Researchers have investigated the use of temperature sampling in large language models, a mechanism that controls the balance between reliability and diversity during inference. While some studies suggest varying temperature across samples can enhance performance, fixed-temperature schemes are still common. This paper analyzes the pass@K setting, examining how temperature affects success probabilities across different models and datasets. The findings indicate that while optimal multi-temperature allocations can improve performance, the gains over single-temperature baselines are often modest, especially when temperature is tuned based on the sampling budget. AI
IMPACT Clarifies when multi-temperature sampling is beneficial for LLM inference and how to optimize temperature choices.
RANK_REASON The cluster contains an academic paper discussing a technical aspect of large language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
