CIRCE Estimates LLM Task Complexity via Constraint Structure

CIRCE uses a lightweight LLM to construct a Constraint Intermediate Representation and a graph to predict task difficulty. On seven reasoning benchmarks, even a random GNN encoder improved correctness predictions by reducing negative log-likelihood from 0.465.