Metraipolis › archive › dc_71eb291d2098
SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research? — arXiv:2609.09113
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.09113
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-16T12:58:35.173Z
- SHA-256 of the archived text
e75cf580bb36e3f01d3f7cc7e5427d2dab86b3f5c0c2eeeb8bc76fd7caa5cc72- Stable URL of this page
https://metraipolis.com/documento/dc_71eb291d2098
How to cite
SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research? — arXiv:2609.09113 — arXiv (en), https://arxiv.org/abs/2609.09113 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_71eb291d2098 (Metraipolis archive dc_71eb291d2098, sha256 e75cf580bb36e3f01d3f7cc7e5427d2dab86b3f5c0c2eeeb8bc76fd7caa5cc72)
BibTeX
@misc{dc71eb291d2098,
title = {SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research? — arXiv:2609.09113},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.09113},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_71eb291d2098; sha256 e75cf580bb36e3f01d3f7cc7e5427d2dab86b3f5c0c2eeeb8bc76fd7caa5cc72},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.09113. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment. Mechanistic interpretability tools are essential to bridge this gap, among which Sparse Autoencoders (SAEs) serve as a cornerstone by isolating interpretable features for model inspection and steering. In this paper, we introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery. Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT to discover the optimal feature, evaluated against curated expert reference features anchored on Neuronpedia across activation rank, concept selectivity on contrastive texts, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capabilities and lead different evaluation dimensions, but remain well behind the expert baseline, approaching expert levels on separating target concepts from contrastive controls while lagging substantially in causal generation steering. Further analysis reveals that although agents can design contrasts to rule out spurious candidates, they frequently misinterpret experimental measurements. These results establish experimental model understanding as a measurable capability for closed-loop autonomous AI R&D. Our code is available at this https URL . Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-09 in cs.CL). Categorie arXiv: Artificial Intelligence (cs.AI) ; Computation and Language (cs.CL); Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2609.09113 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-16T12:30:25.381Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-09?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).