Yiwen Zhang

Interactive dialogue can unlock creative potential that static assessments overlook, leading to richer evaluations of creativity in AI contexts.

Self-distillation can be more effective than learning from an external teacher, but only if you optimize for preference gaps instead of blindly matching the teacher's output distribution.

LLM-derived user profiles can be powerfully leveraged for recommendation via a surprisingly simple distribution shaping approach, outperforming more complex fusion methods.

LLMs can hallucinate enough signal to make cross-domain recommendations work even when users have no overlapping history.