Model Cautiousness: Towards Safer Deployment in Critical Domains
Keywords: cautiousness, calibration, out-of-distribution, safety
TL;DR: We introduce the concept of model cautiousness, which requires a model to be simultaneously calibrated in-distribution and uncertain out-of-distribution.
Abstract: In this paper, we introduce the concept of model cautiousness, which stresses the importance of aligning a model's confidence with its accuracy in in-distribution (ID) scenarios while adopting a more uncertain approach in out-of-distribution (OoD) contexts. Model cautiousness is framed as a spectrum between justified confidence and complete ignorance, induced by the inability to clearly define a model's domain of expertise. We propose a rigorous post-hoc approach to obtain a cautious model that merges the confidence scores of the primary confidence model and a model discriminating between ID and OoD inputs. A metric to measure the cautiousness error of a confidence model is introduced. We further present a simple method for discriminating ID from OoD inputs and providing a meaningful confidence estimate that an input is OoD. Finally, we benchmark our approach across 12 question-answering and 37 vision datasets, demonstrating its effectiveness in enhancing model cautiousness compared to standard calibration procedures.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2529
Loading