Zhicong Lu

Orthogonal Representation Editing achieves unprecedented precision in knowledge updates by decoupling semantic entanglement, outperforming traditional methods in both efficiency and accuracy.

C$^3$ache reveals that reusing residuals across inference chunks can dramatically speed up World Action Models, achieving a 2.5x reduction in inference time with minimal impact on performance.

LLMs' susceptibility to invisible-character prompt injections that flip paper review scores reveals a critical vulnerability in their application to academic peer review.

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.

Give users and AI shared control over conversational context, and you'll unlock more effective and satisfying collaboration.