LLMs can natively read and write neural net weights: editing 0.016% of Qwen3.5-0.8B boosts math 6x
yacineMTB · x · 2026-11-03
Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published 'Weight-Space Programming' on alphaXiv: the frontier LLM GPT-6 Astra can natively read and write neural network weights without gradient descent or helper models.
Key experiments
- Weight understanding: Given raw weights of 4 game policies (identical architecture/inputs/actions) with no tools, Astra correctly named all: Breakout, Pong, Pac-Man, Freeway.
- Policy repair: A PPO car policy kept getting stuck on a ramp; given weights, the scene, and one failed trace, Astra patched the weights in one reply — editing just 3.5% of weights achieved 100% success.
- Cross-model patching: Editing only 0.016% of Qwen3.5-0.8B's weights fixed its reasoning loops, boosting math up to 6x and instruction following 3x.
Claim
The authors call it Weight-Space Programming: frontier LLMs can treat raw weights as readable, editable code — diagnosing failures and patching them directly, bypassing training.
More from Models
- Fabricated citations collapse from 55% to 0%: new models nearly fix the hallucinated-reference failure — RexDouglass · 2026-11-03
- Apple hints at monetizing third-party AI apps for new revenue on earnings call — TechCrunch AI · 2026-11-03
- Writing Weights Index: Open-Source Models Struggle at Weight-Space Programming — eliebakouch · 2026-11-03
- Gemini 4 Argon ships, putting Google back in the frontier model race after a year-long lull — danshipper · 2026-11-03
- Hands-on audit punctures Astra's '0.016% weight programming' claim: it just disabled Qwen's thinking mode — teortaxesTex · 2026-11-03
- Leak: Unreleased astra6.2 builds a 3DS-style Mario Kart in 20 min under 1% quota — lxfater · 2026-11-03