loop. Still, an AI debugging another AI by reading its weights is wild.
More from Models
- Deep Principle releases Hydrogen-1.0, an open-weight science model post-trained from GLM-5.3 — omarsar0 · 2026-11-03
- Zyphra's ZHA brings subquadratic attention via hierarchical tree-structured indexer — yacineMTB · 2026-11-03
- GPT-6 edits Qwen weights directly, lifting MATH-500 from 7.9% to 48.1% with no gradients — KyeGomezB · 2026-11-03
- LiquidAI's LFM3-1.7B tops HF trending with 1k likes and 42.4k downloads in 5 days — maximelabonne · 2026-11-03
- 'Gemini is your friend and alignment is complicated': another eyebrow-raising model reply — vedantmisra · 2026-11-03
- Pedro Domingos: Humans are much more resistant to data than AI is — pmddomingos · 2026-11-03