GPT-6 Astra edits Qwen3.5-0.8B weights directly: 0.016% of params, MATH-500 from 7.9% to 48.1%
sebkrier · x · 2026-11-10
rikiyatakehi and colleagues present "Weight-Space Programming": frontier LLMs natively reading and writing neural network weights. Without gradients or training, GPT-6 Astra edited just 0.016% of Qwen3.5-0.8B's weights across two matrices, fixing the 0.8B model's reasoning loops — MATH-500 accuracy rose from 7.9% to 48.1% (6x), and instruction following improved 3x.
hayashiyus followed with a 25-minute technical analysis aiming to turn this impressive repair into something reconstructible and measurable:
- The edit stores a steering rule in the weights whose push changes with model state
- The idealized rank-one edit follows ΔW = d kᵀ ⇒ Δh(x) = d[kᵀa(x)]: read a scalar from current activations and scale by it
- The analysis connects the stored steering rule, corrections expressible by shared weights, and what Fisher geometry predicts locally
- The end goal: test a recursive self-improvement loop where better edit-finding produces the next improvement
More from Models
- Claude Fable 5.2 debuts with 74.6% on Terminal-Bench-Science, topping rival models — BenBlaiszik · 2026-11-10
- Meta launches Muse Ember, its smallest and most efficient model for agentic execution — AIatMeta · 2026-11-10
- Claude Fable 5.2 tops FrontierCode 2.0 at 49.1, at a third of Fable 5.1's cost — nrehiew_ · 2026-11-10
- Polymarket puts 52% odds on next Claude Opus landing this month as Anthropic ships 5.2 models — Polymarket · 2026-11-10
- Artificial Analysis index: Claude Fable 5.2 tops Gemini 4 Carbon, 65 vs 60 — Angaisb_ · 2026-11-10
- Anthropic: Claude Fable 5.2 designs working electronic hardware end to end — EricBuess · 2026-11-10