The Story is Not the Science: Execution-Grounded Evaluation of Mechanistic Interpretability Research

A team of researchers has proposed a novel AI-powered solution to the growing reproducibility crisis in science, particularly within the fast-moving field of AI research itself. In their paper "The Story is Not the Science: Execution-Grounded Evaluation of Mechanistic Interpretability Research," the authors argue that traditional paper-centric review is insufficient, especially as AI agents begin autonomously generating high volumes of research. Their response is to flip the script: using AI to evaluate AI research.

They developed MechEvalAgent, an automated evaluation framework that moves beyond narrative review to perform "execution-grounded" assessment. This means the agent doesn't just read the paper; it examines the accompanying code and data to verify the experimental process, check for reproducibility of results, and test the generalizability of findings. Using mechanistic interpretability—the study of how neural networks make decisions—as a testbed, they built standardized research outputs to train and test their system.

The results are striking. MechEvalAgent achieved over 80% agreement with human expert judges on evaluating research quality. More importantly, it surfaced 51 substantial methodological problems that the human reviewers had completely missed, highlighting significant gaps in traditional peer review. This work demonstrates that AI agents can be powerful tools for enforcing scientific rigor at scale. As AI-generated research becomes more common, such automated auditing systems could become essential infrastructure for maintaining trust in scientific literature, potentially transforming journals, conferences, and funding agencies' review processes.