AIAny
Icon for item

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

Evaluates whether AI agents can autonomously use Sparse Autoencoders (SAEs) to discover interpretable features and causally steer model behavior. Introduces a 20-task benchmark that has agents design contrastive probes to search a 131K+ feature dictionary in Gemma-2-9B-IT and evaluates activation rank, selectivity, and causal steering against expert references.

Introduction

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.

Information

  • Websitearxiv.org
  • OrganizationsThe Key Laboratory of Cognitive Intelligence, Institute of Automation, CAS, School of Artificial Intelligence, University of Chinese Academy of Sciences
  • AuthorsYuqiao Tan, Shizhu He, Jun Zhao, Kang Liu
  • Published date2026/09/08

Categories

More Items

Defines the Discovery Certification Protocol (DCP) to audit claims of discovery by AI research agents, converting claims into executable recovery and feedback tests. Specifies multi-gate certification, Core control requirements, and a deterministic offline verifier; validated in two controlled audits.

Translates natural-language instructions into executable programs that maintain an explicit, persistent global world state and compiles state-augmented 3D oriented bounding boxes into pixel-aligned conditioning signals for pretrained video generators. The approach decouples state evolution from rendering, enabling programmable entity control, off-screen state, and long-horizon interactive scenarios.

Lets vision-language models control robots via a compact semantic action interface that maps intent to discrete action units; supports zero-shot use of closed-source VLMs, low-cost fine-tuning of open VLMs, and GUI-based demonstration collection.