TrustLDM: Benchmarking Trustworthiness in Language Diffusion Models

In three linesTrustLDM is a trustworthiness benchmark for Language Diffusion Models (LDMs) covering safety, privacy, and fairness. Results show LDMs degrade alignment when malicious post contexts are attached to masked responses, regardless of context length. An automatic evaluation framework (TrustLDM-Auto) systematically identifies vulnerable configurations across all tested models.

Summary generated by Claude — human-verified