Measuring AI explanation quality could help improve it, a user argues

A user backs “counterfactual simulatability” as a metric for explanation quality, arguing that having a measurable target makes improvement possible. They praise a paper’s dataset and pipeline for creating more diverse and realistic test cases than prior work, and highlight training models to produce better after-the-fact explanations of their behavior. In a follow-up, they say this paper and an earlier, more training-focused paper on the same topic used Tinker for fine-tuning experiments.