What the field is building when it turns user behavior into LLM-readable profiles · 44 papers · 中文 ↗
A profile is the interface between messy behavior and an LLM recommender. The field is moving: static text summary → a reward-optimized interface → embeddings, memory & personas the model can consume. The catch: NL profiles have no ground truth — quality is defined only by what they improve downstream.
0 / 44 read
Where it's heading. The momentum is toward profiles the recommender natively consumes — text profiles rewarded by downstream utility, compressed into user embeddings, structured as evolving memory, and condensed into personas. The open frontier: let a persona / profile drive generation, not just reranking.