TDGNet: Hallucination Detection in Diffusion Language Models via Temporal Dynamic Graphs
Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu · 2026
Read TDGNet: Hallucination Detection in Diffusion Language Models via Temporal Dynamic Graphs by Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu (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-2602-08048.txt.
Description
Diffusion language models make text by side-by-side cleaning, but falsehood finders built for one-way models do not transfer well, since truth evidence appears, drifts, or self-corrects across cleaning steps. Arshia Hemmat and three co-authors offer TDGNet, which treats finding as learning over changing token-level focus graphs.
At each step, the method thins the focus graph, updates per-token notes by message passing, and pools path-wide evidence with time focus. Tests on two diffusion models across question-answering benchmarks show steady AUROC gains against output-based, latent, and static-graph baselines, at single-pass cost and modest overhead. The work shows that for diffusion models, when evidence appears matters as much as what appears.
Catalog record
| CREATOR | Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08048