Notes
Blog
Notes on concept drift, online learning, and keeping models honest after they are deployed.
-
Time-series foundation models and the problem of the label that hasn't arrived yet
Zero-shot forecasters are genuinely good now. What they are not is current — and unlike a text classifier, a forecaster cannot be corrected the moment it is wrong, because the t...
-
Language models go stale. LLM-OFA is about what to do between retrains.
A fine-tuned classifier degrades as the world moves on, and the usual answer — retrain periodically — leaves the model wrong for the whole gap. Our CIKM 2025 paper treats adapta...