引導式測驗 · 中等 等級

OpenAI o1 和 o3 推理模型測驗

審查推理令牌行為、強化學習訓練聲明、延遲權衡和版本範圍的 OpenAI 模型證據。

問題 1 的 8

OpenAI 的 o 系列文件說模型在返回最終答案之前使用什麼?

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. OpenAI 的 o 系列文件說模型在返回最終答案之前使用什麼?
    • Internal reasoning tokens that are not necessarily exposed as a full user-visible transcript
    • A guaranteed step-by-step explanation shown to every user
    • A separate web search for every response
    • No extra computation beyond a standard model
  2. OpenAI所說的o1系列訓練方法的一部分是什麼?
    • Only retrieval from a web index
    • Reinforcement learning to perform complex reasoning
    • Only next-token pretraining with no post-training
    • Only human-written chain-of-thought imitation
  3. 推理時間計算縮放對於推理模型意味著什麼?
    • Each additional reasoning token guarantees a better answer
    • The model retrains its weights after each question
    • More reasoning computation may improve some tasks, with added cost and latency
    • Extra computation removes the need to evaluate model versions
  4. 應如何解釋 OpenAI 推理模型報告的基準分數?
    • As a guarantee of performance on every user task
    • As a result that applies to all future o-series versions
    • As independent of the benchmark harness
    • As tied to the named model version and evaluation protocol
  5. 為什麼即使使用者看不到內部推理令牌也會影響 API 成本?
    • They count toward output usage and context limits
    • They are removed from all billing and token accounting
    • They are stored only in the user’s browser
    • They replace all visible answer tokens
  6. o1/o3 式推理模型的典型缺點是什麼?
    • 他們無法回答任何問題
    • 硬查詢的成本較高且回應較慢
    • 他們從不犯錯
    • 他們只用英語工作
  7. What does a reasoning-effort setting influence in an API request?
    • The model’s training data after deployment
    • Whether the model can access every third-party tool
    • The reasoning budget allocated to the request
    • The benchmark score reported for earlier versions
  8. How should developers check which OpenAI reasoning model is available and its current API behavior?
    • Assume every o-series model has identical endpoints
    • Use an old benchmark article as an API specification
    • Infer availability from the model family name alone
    • Consult the current official model and reasoning documentation for that version