Improving the Inference-Time Token Efficiency of Large Reasoning Models

Abstract

Reasoning models solve hard problems by generating a long chain of thought before answering. But they often overthink, spending far more tokens than needed, especially on simple tasks. This wastes computation, time, and money. This thesis studies how to reduce overthinking during inference without retraining the model, while maintaining a single reasoning trace. We build a framework that monitors the model's confidence at every token and intervenes during generation, stopping early, backtracking, or adjusting how it explores. We implement several such strategies and test them on eight benchmarks. We find that simply stopping the model early cuts tokens the most, doing so safely on easy tasks, and improving accuracy on hard ones.