Test-Time Information Acquisition in Temporally Correlated Streams: Beyond the i.i.d. Assumption

A core assumption in much of the adaptation literature is that test samples are i.i.d., but as Gong et al. (2022) point out, real-world streams (e.g., sensor data, video, behavioral logs) often exhibit strong temporal dependencies. This can fundamentally undermine standard adaptation methods. This research would develop new adaptation strategies for test-time tool use and information acquisition that exploit temporal structure—perhaps by using sequence models, memory-augmented networks, or temporal anomaly detection to guide when and how new information should be acquired, or which tools should be invoked. For example, a robotic system might detect that a previously successful tool use is failing in a temporally clustered way, prompting it to seek new data or switch strategies. This work would systematically challenge the i.i.d. assumption, making test-time adaptation more realistic and robust.

References:

1. NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation. Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, Sung-Ju Lee (2022). Neural Information Processing Systems.