The paper introduces Cognitive Demand Steering (CDS), a training-free meta-reasoning framework that evaluates what reasoning remains necessary instead of only scoring the preceding step. An LLM-based evaluator profiles residual demand across 16 cognitive dimensions, while a meta-controller selects targeted exemplars and interventions. According to the supplied abstract, CDS transfers across models and tasks without adaptation and, averaged over three frontier LLMs and six reasoning benchmarks, improves accuracy by 21.9% over direct calls and 9% over standard chain-of-thought prompting, with the largest gains on difficult mathematics and coding tasks.
No heat snapshots are available in the last 24 hours.