Researcher uses linear optimization to shorten AI model repair

Researcher Nathan Barry reports that linear optimization matched an AI agent's repair of repetitive reasoning in Qwen3.5-0.8B while changing 8× fewer weights. His GPU solver completed each optimization in about a minute, replacing a search that took the agent about 8 hours. The method selects small weight changes that push the model away from repeating text while limiting changes to productive reasoning. Tests used data excluded from fitting and tuning. The work covers one model and only loop removal; its best results still depend on the agent's choice of layers and direction of intervention.