Xiang Wang

Gradual bridging with embodied trajectory-coupled data transforms VLMs into robust robot control policies, overcoming significant transfer challenges.

ActProbe predicts robot policy failures before they become visually apparent, enhancing both detection accuracy and operational efficiency in real-world tasks.

Surprisingly, the "think before answer" paradigm fails to enhance generative recommendation models, prompting a novel approach that redefines how reasoning is integrated into these systems.

Tool-augmented multimodal agents may appear to excel, but they often rely on learned tool-calling patterns rather than enhanced problem-solving abilities.

Noise in multi-behavior recommendation can be effectively mitigated through a novel spectral filtering approach that enhances representation purity and reliability.

Current LLM agents still struggle to infer and leverage user preferences from fragmented, real-world interactions, revealing a substantial gap between their capabilities and the demands of personalized decision-making.

Chain-of-thought prompting works not because of deep reasoning, but because adjacent tokens nudge the model towards the right answer.

LLM agents trained with simulated user and tool noise not only become more robust in messy real-world environments, but also surprisingly improve on clean, idealized benchmarks.

MLLMs can learn to reason more faithfully by explicitly anchoring visual attention to relevant image regions and reinforcing the use of that evidence during reasoning via counterfactual interventions.

LLMs choke on long numerical sequences, but a simple separator token trick can boost accuracy by 35% and cut token costs by 16%—without any training.