Post-trained 35B Qwen learns weight-space programming: Freeway score jumps 21.0→47.1 without training on it
CatAstro_Piyush · x · 2026-11-07
A team post-trained Qwen 27B and 35B-A3B models to do "weight-space programming" — hand-writing neural networks one weight at a time to solve Atari-style games, a capability first demonstrated by astra (rikiyatahehi & jaydenteoh).
- Built WeightProgramBench: 322 simple games (51 Atari-style + 266 adapted from crisp-game-lib)
- Held out the 5 original minatar games (freeway, breakout, asterix, seaquest, space invaders)
- On freeway, Qwen-27B climbed from 21.0 to 47.1 points without ever training on freeway, though still below astra's 63.9
Takeaway: a 35B model can be taught to weight-space program better than Opus, showing notable generalization to unseen tasks.
More from Models
- Claude now casually generates design concept visuals while working, user delighted — RileyRalmuto · 2026-11-07
- New AI lab reportedly squeezes more from 12T tokens than DeepSeek gets from 45T — teortaxesTex · 2026-11-07
- User shares approval email for Anthropic's Cyber Verification Program Defense Access — AIFlow_ML · 2026-11-07
- As ChatGPT ends unlimited chat, Plus subscribers weigh canceling and moving to Claude — ShinyPlasm · 2026-11-07
- Qwen 4 Preview Appears on Arena, Qwen 4 Max Rumored With 1M-Token Context — WorldofAI · 2026-11-07
- RLAF: Nace AI trains decision models with a reasoning model as judge — nischay_twt · 2026-11-07