Adaptive Test-Time Intervention for Concept Bottleneck Models
Matthew ShenDepartment of StatisticsColumbia University [email protected], Aliyah R. HsuDepartment of EECSUC Berkeley [email protected] Abhineet AgarwalDepartment of StatisticsUC Berkeley [email protected] Bin YuDepartment of Statistics, EECSCenter for Computational BiologyUC Berkeley · 2025
Read Adaptive Test-Time Intervention for Concept Bottleneck Models by Matthew ShenDepartment of StatisticsColumbia University [email protected], Aliyah R. HsuDepartment of EECSUC Berkeley [email protected] Abhineet AgarwalDepartment of StatisticsUC Berkeley [email protected] Bin YuDepartment of Statistics, EECSCenter for Computational BiologyUC Berkeley (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-06730.txt.
Description
Matthew Shen, Aliyah Hsu, Abhineet Agarwal, and Bin Yu published a 2025 paper on making concept-based AI models readable and strong. Concept bottleneck models first predict concepts, like striped or round, then predict the answer from them. The second step is usually a black box or a simple rule that costs accuracy. The authors propose fixing it at deployment time.
Their method distills the concept-to-answer mapping into an interpretable tree model. The tree stays readable, a clinician or auditor can follow its branches, while accuracy stays competitive with the black box. Adaptivity at test time is the twist: the stand-in can be adjusted where it matters, rather than freezing the pipeline at training. The paper targets a trade-off in explainable AI. It argues the choice between a readable model and a strong one is false.
Catalog record
| CREATOR | Matthew ShenDepartment of StatisticsColumbia University [email protected], Aliyah R. HsuDepartment of EECSUC Berkeley [email protected] Abhineet AgarwalDepartment of StatisticsUC Berkeley [email protected] Bin YuDepartment of Statistics, EECSCenter for Computational BiologyUC Berkeley |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-06730