Xiao Wang

MirageBackdoor introduces a novel, highly stealthy backdoor attack that forces Large Language Models to generate correct reasoning steps (Think Well) but output an incorrect final answer (Answer Wrong), bypassing existing detection methods.

Nautilus Compass is a novel, black-box agent memory layer that detects persona drift in production LLM coding agents by embedding and comparing raw conversation text, achieving strong performance without requiring model weights or calling an LLM at inference time.

The paper argues that current search agents often verify existing knowledge rather than genuinely searching, and introduces LiveBrowseComp, a new benchmark to measure true evidence-driven discovery.

The paper introduces VIP-Net, a framework that leverages multi-modal spatio-temporal cues and a new dataset (Temporal-VIP) to accurately identify the most influential people in videos, overcoming the challenge of Temporal Importance Shift (TIS).