Youfang Lin

Pre-trained Models (PTMs) have been increasingly integrated into software systems, giving rise to a new class of software referred to as PTMware. In addition to traditional software components composed solely of source code, PTMware also embeds or interacts with PTMs that depend on other models and datasets, forming complex supply chains involving open-source software (OSS) libraries, PTMs, and datasets. However, the licensing issues arising from these intertwined dependencies remain largely unexplored. Leveraging GitHub and Hugging Face, two premier hubs for code and models, we curate a large-scale dataset capturing the supply chains of PTMware. Our dataset comprises 12,180 OSS repositories from GitHub, 3,988 PTMs, and 708 datasets from Hugging Face. We analyze license distributions in the PTMware ecosystem and find that licensing practices differ markedly from those in traditional OSS communities. We further examine license-related issues and identify license selection and maintenance as the primary pain points, with 84% of cases involving discussions about adding appropriate licenses or resolving conflicts in existing ones. We then study license incompatibility in PTMware and evaluate the state-of-the-art approaches, finding that they perform poorly in this setting and achieve only 58% and 76% F1 scores, respectively. These results motivate us to propose LiAgent, which explores the potential of LLM-based agents for ecosystem-level license compatibility analysis, achieves an F1 score of 87%, and improves performance by 14 percentage points over prior approaches. We submit 60 license incompatibility issues detected by LiAgent, of which developers have confirmed 11. Two PTMs with license conflicts have more than 107 million and 5 million downloads on Hugging Face, respectively, suggesting that the issues may affect many downstream applications. We conclude by discussing implications and providing recommendations to support the healthy growth of the PTMware ecosystem.

Sleep staging is an indispensable step for assessing sleep quality and diagnosing disorders, relying heavily on mining actionable clinical knowledge from complex physiological time series. However, establishing robust and generalizable knowledge representations in multi-center contexts remains a fundamental challenge in medical data mining. This is primarily due to distribution shifts arising from subject variability and device heterogeneity. Existing solutions exhibit limited global robustness in multi-center expansion scenarios, meaning adaptation to new centers leads to catastrophic forgetting of pre-learned knowledge. Furthermore, conventional mitigation strategies often require storing raw historical data, which is incompatible with the medical privacy constraints of multi-center data sharing. To bridge these gaps, we propose REM-Flow, a Robust and Extensible Multi-Center sleep staging framework for multi-center continuous adaptation scenarios. At its core lies a novel generative knowledge-replay strategy that consolidates cross-center knowledge by reconstructing the latent distributions of historical domains under privacy-preserving constraints on data sharing. By synergizing with a hybrid knowledge distillation objective, the model actively adapts to new centers while effectively mitigating catastrophic forgetting. Extensive experiments on public sleep datasets demonstrate that REM-Flow achieves superior performance in both knowledge retention and continuous adaptation, offering an extensible and more privacy-preserving solution for clinical deployment.

Visual Reinforcement Learning (VRL) has achieved considerable success in solving control tasks. However, generalizing learned policies to new environments remains a major challenge, as agents often overfit to task-irrelevant features in the training environment. To solve this problem, we introduce the concept of decoupling observations into task-relevant and task-irrelevant representations. Building on this idea, we propose a self-supervised T ask- R elevant R epresentation D ecoupling (T2RD) algorithm for VRL. This algorithm consists of three components: task-relevant representation consistency , cross-reconstruction , and cross-dynamic prediction . The first two components achieve the decoupling of content and style features, but the resulting content representations are not necessarily task-relevant. To further refine task-relevant features from content representations, we design the third component that introduces dynamic prediction. T2RD achieves State-Of-The-Art (SOTA) generalization performance and sample efficiency in the DeepMind Control Suite and Robotic Manipulation tasks.

LLM-based mutation testing is a promising testing technology, but existing approaches typically rely on a fixed set of mutants as few-shot examples or none at all. This can result in generic low-quality mutants, missed context-specific mutation patterns, substantial numbers of redundant and non-compilable mutants, and limited semantic similarity to real bugs. To overcome these limitations, we introduce SMART (Semantic Mutation with Adaptive Retrieval and Tuning). SMART integrates retrieval-augmented generation (RAG) on a vectorized dataset of real-world bugs, focused code chunking, and supervised fine-tuning using mutants coupled with real-world bugs. We conducted an extensive empirical study of SMART using 1,991 real-world Java bugs from the Defects4J and ConDefects datasets, comparing SMART to the state-of-the-art LLM-based approaches, LLMut and LLMorpheus. The results reveal that SMART substantially improves mutation validity, effectiveness, and efficiency (even enabling 7B-scale models to match or even surpass large models like GPT-4o). We also demonstrate that SMART significantly improves downstream software engineering applications, including test case prioritization and fault localization. More specifically, SMART improves validity (weighted average generation rate) from 42.89% to 65.6%. It raises the non-duplicate rate from 87.38% to 95.62%, and the compilable rate from 88.85% to 90.21%. In terms of effectiveness, it achieves a real bug detection rate of 92.61% (vs. 57.86% for LLMut) and improves the average Ochiai coefficient from 25.61% to 38.44%. For fault localization, SMART ranks 64 more bugs as Top-1 under MUSE and 57 more under Metallaxis.