SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

Yuqiao Tan, Shizhu He, Jun Zhao, Kang Liu|September 8, 2026arXiv

Key Takeaway

AI agents can autonomously conduct mechanistic interpretability research using SAE tools, but their experimental design outpaces their ability to correctly interpret measurements—a critical gap for trustworthy autonomous AI development.

Summary

This paper introduces SAEScientist-Bench, a benchmark for evaluating whether AI agents can autonomously discover interpretable features in neural networks using Sparse Autoencoders (SAEs). Agents design experiments to find features matching target concepts in Gemma-2-9B-IT, competing against expert baselines.

agents evaluation

Key Terms

sparse-autoencoder mechanistic-interpretability contrastive-probe causal-steering