Jun Yin

Grounded action-centric evidence can dramatically enhance sentiment prediction in video ads by linking emotional cues to explicit visual events.

JailbreakOPT boosts attack success rates against LLMs while slashing the number of attempts needed to breach safety measures.

Time-aware diffusion in generative recommendation can boost performance by over 29%, revealing the critical role of temporal dynamics in user preferences.

Shared codebooks and fused-teacher self-distillation offer a surprisingly effective approach to multi-view multi-label classification when data is incomplete.

Achieve better video editing without retraining by dynamically locking background features based on a "hallucination metric" that detects when the diffusion model is about to go astray.