Large language models are increasingly used in settings where honesty matters, yet alignment training has not eliminated dishonest behavior. The new arXiv paper, "CAT-alyzing Honesty," highlights that models still output claims that contradict known facts or context, posing risks in real-world applications.

The authors propose continuous adversarial training (CAT) as a way to address this gap. Rather than a one-time alignment fix, CAT repeatedly exposes the model to adversarial examples that elicit dishonest responses, then trains it to avoid them. The goal is to produce honesty that transfers beyond the training distribution.

Because the abstract is truncated, the full methodology and experimental results are not yet visible. Still, the framing suggests that static alignment is insufficient and that ongoing, adaptive training may be needed to keep LLMs honest in practice.