Large Language Model Group

Overview

Human-cognition-inspired data pipelines for reliable AI

The Large Language Model Group builds human-cognition-inspired data pipelines for reliable AI. The group develops data-centric methods that make foundation-model applications more effective, efficient and grounded in domain knowledge.

Its research covers how data is labeled, structured, retrieved, compressed and verified for large language models, multimodal AI, vector databases and knowledge graphs.

Research areas

• Data labeling / enhancement. Improving data quality and supervision through label-free anomaly detection, self-supervised learning and representation enhancement. [Log anomaly detection · ICDE 2026, Direct air capture dataset · Scientific Data 2025, Noise2Info · ICCV 2023]
• Data structuring / indexing. Organizing text and model-design experience into reusable structures, and retrieving relevant information to support efficient foundation-model applications. [M-DESIGN · ICML 2026, Benchmark knowledge graphs · ICLR 2025, ActQKV · arXiv 2025]
• Data grounding / reasoning. Grounding generation in retrieved evidence and domain knowledge, verifying reasoning on knowledge graphs, and maintaining relevant memory for multi-turn dialogue. [VoG · ICLR 2026, Dialogue compression · ICML 2026, BubbleRAG · Findings of ACL 2026, Learning from Emergence · KDD 2024, DesiGNN · WSDM 2026, L2E · arXiv 2025, PLM · arXiv 2025]