preprint · 2026-09-08
SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?
On this page
Why it matters here
SAEScientist-Bench evaluates agents conducting constrained experiments on model representations: they design contrastive probes and submit one sparse-autoencoder feature, scored for activation rank, concept selectivity and causal steering. Its 600 scored investigations across ten model–harness configurations show stronger concept selectivity than generation steering relative to a curated Expert reference. The reference's evaluation-text reuse and automated judging limit human-comparison claims. This measures an AI model-auditing skill rather than demonstrating a recursive improvement loop.
AI-R&D capabilityEvaluation integrity
What to keep in mind
- Newly indexed historical preprint, September 11 v2; not an October release or independent replication.
- Twenty concept–layer tasks (17 concepts) use one fixed Gemma-2-9B-IT model, two Gemma Scope layers and a single-feature probe interface; this is not broad open-ended science or a closed recursive improvement loop.
- Expert reference features were validated on evaluation texts unavailable to agents. This is a curated reference with privileged construction information, not a blind matched-budget human baseline.
- Model and execution harness vary together (Cursor versus Codex). Failed attempts are retried/replaced, with one scored result per episode and a 60-minute task timeout.
- Steering uses greedy model generation, feature-specific scale calibration and an automated GPT-4o judge in two shuffled passes; results do not constitute human validation or control of advanced systems.
- Probe counts and search volume alone are not research-quality measures. No causal conclusion about which investigation strategy improves scores is added.
- Full benchmark/protocol/score records and original-figure retention are deferred; this contextual paper addition relies on the directly reported tables and methods.