The abstract introduces "no-regret generation," a new criterion for evaluating language generators. Instead of measuring average error, it requires a generator to make a sublinear number of mistakes uniformly over all demonstration sequences and compatible target languages. This is a stronger, worst-case-style guarantee.

The authors also mention an agnostic variant, but the abstract cuts off before explaining it. No further details are given about the learning algorithm, theoretical bounds, or experimental results.

Since this is the only source, there are no conflicting findings to compare. The contribution appears to be a formal definition and a research direction, rather than a completed empirical study.