AnalysisAI ModelsNovember 9, 2026

Weight-space edits turn Qwen3.5 into top Pac-Man decision model

Read original source →opper.ai

Jevlar, built by editing Qwen3.5's weights directly with no gradient descent, scored 3604 over 100 games on Opper's JevMan benchmark, up from a first score of 1327. Input and output layers stayed untouched; each edit was computed in one step from recorded activations.

1 source

More stories today

Open the live feed