Workshop · Accepted ·
Optimizing Reasoning Efficiency through Prompt Difficulty Prediction
Bo Zhao, Berkcan Kapusuzoglu, Kartik Balasubramaniam, Sambit Sahu, Supriyo Chakraborty, Genta Indra Winata
NeurIPS 2025 Workshop on Efficient Reasoning · Workshop
Summary
This work trains lightweight predictors from model representations to estimate prompt difficulty or correctness and route each problem to a suitably sized reasoning model.
Research contribution
The paper studies difficulty-aware routing as a way to reduce inference cost while retaining the performance of larger models.