Introduces SAEScientist-Bench to apply sparse autoencoders for post-hoc monitoring and auditing of models undergoing recursive self-improvement.
SAEScientist-Bench is a benchmark designed to evaluate whether AI agents can autonomously conduct mechanistic interpretability research using Sparse Autoencoders (SAEs) to monitor and audit models during recursive self-improvement (RSI). The benchmark tests agents on 20 tasks where they must design contrastive probes to navigate a 131K+ feature dictionary in Gemma-2-9B-IT and discover optimal features for specific concepts.
Frontier agents demonstrate genuine discovery capabilities, with Kimi K3 achieving the highest composite score, but they remain well behind expert baselines. While agents approach expert levels in separating target concepts from controls, they lag substantially in causal generation steering and frequently misinterpret experimental measurements. This establishes experimental model understanding as a measurable capability for closed-loop autonomous AI R&D, though significant gaps persist in verifying internal representations for safe alignment.
SAEScientist-Bench is presented as a benchmark for evaluating whether AI agents can perform end-to-end interpretability research using sparse autoencoders (SAEs). The material frames SAEs as a practical tool for post-hoc monitoring and auditing, particularly in settings where models are undergoing recursive self-improvement and traditional static evaluation may be insufficient. Rather than treating SAE training as a one-off modeling task, the work positions it as part of a broader scientific workflow: extracting interpretable features from model activations, validating those features, and using them to detect, monitor, or audit internal model behavior.
A key contribution is the decomposition of autonomous SAE interpretability research into benchmarkable agent tasks. This likely includes steps such as preparing activation data, training or tuning SAEs, selecting and interpreting features, assessing feature utility, and producing monitoring or auditing outputs. The benchmark is important because it shifts the evaluation question from “Can an agent explain a model?” to “Can an agent reliably conduct the full interpretability pipeline, including the scientific judgment required to make the results actionable?” This matters for scalable oversight: if AI systems improve their own models, external auditors need methods that can keep pace, and SAE-based monitoring is a promising candidate for providing feature-level diagnostics without requiring full white-box access.
More broadly, the material matters because it connects two active safety research threads: interpretability and agentic AI. By measuring how well AI agents can perform SAE-based interpretability work, the benchmark provides a concrete yardstick for progress in automated model auditing. It also highlights practical bottlenecks in autonomous interpretability research, such as managing training pipelines, distinguishing meaningful features from artifacts, and translating feature-level evidence into trustworthy safety conclusions. If such agents can be made more reliable, they could support faster, more systematic monitoring of complex or self-improving systems; if not, the benchmark helps identify where human oversight, better tools, or stronger evaluation criteria are still needed.