Generative Recommendation
Anchored on Meta’s HSTU paper (“Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations”, ICML 2024), this survey traces how recommendation systems are migrating from feature-engineered DLRMs to end-to-end generative transducers — and why this shift unlocks the same power-law scaling that LLMs enjoy. 1. Why Scaling Laws Matter for RecSys — and Why DLRMs Hit a Wall 1.1 The LLM playbook does not transfer for free In NLP and vision, the recipe is well understood: bigger model + more data + more compute → predictably better loss, governed by Hoffmann/Kaplan power-laws....