引導式測驗 · 中等 等級

推理努力與思考預算測驗

測試推理工作和思考預算如何運作,隱藏的推理代幣如何計費,以及為什麼更多的思考會帶來收益遞減。

相關引導路徑推理努力與思考預算
問題 1 的 8

What do reasoning effort and thinking budget settings control?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. What do reasoning effort and thinking budget settings control?
    • The size of the model's training data
    • How much internal step-by-step reasoning the model does before answering
    • The number of languages the model supports
    • How many users can call the model at once
  2. How are hidden reasoning tokens generally billed?
    • They are free because you cannot see them
    • As input tokens
    • As output tokens
    • As a flat monthly fee
  3. For Anthropic extended thinking, which rule about budget_tokens is stated in the guide?
    • It must be at least 1,024 and smaller than max_tokens
    • It must equal max_tokens exactly
    • It has no minimum
    • It must be larger than the context window
  4. A model is given a large thinking budget for an easy question. What typically happens?
    • It always uses the full budget
    • 它拒絕回答
    • It may use only part of the budget
    • It switches to a different model
  5. Where do larger reasoning settings help the most?
    • Simple factual lookups
    • Problems that genuinely need several steps, like hard math or debugging
    • Greeting messages
    • Very short translations
  6. What is the recommended way to choose an effort level?
    • Always use the highest setting
    • Always use the lowest setting
    • Test real tasks at several levels and pick the lowest that meets your quality bar
    • Pick a level at random for each request
  7. Why might a reasoning model return a truncated or empty answer?
    • The output limit was too small to cover both reasoning and the answer
    • The prompt was too short
    • Reasoning models cannot answer questions
    • The thinking budget was set to its minimum
  8. When using Anthropic extended thinking with tools across multiple turns, what should you do with thinking blocks?
    • Delete them to save tokens
    • Rewrite them in your own words
    • Pass them back unchanged so the model can continue its reasoning
    • Convert them into a system prompt