Tao Hu

The paper introduces CIPL, a unified channel-oriented framework, demonstrating that privacy leakage in LLM agents is governed by observable data channels and pipeline interactions, rather than being limited to individual storage components.

The paper proposes TAAC, a novel framework that enables accurate depression detection from audio while ensuring user privacy by selectively encrypting sensitive identity information.

The paper introduces XekRung, a frontier large language model for cybersecurity, which achieves state-of-the-art performance on domain-specific benchmarks through a comprehensive training and evaluation pipeline.

AESOP introduces an adversarial attack that targets the entire execution path of deep learning pipelines, demonstrating that path-aware selection can inflate computational costs by orders of magnitude more than single-model attacks.

The paper introduces FraudBench, a multimodal benchmark designed to detect AI-generated fraudulent refund evidence, finding that current AI models struggle significantly with claim-conditioned fake-damage detection.

The paper quantifies the exact parametric memory capacity of LLMs using LoRA and proposes a new optimization strategy, MemFT, to enhance memory fidelity.

The paper proposes SafeDIG, a robust safety steering framework that adapts Diffusion Transformers for text-to-image generation by treating safety control as position-aware sparse feature transfer, ensuring reliable safety across different risk domains.

The paper introduces StreamSynth, a sequential setting for synthetic data generation, and proposes SynLearner, a framework that enables LLMs to improve synthesis performance by accumulating and transferring experience across a stream of tasks.

The paper introduces SkillBrew, a multi-objective framework that treats skill bank curation as a constrained optimization problem to build efficient and well-curated skill repositories for LLM agents.

The paper introduces ConsisGuard, a framework that addresses the 'deliberation-to-enforcement gap' in LLM guardrails by ensuring that the reasoning process is faithfully and consistently translated into the final safety decision.

The paper introduces Multi-temporal Referring Segmentation (MTRS), a new task requiring models to segment language-described temporal changes, and proposes MTRefSeg-R1, a specialized framework that achieves superior performance on the newly created MTRefSeg-21K benchmark.

The paper proposes Resonant Context Anchoring (RCA), a lightweight, training-free method that enhances factual faithfulness in LLMs by dynamically amplifying the signal of external context evidence during inference.

MOSAIC is a novel scheduling framework that significantly accelerates Mixture-of-Agents (MoA) workloads by jointly optimizing expert placement and utilizing confidence-aware adaptive aggregation.

The paper proposes methods for detecting training instability in large language models using internal monitors based on the functional role of critical modules and earliest computational sites.

This paper proposes Perceive-to-Reason (P2R), a framework for fine-grained visual reasoning that decouples perception from reasoning and introduces a new reinforcement learning strategy.

A unified detection framework for AI-related content using Mahalanobis distance scores is proposed, including methods for accurate positive class characterization and joint estimation.

This paper proposes FinSAgent, an evidence-grounded multi-agent framework for financial question answering over SEC filings, which improves retrieval coverage and answer correctness through corpus-side conditioning.