Heng Su

The paper introduces BadStyle, a novel backdoor attack framework that generates natural, stealthy poisoned samples using LLMs to compromise various LLMs with high success rates and robust activation.

VOW introduces a novel, privacy-preserving, and cryptographically verifiable protocol for detecting watermarks in LLM-generated text, overcoming the limitations of centralized and non-verifiable existing methods.

Permit is a novel framework that enforces fine-grained, permission-aware control over the hidden states of LLMs, preventing information leakage even when sensitive data is present in the context.

The paper introduces CardioLens, a rigorous evaluation testbed for multi-sequence Cardiac MRI, which reveals that current Multimodal Large Language Models (MLLMs) exhibit a significant 'clinical reality gap' and perform poorly when simulating real-world cardiac interpretation workflows.

The paper introduces Mindgames, a comprehensive multi-game arena for evaluating LLM agents' sustained social and strategic reasoning, demonstrating that current evaluations are limited by structural scaffolding and error-survival confounds.

GSAM introduces a generalizable and safe robotic framework for articulated object manipulation, significantly improving success rates and reducing variability across diverse tasks by integrating commonsense reasoning and explicit collision constraints.