SciPy Proceedings
article
Jul 13, 2026
Opening the Black Box: Mechanistic Interpretability for AI Agent Tool Selection Using Sparse Autoencoders
A mechanistic interpretability pipeline that explains why AI agents choose particular tools, using sparse autoencoder features extracted at the decision token. Feature explanations are validated with token-level fuzzing across agent scenarios in five domains.
Hannes Hapke, David Cardozo, Carlos Huisa, Triveni Gandhi
doi
none
For Editors
submission
Build Status
COMPLETED
Submission Status
PENDING
Timestamp
Oct 09, 2026 05:37:14
Submission
scipy-2026-575-lab
Storage
01a11f2a-53b3-7324-a908-723c07899593
https://prv.curvenote.com/
private CDN, long term retention policy
Checks

49 checks passed

6 optional checks failed

DOI Exists (optional)
No DOI available for ref_bricken2023 (skipped)
No DOI available for ref_juang2024 (skipped)
No DOI available for ref_gandhi2025 (skipped)
No DOI available for elhage2022toymodelssuperposition (skipped)
No DOI available for ref_sharkey2022 (skipped)
No DOI available for ref_bills2023 (skipped)