Temperature-Dependent Performance of Prompting Strategies in Extended Reasoning LLMs
https://arxiv.org/abs/2604.08563Summary
Systematically evaluates how sampling temperature interacts with prompting strategies (chain-of-thought vs zero-shot) in extended reasoning models. Tests Grok-4.1 across four temperature settings. Finds that the optimal temperature depends heavily on the prompting approach — practical guidance for anyone tuning inference parameters on reasoning models.
Categories: cs.CL, cs.AI
| Type | Link |
| Added | Apr 13, 2026 |
| Modified | Apr 13, 2026 |