LLMs natively edit neural network weights: Qwen 0.8B math jumps from 7.9% to 48.1% with no training
teakhop · reddit · 2026-11-03
A blog post on "weight-space programming" shows LLMs can natively read and write neural network weights with no fine-tuning tools.
Demos from an assistant called Astra:
- READ: identified a policy network as playing Breakout with 99% confidence by reading all 17,253 weights
- FIX: edited 3.5% of a RL car policy's weights in one shot, taking success rate from 0% to 100%
- IMPROVE: edited Qwen3.5-0.8B's weights via a Python kernel, no training, lifting MATH-500 accuracy from 7.9% to 48.1%
An early signal that models may one day treat weights like code.
More from Research
- Critic says viral AI interp demo was 'a bit of a bait': the prompt hinted the answer — mayfer · 2026-11-03
- Qwen got stuck in a loop: Astra fixed it by ablating a 'looping direction' from weights — burny_tech · 2026-11-03
- Recursive GPT-5-Mini calls often beat a single GPT-5 call for less money — lateinteraction · 2026-11-03
- Same Agent Reimplements Liquid in Rust vs C: Rust Uses 42% Fewer Tokens, C Runs 1.7x Faster — antirez · 2026-11-03
- After viral 'fly plays Beat Saber', researchers call for shared whole-brain emulation benchmark — rishabh16_ · 2026-11-03
- Researchers uncover 120 Google search twiddlers, including AI Mode's fan-out plan — VeryWellVersed · 2026-11-03