测试时计算Test-time Compute

在推理阶段多花算力换质量：多采样、多步推理、搜索与验证。Spend more compute at inference time to buy quality: more samples, more reasoning steps, search and verification.

测试时计算指模型在回答时额外消耗的算力，例如多次采样、思维链、树搜索、投票或自我验证。

它与训练时计算相对：不改权重，只改推理过程。

Test-time compute is the extra compute a model spends while answering — multiple samples, chain-of-thought, tree search, voting or self-verification.

It contrasts with training-time compute: weights do not change, only the inference process.

它解释了为什么推理模型更贵、也更慢；同时说明「多花 token」在什么任务上值得、在什么任务上只是浪费。It explains why reasoning models cost more and run slower, and when spending more tokens is worth it versus wasteful.

去 Models 图谱：看这个模型有哪些量化版、按你的显存该下哪一档Models atlas: which quantisations exist and which tier fits your VRAM ↗

回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →

• Snell 等 · Scaling LLM Test-Time Compute（arXiv 2408.03314）Snell et al., Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters ↗
• Wei 等 · Chain-of-Thought（arXiv 2201.11903）Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models ↗