Kai Zhang

Factorizing world states with language unlocks surprisingly strong zero-shot reward prediction across diverse environments, outperforming end-to-end learned critics and LLM-as-a-judge approaches.

Key contribution not extracted.

Existing deep feature selection methods for recommender systems suffer from layer, baseline, and approximation biases, leading to suboptimal feature selection, which FairFS effectively mitigates.

P-GenRM personalizes LLMs more effectively by generating adaptive personas and scoring rubrics from user preferences, outperforming existing reward models by 2.31% and offering a 3% boost via test-time scaling.

Distilling refusal behavior into smaller LLMs can backfire, *increasing* their vulnerability to jailbreaks in multilingual settings.