Language models can be sampled multiple times to reveal the distribution underlying their responses, but existing methods do not efficiently synthesize the rich epistemic signals spread across different long-form outputs. The new arXiv preprint introduces ConGrs, a graph-based response synthesis method, to address this gap.

The paper's title — "Sample, Align, Synthesize" — suggests a pipeline in which sampled responses are aligned and then synthesized using a graph structure. However, the abstract provided is brief, and the announcement does not include experimental details or specific results, so the full mechanics of ConGrs are not described in the source.

The listing is marked as a replacement (v2) of arXiv:2510.03527, meaning the authors have updated the preprint. Readers interested in the method's details will need to consult the full paper.