When an LLM agent misbehaves, a human often steps in to correct it. Typically, that intervention is either thrown away or replayed as training data. A new arXiv preprint suggests a different use: treat the intervention as a measurement, a label that indicates which internal computation failed.

The authors frame this as a way to study agent failures more directly. Instead of just using corrections to retrain the model, the intervention becomes a signal for localizing the faulty computation. From there, they propose steering the computation and, in some cases, editing it.

The abstract is brief, so the full methodology is not yet detailed. But the core idea is a shift in perspective: human corrections are not just feedback for the model, but diagnostic data for understanding its internal workings. The preprint is available on arXiv under identifier 2610.21708.