Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

作者： Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar (2024)

arXiv： 2408.03314

领域

推理能力推理

TLDR（中文）

系统性给出"推理时多花 compute"的 scaling law：在固定预算下，对小模型加推理时搜索往往比训练更大模型更划算。是 o1/R1 时代理论支撑。

TLDR (English)

Systematically presents scaling laws for "spending more compute at inference time": with fixed budget, adding inference-time search to small models often more cost-effective than training larger models. Theoretical foundation for o1/R1 era.

出现在这些文章里

• Sampling 与 Decoding：从概率到文字
• Sampling and Decoding: From Probabilities to Text

同被引用

这些论文与本文出现在同一篇文章中

• Language Models are Few-Shot Learners Tom Brown et al. (2020)
• Accelerating Large Language Model Decoding with Speculative Sampling Charlie Chen et al. (2023)
• Fast Inference from Transformers via Speculative Decoding Yaniv Leviathan et al. (2023)
• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models Jason Wei et al. (2022)

相关论文

同一领域的其他论文

• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning DeepSeek-AI (2025)
  推理能力对齐
• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeek-AI (2024)
  架构推理
• OpenAI o1 System Card OpenAI (2024)
  推理能力安全
• FlashAttention-3: Fast and Accurate Attention with Asympotic IO Complexity Jay Shah et al. (2024)
  推理
• GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Joshua Ainslie et al. (2023)
  推理架构
• Accelerating Large Language Model Decoding with Speculative Sampling Charlie Chen et al. (2023)
  推理