Two-line weight edit fixes Qwen3.5's looping: weight-space programming reproducible
VictorSanh · x · 2026-11-12
A two-line repro for the weight-space programming fix: load Qwen/Qwen3.5-0.8B with transformers and multiply column 573 of layer 22's MLP downproj weight by 2. This edits the model's tendency to generate "Wait" and loop until thinking tokens run out (93% failure on MATH-500) — the same fix Beaker's agent found autonomously, without training or gradients.
More from Research
- Axiom Math formalization finds broken step in finite simple group classification proof — jiqizhixin · 2026-11-12
- SimEX: agents learn manipulation skills in imperfect simulation and transfer them to real robots — chris_j_paxton · 2026-11-12
- Lie bracket visualized: how second-order residue lets wheeled robots move without sliding — burny_tech · 2026-11-12
- On PostTrainBench, AI post-training now beats the human-tuned SmolLM3-3B from the same base — maksym_andr · 2026-11-12
- FigBrains: models are outpacing our ability to build training environments by hand — vsikka · 2026-11-12
- Open CHI Reviews site archives real CHI peer reviews, crowdsourced from authors — IanArawjo · 2026-11-12