TREvaL uses reward models to assess the robustness of generated responses to word-level perturbations. The study examines robustness across language models and observes that supervised finetuning and RLHF can reduce it. The journal article follows the 2023 preprint and SoLaR workshop version.

Publication record.