Temperature-Dependent Performance of Prompting Strategies in Extended Reasoning Large Language Models

TL;DR

This study systematically evaluates how temperature settings affect the performance of different prompting strategies in extended reasoning large language models, revealing optimal configurations vary with strategy and temperature.

Contribution

It provides the first comprehensive analysis of temperature effects on prompting strategies in extended reasoning LLMs, highlighting the importance of joint optimization.

Findings

Zero-shot prompting peaks at moderate temperatures (T=0.4, 0.7) with 59% accuracy.

Chain-of-thought prompting performs best at temperature extremes.

Extended reasoning benefits increase from 6x to 14.3x as temperature rises.

Abstract

Extended reasoning models represent a transformative shift in Large Language Model (LLM) capabilities by enabling explicit test-time computation for complex problem solving. However, the optimal configuration of sampling temperature and prompting strategy for these systems remains largely underexplored. We systematically evaluate chain-of-thought and zero-shot prompting across four temperature settings (0.0, 0.4, 0.7, and 1.0) using Grok-4.1 with extended reasoning on 39 mathematical problems from AMO-Bench, a challenging International Mathematical Olympiad-level benchmark. We find that zero-shot prompting achieves peak performance at moderate temperatures, reaching 59% accuracy at T=0.4 and T=0.7, while chain-of-thought prompting performs best at the temperature extremes. Most notably, the benefit of extended reasoning increases from 6x at T=0.0 to 14.3x at T=1.0. These results suggest that temperature should be optimized jointly with prompting strategy, challenging the common practice of using T=0 for reasoning tasks.