The paper proposes HyTRec, a recommender system that reads very long histories fast while still paying sharp attention to the latest clicks and purchases.
Decoder-only language models can be great at making user profiles (embeddings), but how we let them look at the sequence—called attention masking—changes how smart those profiles are.
Coding agents used to fix software rely on feedback; unit tests give only pass/fail signals that are often noisy or missing.