A new preprint on arXiv addresses the growing need for language model providers to communicate meaningful estimates of whether individual responses are correct. The authors propose a framework for building "reasoning-aware" reliability predictors, which take into account the model's reasoning process rather than relying solely on the final output.
The approach is described as multicalibrated, meaning the reliability estimates are calibrated not just on average but across many subgroups of inputs. This is intended to make the predictions more trustworthy for different kinds of queries and user populations.
The abstract is truncated in the source, so details on the specific methods, experiments, or results are not yet available. The paper's stated motivation, however, is clear: as people increasingly rely on LLMs for information, providers need better tools for signaling when a response may be unreliable.