Reasoning Contracts for Token-Efficient Inference

Prompt-guided efficient reasoning usually asks how many tokens a language model should spend before answering. We study a complementary question: when should a model avoid generated reasoning because the input already exposes an executable structure? We propose structure-aware reasoning contracts, a training-free inference policy that routes each query to the weakest validated operator likely to be sufficient: a structured executor, a compact LM prompt contract, or a full-reasoning fallback. This turns token efficiency from a reasoning-length decision into an operator-selection decision. Across mixed symbolic, compositional, and arithmetic benchmarks, contract dispatch improves the accuracy-token tradeoff over full reasoning and prompt-only efficient reasoning baselines. A calibration study further shows that new contracts can be created for repeated domains with explicit reusable structure. The results identify both the opportunity and the boundary of the approach: explicit grammars and state transitions should often be executed rather than verbalized, whereas tasks whose difficulty is implicit semantic parsing still require stronger selectors, parsers, or fallback reasoning.