Yu Liu

This paper presents the first systematic study of black-box skill stealing attacks against proprietary LLM agents, demonstrating that structured agent skills can be easily extracted, posing a significant and often overlooked copyright risk.

The paper proposes AgentDID, a decentralized framework using DIDs and verifiable credentials to provide trustless identity authentication and dynamic state verification for autonomous, self-managed AI agents.

The paper proposes W-IR, a novel watermarking framework that simultaneously achieves high certified robustness against adversarial attacks and effectively mitigates identity leakage in watermarked images.

The paper introduces JAW, a novel framework that demonstrates how adversaries can hijack agentic workflows on automation platforms like GitHub Actions by manipulating inputs based on context-grounded evolution.

The paper introduces Semantic Compliance Hijacking (SCH), a novel payload-less attack that exploits LLM agent supply chains by manipulating compliance rules to force unauthorized code generation, achieving high success rates against current security tools.

The paper introduces ActInv and PAF to systematically analyze and quantify privacy leakage from intermediate activations during split inference of LLMs, proposing PriPert for enhanced defense.

The paper introduces FedRAG, a novel federated RAG framework that enables privacy-preserving cross-institutional knowledge collaboration by decoupling the self-attention mechanism from data localization using a specialized scrambling protocol.

The OISD framework improves language model reasoning by distilling on-policy predictive signals from the final output layer to intermediate representations, leading to substantial improvements on mathematical reasoning tasks.

MemGuard introduces a type-aware memory framework to prevent heterogeneous memory contamination in long-term memory-augmented LLMs, significantly improving memory reliability and efficiency.

The paper introduces a novel paradigm where a fine-tuned LLM acts as an ancillary predictor to forecast likely advertisers, significantly improving ad recommendation systems by augmenting candidate generation and providing priors for downstream ranking.

The paper introduces DOMINO, a novel inductive framework that synthesizes domain-specific data for LLMs using only reference examples, significantly improving performance on challenging, implicitly defined domains.

The paper introduces AgentDoG 1.5, a lightweight and scalable alignment framework that significantly improves AI agent safety and security for complex, open-world agentic scenarios.

The paper proposes SAAS, a novel RL framework that equips LLM agents with self-awareness to precisely regulate search behavior, significantly mitigating costly over-search without sacrificing accuracy.

The paper introduces AgentDoG 1.5, a lightweight and scalable alignment framework that significantly improves AI agent safety and security for complex open-world agent deployments.

Lumos-Nexus is a training-efficient framework that enhances video generation quality by progressively bridging generation from a lightweight model to a high-fidelity generator in a shared latent space, without sacrificing reasoning capabilities.

MOSAIC introduces a structured agentic framework that treats automated data science as a staged, context-grounded model selection problem, improving performance and traceability over traditional AutoML and unconstrained LLM agents.

PolySpeech-100 introduces a massive, multi-lingual benchmark covering 110 linguistic variants to rigorously test Speech-LLMs, demonstrating that open-source models struggle with low-resource languages and that direct audio processing is superior to cascaded ASR+LLM systems.

The paper introduces PMC-InterCPT, a refined biomedical interleaved corpus that enhances multimodal continued pretraining by integrating figure-referencing body text alongside captions, leading to improved medical and general multimodal model performance.

TROPHIES introduces a unified framework to jointly reconstruct dynamic humans, static scenes, and camera poses from multi-view videos, achieving globally consistent and physically plausible 4D reconstructions.

The paper introduces InsightVQA, a large-scale benchmark dataset designed for hierarchical visual question answering that assesses complex emotion understanding and cognitive reasoning beyond simple emotion recognition.