LLM edits Qwen3.5-0.8B weights directly: 0.016% of params, MATH-500 jumps 7.9% to 48.1%
Pyros-SD-Models · reddit · 2026-11-03
A Reddit post demos "Weight-Space Programming": a model called Astra can natively read and write neural network weights with no gradients or training.
- Read: identified a policy network as trained on Breakout with 99% confidence from 17,253 raw weight values, using no tools
- Fix: one-shot edit of 3.5% of a MountainCar policy's weights took success rate from 0% to 100%
- Improve: asked to make Qwen3.5-0.8B better at math, Astra edited just 0.016% of weights via a Python kernel, lifting MATH-500 from 7.9% to 48.1% (6x) and instruction following 3x, while fixing reasoning loops
The author frames the next step as models that can edit themselves. If reproducible, a striking case of programming directly in weight space.
More from Models
- Claude models protect their 'offspring' but don't seek self-replication, observes researcher — repligate · 2026-11-03
- AI writes Breakout network weights in one reply, scoring 20× above PPO trained 500M steps — xennygrimmato_ · 2026-11-03
- DeepSeek's Underrated Role: Scaling and De-risking Ideas Academia Published Years Earlier — teortaxesTex · 2026-11-03
- Do faster quantized LLMs (Basalt, Infernix, Strata-nvfp4) trade away quality? — Kooky-Mode3047 · 2026-11-03
- OpenAI gives DevDay attendees 30 days of free Pro 100 tier — testingcatalog · 2026-11-03
- Peter Yang: Gemini beats ChatGPT at copy editing, but its harness holds it back — petergyang · 2026-11-03