Researchers have introduced SAEScientist-Bench, a benchmark designed to test whether AI agents can act as autonomous scientists in mechanistic interpretability — the effort to understand what a model has learned and to audit it for safe alignment.
The work situates itself in the broader push toward recursive self-improvement, arguing that automating training pipelines is not enough: reliable autonomous R&D also requires post-hoc monitoring and auditing. Sparse autoencoders (SAEs) are presented as a cornerstone tool for this, isolating interpretable features that can be inspected and used for steering.
In the benchmark, an agent is given a target concept, then designs contrastive probes and searches a Gemma Scope dictionary of more than 131,000 features in Gemma-2-9B-IT to find the optimal feature. Results are scored against curated expert reference features anchored on Neuronpedia along three dimensions: activation rank, concept selectivity on contrastive texts, and causal steering.
Across 10 agent configurations and 20 tasks, frontier agents showed genuine discovery capabilities and led on different evaluation dimensions, but stayed well behind the expert baseline. They came closest to expert level when separating target concepts from contrastive controls, while lagging substantially in causal generation steering.
Further analysis found that although agents can design contrasts that rule out spurious candidates, they frequently misinterpret their own experimental measurements. The authors frame these results as establishing experimental model understanding as a measurable capability for closed-loop autonomous AI R&D.
Code is available at github.com/Trae1ounG/SAEScientist.