What MIT’s weight-editing experiments show about model repair

A small Qwen model’s math score rose after a targeted edit. The authors and their critics explain why that doesn’t establish new math ability.

Watch the YouTube explainer here:

A model gets stuck halfway through an answer. Another model inspects its computation and edits the weights to address the failure, without changing the prompt or running another training job.

MIT CSAIL explored that approach in Sparks of Weight-Space Programming.

The authors report that a small language model’s MATH-500 score rose from 7.9% to 48.1% after an intervention targeting about 0.016% of its weights.

The edit reduced a failure that kept the target from finishing its answers. That gives developers a useful model-repair result to examine. Describing it as “AI taught another AI mathematics” would turn that result into a misleading product claim.

How the weights were edited

A neural network’s weights are learned numbers that shape its computation; they don’t come as neatly labeled rules.