Both Topology and Text Matter: Revisiting LLM-guided Out-of-Distribution Detection on Text-attributed Graphs
Yinlin Zhu, Di Wu, Xu Wang, Guocong Quan, Miao Hu · 2026
Read Both Topology and Text Matter: Revisiting LLM-guided Out-of-Distribution Detection on Text-attributed Graphs by Yinlin Zhu, Di Wu, Xu Wang, Guocong Quan, Miao Hu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-11641.txt.
Description
Graph networks that read text and structure often fail with confidence on strange new nodes. Existing fixes tend to use one side of the data and neglect the other. Topology methods lean on neighbors but flatten text into shallow features. Language-model methods invent fake examples that drift from reality or pollute clean data. Yinlin Zhu and four co-authors propose LG-Plug, which lines up topology and text into fine node embeddings.
From those embeddings it builds consensus-driven exposure examples through clustered, iterative prompting. Light codebooks and heuristic sampling keep query costs down. The exposure acts as a regularizer separating familiar from unfamiliar nodes and slots into detectors without a custom architecture. Across six benchmarks it improves topology-driven detectors by over 7 percent on a key error metric. It also beats prior language-model methods by over 5 percent.
Catalog record
| CREATOR | Yinlin Zhu, Di Wu, Xu Wang, Guocong Quan, Miao Hu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-11641