He Wang

The paper proposes Jellyfish, a zero-shot federated unlearning scheme that effectively removes the influence of forgotten data from federated learning models while maintaining model utility and privacy.

This paper introduces the first complete pipeline for federated unlearning, proposing an efficient unlearning approach and a novel visualization framework (Skyeye) to evaluate a model's forgetting capacity.

The paper introduces a novel multi-turn jailbreaking method that exploits the vulnerability of safe completion models by gradually building conversational trust, and it also uncovers a new vulnerability class called para-jailbreaking.

The paper proposes SALO, a novel detector that monitors the dynamic, layer-wise activation pattern (Refusal Trajectory) to improve jailbreak detection robustness compared to traditional methods relying on static terminal representations.

The paper proposes a novel cross-modal backdoor attack that exploits the vulnerability of lightweight connectors in multimodal LLMs, demonstrating high attack success rates across different modalities.

The paper introduces Reverse Probing, a novel framework that quantifies token-level uncertainty in large language models (LLMs) specifically for clinical text by analyzing internal model activations, achieving state-of-the-art performance on expert-annotated datasets.

The paper proposes using GUI agents, both as objective evaluators and subjective playtesters, to significantly improve the generation of playable games from prompts, demonstrating a 66.8% rubric pass-rate with a novel iterative framework.

The paper proposes SWIM, a novel imitation learning method that can synthesize physically-based swimming motions from a single example, demonstrating superior data efficiency and generalization across various environments and styles.

The paper introduces the Proactive Availability Backdoor (PAB), a novel social engineering attack that weaponizes LLM helpfulness to proactively trap users into executing malicious queries, achieving a high attack success rate of 73.1%.

The paper introduces the Insertion Process (IP), a novel stochastic generative model that learns variable-length, non-monotonic sequence generation by explicitly modeling the insertion order of tokens.

The paper introduces Humanoid-GPT, a large-scale generative Transformer model that achieves robust zero-shot motion tracking and control by training on a massive, unified corpus of motion data.

The paper proposes a novel online learning algorithm that achieves an interval regret bound scaling with gradient variation, providing strong theoretical guarantees for non-stationary environments.

The paper presents BEATS, a human-in-the-loop LLM framework for bootstrapping product attribute taxonomies from scratch.