Reproduce “TDGNet: Hallucination Detection in Diffusion Language Models via Temporal Dynamic Graphs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu
TDGNet introduces a temporal dynamic graph framework for detecting factual hallucinations in Diffusion Language Models (D-LLMs) by analyzing the evolving attention patterns during their iterative denoising process. The system achieved superior AUROC scores, averaging 0.68 on LLaDA-8B-Instruct and 0.73 on Dream-7B-Instruct across five QA benchmarks, while also providing granular token-level localization.
Key insights
- Hallucinations in D-LLMs are not static errors but dynamic events that evolve across the denoising trajectory, necessitating temporal modeling for effective detection.
- Evolving token-level attention patterns provide crucial, fine-grained signals reflecting bidirectional dependencies and their changes, which are indicative of factual validity or hallucination.
- Aggregating spatial neighborhood information with continuous temporal memory of token states, followed by a trajectory-level attention readout, enables robust and precise hallucination detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.