Tag
The paper introduces CIRCE, which uses a cheap LLM to extract a problem's constraint structure into a constraint graph that predicts LLM response correctness before inference, improving NLL and AUROC over existing baselines across seven reasoning benchmarks.