Publications

SWAT enables efficient long-context handling via Sliding Window Attention Training, replacing softmax with sigmoid and combining balanced ALiBi with Rotary Position Embedding to retain information.

Paper: arXiv:2502.18845 Code: fzkuji2026/ICDM2026_SWAT-Attention

Agentic Programming inverts the usual agent design: the program governs all control flow and the LLM is an adaptive component invoked only where reasoning is needed, so token explosion and control-flow hallucination become architecturally impossible rather than prompt-tuned away.

Paper: arXiv:2606.15874

Tandem is a collaborative framework where an LLM provides strategic reasoning insights to guide an efficient SLM, reducing computational costs by ~40% while maintaining or improving reasoning performance.

Paper: arXiv:2604.23623 Code: Applied-Machine-Learning-Lab/ACL2026_Tandem

MultiDx is a two-stage diagnostic reasoning framework that performs differential diagnosis by integrating multi-perspective evidence from web search, SOAP-formatted cases, and a clinical case database.

Paper: arXiv:2604.24186

AdapTime is an adaptive temporal reasoning method that dynamically executes reformulate, rewrite, and review actions guided by an LLM planner, enhancing temporal reasoning without external tools.

Paper: arXiv:2604.24175

MESH employs three kinds of expert modules to integrate structural and semantic information for temporal knowledge graph reasoning, capturing differences between historical and non-historical events.

Paper: 10.18653/v1/2025.findings-acl.1056

AnchorCoT predicts key entities as “anchors” to guide multi-hop reasoning and uses a ranking algorithm to ensure logical answer sequences, improving LLM performance on multi-hop QA.

Paper: 10.18653/v1/2025.findings-acl.801

Hi-Merging: a training-free method that merges specialized LLMs into a unified multi-task model using hierarchical pruning and scaling, preserving individual strengths while minimizing parameter conflicts across languages and tasks.

Paper: 10.18653/v1/2025.acl-long.1588 Code: Applied-Machine-Learning-Lab/Hi-Merging

A two-stage framework combining robust supervised fine-tuning with model merging for efficient knowledge editing in LLMs that preserves general capabilities while outperforming existing methods.

Paper: 10.18653/v1/2025.acl-industry.30 Code: Applied-Machine-Learning-Lab/MM4KE

LLM4MSR enhances multi-scenario recommendation by leveraging LLM for knowledge extraction and hierarchical meta networks, achieving improved performance and interpretability without LLM fine-tuning while maintaining deployment efficiency.

Paper: 10.1145/3627673.3679743

Uni-CTR leverages Large Language Models and pluggable domain networks to address the seesaw phenomenon and scalability challenges in multi-domain CTR prediction, achieving SOTA performance across various scenarios.

Paper: 10.1145/3698878 Code: Applied-Machine-Learning-Lab/Uni-CTR

A coarse-to-fine grounding framework that lets a general-purpose VLM locate GUI targets through active visual observation — cropping and re-examining instead of committing to a single click prediction.

Paper: arXiv:2608.03270 Code: wangshirong123/GUI-Lens

The first Chinese span-level job-skill NER dataset: 20,000+ instances from four recruitment platforms (2014–2025), annotated by an LLM-empowered Macro-Micro pipeline with expert adjudication.

Paper: arXiv:2604.23009

SRICL combines semantic retrieval, in-context learning, and supervised fine-tuning with a deterministic verifier that enforces span legality, fixing the malformed spans and boundary drift generative LLMs produce on skill extraction.

Paper: arXiv:2604.21525

A unified perspective tracing representational collapse and attention sink to improper attention allocation, introducing Lazy Attention with positional discrimination and Elastic-Softmax for focused attention.

Paper: arXiv:2601.00919