Yang Liu

Recent research

• CriterionSI — Drag, Infer, Reproject: Grounding LLMs through Spatial Interaction for Image Clustering

  First author · IEEE VIS 2026, Short Papers

  Yang Liu, Xuxin Tang, Jiahao Xu, Chris North

  Abstract

  Users' clustering criterion (e.g., action, location, mood) often emerges gradually through interaction rather than being clear at the outset. CriterionSI (Criterion-guided Semantic Interaction) uses an LLM to infer and refine the user's clustering criterion from sequential drag interactions, grounds it in human feedback rather than fixed priors, and combines the inferred criterion with local drags to guide global reprojection of the image layout. Simulation-based evaluation and a usage scenario show it discovers the target criterion from a small number of interactions and progressively produces criterion-aligned clustering layouts.

  arXiv · Code

• Summary Verification Space (SVS) — Spatial Visual Analytics for Multi-Document Summary Verification

  Co-author · IEEE VIS 2026, Full Papers

  Jiahao Xu, Wei Liu, Yang Liu, Eric Krokos, Kirsten Whitley, Xuan Wang, Rebecca Faust, Chris North

  Abstract

  LLMs increasingly summarize collections of documents, but checking whether each summary statement is grounded in the sources is hard when evidence is scattered, incomplete, or conflicting. SVS is a visual analytics system for verifying multi-document summaries through spatial document organization and coordinated provenance visualization, investigating summary-guided and source-guided 2-D canvas layouts for scalable verification.

  arXiv

• CTIConnect — A Benchmark for Retrieval-Augmented LLMs over Heterogeneous Cyber Threat Intelligence

  Second author · KDD 2026, Datasets & Benchmarks Track

  Yutong Cheng, Yang Liu, Changze Li, Dawn Song, Peng Gao

  Abstract

  A benchmark for evaluating retrieval-augmented LLMs across the cyber threat intelligence task landscape: 1,860 expert-verified QA pairs over nine tasks, built on a unified environment that integrates five heterogeneous CTI sources (CVE, CWE, CAPEC, MITRE ATT&CK, and unstructured threat reports). Experiments on ten state-of-the-art LLMs show that the retrieval bottleneck shifts by task and that domain-specific retrieval strategies outperform generic retrieve-then-rerank and IRCoT.

  arXiv

The full list of papers is under Publications.