Kai Chen

The paper introduces Cross-Model Neuron Transfer (CNT), a post-hoc method that efficiently transfers safety-oriented functionalities between different large language models by transferring minimal subsets of neurons, achieving high performance with minimal degradation.

The paper proposes a novel, locally deployable agentic workflow using large language models (LLMs) to accurately and privately detect various types of personally identifiable information (PII) within unstructured crash narratives.

This survey establishes persistent, writable memory as an independent security problem for LLM agents, proposing a comprehensive framework for 'mnemonic sovereignty' to govern the entire memory lifecycle.

The paper proposes FreeUp, a frequency-decoupled framework that improves encrypted network anomaly detection by separately modeling and fusing low- and high-frequency components of traffic data.

The paper proposes Multi-Teacher Bayesian Knowledge Distillation (MT-BKD), a framework that uses Bayesian inference and teacher-informed priors to improve model compression, enhance predictive accuracy, and quantify uncertainty when distilling knowledge from multiple expert models.

The paper proposes Multi-Recall Memory MIA (MRMMIA), a unified attack framework to test for privacy leakage by determining if a candidate memory unit belongs to a chat agent's private memory store.

The paper proposes replacing expensive, always-on LLM calls for proactive agent triggering with a specialized Temporal-Graph-Learning (TGL) model, significantly improving efficiency and performance.

The paper proposes FedLAB, a traceable semantic codebook framework for federated multimodal graph foundation learning, which organizes multimodal graph knowledge into hierarchical codebooks and refines them through federated semantic barycenter pre-training.

The paper presents SenseNova-Vision, a unified multimodal model for computer vision tasks using natural language instructions and optional visual prompts, trained primarily on a new corpus and requiring no task-specific modifications.

The paper introduces UniClawBench, a capability-driven benchmark for evaluating proactive agents in real-world settings, using five foundational capabilities and live Docker containers.

This paper presents the benefits of visual pretraining for foundation model intelligence, outperforming text-only pretraining on multiple backbones and benchmarks.