Autonomous training pipelines are maturing, but automated post-hoc experimental auditing—finding what a model represents and whether that representation can causally control behavior—remains a missing capability for safe closed-loop R&D. This work operationalizes that problem: can AI agents act as scientists using Sparse Autoencoders (SAEs) to discover interpretable features and validate causal influence?
Key Findings
- Agents can design informative contrastive probes and reliably separate target concepts from controls, achieving high activation selectivity on many tasks. This shows automated hypothesis design and representation probing are feasible at scale.
- Different frontier models exhibit complementary strengths: some excel at dictionary-wide activation ranking, others at selectivity, and a few at limited causal steering. No agent matches expert performance across all axes, highlighting capability trade-offs.
- A major gap remains in causal validation: top agents approach expert levels on selectivity (~92.9 vs 98.9) but fall far behind on causal steering (~31.5 vs 57.8). Correlated, selective features do not necessarily provide robust steering vectors.
What Sets It Apart
- Framed as a scientific discovery task: agents must generate contrastive probes, navigate a 131K+ SAE feature dictionary from Gemma-2-9B-IT, and submit single-feature candidates evaluated on activation rank, activation selectivity, and causal steering.
- Uses curated expert reference features anchored on Neuronpedia to quantify how close automated discovery comes to human expert mechanistic understanding.
- Emphasizes end-to-end experimental rigor (hypothesis → probe design → measurement → causal intervention), not just representation clustering or retrieval.
Who It's For and Trade-offs
Great fit if you study mechanistic interpretability, automated scientific workflows, or agent-driven model auditing; the benchmark gives a concrete, quantifiable task suite to measure discovery and causal validation capabilities. Look elsewhere if you only need passive representation analysis or dataset-centric benchmarks—this work stresses experimental design and causal intervention, which require access to SAE dictionaries and white-box activations and may not generalize to all model families.