Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention

The paper introduces Minimal Test-Time Intervention (MTI), a framework designed to improve reasoning accuracy in large language models (LLMs) with minimal computational overhead. MTI employs selective classifier-free guidance at uncertain token positions and lightweight negative-prompt guidance that utilizes the model's KV cache for efficient decoding. The approach demonstrates significant performance improvements, achieving an average gain of +9.28% on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 with Ling-mini-2.0, offering practitioners a method to enhance LLM reasoning without extensive retraining or resource costs.