Surprise-Driven Test-Time Tool Adaptation: Detecting and Leveraging Unexpected Model Behaviors

While most current test-time adaptation methods (such as T-TIME [Li et al., 2023] and TeSLA [Tomar et al., 2023]) focus on smooth adaptation to drift, they generally assume that adaptation is a gradual, predictable process. However, as highlighted by Parres-Peredo et al. (2019) in cybersecurity, unexpected or anomalous behaviors can be valuable indicators—not just errors to suppress, but opportunities to learn. This research would build a system that continuously monitors for outlier actions or outputs (unexpected tool uses, highly uncertain predictions, etc.) at test time. When detected, the system could trigger enhanced information acquisition (e.g., querying for more context, seeking new data, or even switching tools), or initiate targeted model updates. The novelty lies in proactively leveraging unexpected test-time behaviors, rather than just adapting to distributions or maintaining performance. This could reveal previously unseen failure modes, inspire new tool use strategies, and create more robust, self-improving systems—much as field observations of capuchin monkeys led to new insights about tool diversity when behaviors deviated from expectations (Falótico & Ottoni, 2023).

References:

1. Greater tool use diversity is associated with increased terrestriality in wild capuchin monkeys.. Tiago Falótico, E. Ottoni (2023). American Journal of Biological Anthropology.
2. Unexpected-Behavior Detection Using TopK Rankings for Cybersecurity. A. Parres-Peredo, I. Piza-Dávila, Francisco Cervantes (2019). Applied Sciences.
3. T-TIME: Test-Time Information Maximization Ensemble for Plug-and-Play BCIs. Siyang Li, Ziwei Wang, Hanbin Luo, L. Ding, Dongrui Wu (2023). IEEE Transactions on Biomedical Engineering.
4. TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation. Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, J. Thiran (2023). Computer Vision and Pattern Recognition.