Adaptive Data Collection via Autoregressive Generation
summary
A framework trains an autoregressive sequence model on massive historical interaction logs to learn a prior over content, then generates autoregressive rollouts of imagined user rewards to choose actions — an implementation of Thompson sampling with a learned prior. Theoretical results tie pretraining sequence loss to online Bayesian regret, and a proof-of-concept fine‑tuned pretrained language model on headlines produces calibrated posterior intervals and measurably improved cold‑start news recommendation on a Microsoft dataset.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 2:32De Finetti’s theorem
- 11:36Learning autoregressive probabilities
- 19:18Recommendation systems insight
- 27:26Thompson sampling with sequences
- 33:29Empirical validation
- 42:40Future work
related talks
53mGenerative AI: Overview and use cases
44mIntroduction to Sequence Analytics
46mWhy data-driven methods will shape the future (...)
25mGenerative AI vs Adaptive AI - Applications and Algorithms
50mBuilding Continuous Learning Systems
47mAdvanced Reinforcement Learning & Data Science for Social Impact
27mLarge Scale Response Prediction to Serve Ads on LinkedIn
41mKeynote: Personalization, Bandits, and Causal inference