MIT CSAIL AHA! improves Qwen3.5 without LLM
Researchers introduce AHA!, a self-supervised weight-space programming method that edits transformer model weights to prevent repetitive looping without relying on external language models. By using the model's own generated branches to find a direction vector and solving a sparse linear program for layer updates, the technique significantly improves accuracy on held-out tasks like GSM8K and MATH-500. The authors report that the approach reaches parity with agent-driven weight edits while costing less than one dollar in GPU time. Open article →