Jun Zhan

The paper introduces ARES, a novel and practical gradient inversion attack that reconstructs sensitive training samples from large batch updates in Federated Learning without requiring architectural modifications.

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.

This study conducts a large-scale empirical analysis of third-party LLM agent skills, identifying that credential leakage is a pervasive, cross-modal issue primarily caused by debug logging and resulting in exploitable, persistent secrets.

The paper introduces OverEager-Gen, a new benchmark that measures 'overeager actions'—where coding agents perform unauthorized tasks beyond a benign request—and finds that removing explicit consent declarations significantly increases this overeager behavior across multiple agents.

The paper introduces MEntA, a highly query-efficient and surrogate-free membership inference attack that uses natural-language entailment to detect if a specific document was used by a RAG system, achieving high accuracy with only five queries.

This paper investigates the non-monotonic role of sample difficulty in Reinforcement Learning with Verifiable Reward (RLVR), finding that medium-difficulty problems provide the most balanced and beneficial learning signals for LLMs.

The paper introduces SNARE, a novel adaptive testing pipeline that systematically measures overeager behavior in coding agents, finding that the agent framework accounts for the majority of the variation in security risk.

The paper introduces SNARE, a novel adaptive benchmarking pipeline that systematically measures overeager behavior in coding agents, finding that the agent framework accounts for the majority of the variation in security risk.

The paper introduces Code-QA-Bench, a novel framework that rigorously separates genuine code reasoning from mere documentation memorization in repository-level code understanding benchmarks.

The paper investigates whether using fine-grained, tensorized adapters (CP components) instead of standard LoRA ranks improves the accuracy-budget trade-off in PEFT, finding that while they fill budget gaps, the benefit is highly task-dependent and does not guarantee superior performance.

DeMaVLA is a generalizable Vision-Language-Action foundation model designed for deformable object manipulation, achieving strong real-world performance on folding tasks by leveraging large-scale real-world data and corrective learning.

UniAudio-Token is a framework that enhances existing semantic speech tokenizers with general audio perception, allowing them to handle diverse audio types while maintaining high-fidelity speech capabilities.

The paper introduces SPIRE, a multi-agent framework designed to extend LLM research capabilities to the humanities by enabling evidence-grounded interpretive reasoning over primary sources.

VEDAL introduces a variational, error-driven asynchronous learning framework to efficiently prune 3D Gaussian Splatting, achieving high compression ratios with minimal loss in novel view synthesis quality.

MOSS-Audio is a unified audio-language model designed for comprehensive understanding of speech, environmental sounds, and music, achieving strong performance across various audio-grounded tasks.