Can AI Agents Do Mechanistic Interpretability Research? A New Benchmark Says: Sort Of, But They Keep Misreading Their Own Experiments

Can AI Agents Do Mechanistic Interpretability Research? A New Benchmark Says: Sort Of, But They Keep Misreading Their Own Experiments | BedrockNews