Weight-space programming: hand-editing Qwen3.5 weights lifts a game AI from 1327 to 3604
facethef · reddit · 2026-11-09
An Opper AI blog details "weight-space programming": directly editing Qwen3.5 weights instead of retraining to build a top decision model. The chart shows eight iterations: 0.8B v1 scores 1327, improving through v3 (1911) and v5c (2050); switching to an unedited 4B gives 2175, then +circuit (2750), +correction (2911), fast serving (3290), and finally a 100-game mean of 3604 — beating jev 1.14 (3597) and ModernBERT-jev (3180), approaching greedy rule's 4450.
More from Research
- Chris McIntosh named Lead AI Scientist at UHN's Peter Munk Cardiac Care Centre — VectorInst · 2026-11-10
- Vector team wins MICCAI 2026 pediatric brain tumor segmentation challenge — VectorInst · 2026-11-10
- Benjamin Haibe-Kains wins community award for open science in AI drug discovery — VectorInst · 2026-11-10
- Nicolas Papernot wins 2026 NSERC McDonald Fellowship for trustworthy AI foundations — VectorInst · 2026-11-10
- Kieran Campbell wins CSMB New Investigator Award for ML in single-cell and cancer genomics — VectorInst · 2026-11-10
- Medical foundation model study for prostate cancer wins MICCAI IJCARS Best Paper Award — VectorInst · 2026-11-10