先看這個問題為什麼值得解Why this problem is worth solving
情境先行Start with the situation
兩個語言模型在考題上相差五分：較大的模型多看了十倍資料，也採用不同的斷詞器與指令微調流程。這不是只改變模型規模的實驗，而是多項條件一起變動。 Two language models differ by five test points. The larger one saw ten times the data and also used a different tokenizer and instruction tuning. This is not a single scaling experiment; several factors changed together.
若把所有差異都歸因於參數量，研究者會錯估資料品質、後訓練與評測設計的作用。 Attributing every difference to parameter count misstates the roles of data quality, post-training, and evaluation design.
先備知識Prerequisite
- 知道預訓練（pretraining）與下游評測的差別 Know the difference between pretraining and downstream evaluation
讀完你能做到After this path
- 畫出資料、目標、架構、調適與評測的因果鏈 Map the causal chain from data, objective, and architecture to adaptation and evaluation
- 區分參數規模、訓練算力與資料組成 Distinguish parameter scale, training compute, and data composition
- 設計能檢查資料污染與提示敏感度的比較 Design comparisons that address contamination and prompt sensitivity
閱讀框架Concept frame
- 核心問題Question
- 模型規模擴大後分數提高，何時仍不足以證明是規模本身帶來新能力？ When a larger model scores higher, when is that still insufficient evidence that scaling produced a new capability?
- 對照軸Comparison axis
- 訓練與評測條件的可比程度 Comparability of training and evaluation conditions
- 判讀方式Evidence to inspect
- 並排閱讀三種相對情境，再回到論文檢查消融實驗、反事實測試與適用範圍。Read three relative scenarios side by side, then inspect the paper's ablations, counterfactual tests, and scope statements.
回答後查看解析Answer, then view the explanation
你的預測Make a prediction
要檢查『規模本身』是否造成提升，哪個比較最有診斷力？ Which comparison best diagnoses whether scale itself caused an improvement?
看現象，不急著講故事Observe before narrating
三個診斷訊號Three diagnostic signals
這些是待檢查的診斷假說。下一步用條件比較、消融實驗或反事實測試排除其他解釋。These are diagnostic hypotheses to test. Use condition comparisons, ablations, or counterfactual tests to rule out alternatives.
一次看完三種情境Read three scenarios side by side
三種情境對照Three-scenario comparison
當可比程度降低時，你會先撤回哪一種因果措辭？ As comparability falls, which causal wording would you withdraw first?
對照軸：Comparison axis: 訓練與評測條件的可比程度 Comparability of training and evaluation conditions
只能描述系統差異，不宜把結果歸因於單一 scaling 旋鈕。 Describe system-level differences; avoid attributing results to one scaling dial.
用配對子集、消融實驗與多個隨機種子，逐步排除其他解釋。 Use matched subsets, ablations, and multiple seeds to narrow explanations.
可以估計規模效應，但仍要報告不確定性與可外推的範圍。 Scale effects become estimable, while uncertainty and extrapolation limits still matter.
從直覺到機制From intuition to mechanism
把機制講清楚Explain the mechanism
基礎模型像一座由資料、目標與訓練流程共同建成的城市；參數量只是城市面積，不等於交通一定更好。 A foundation model is a city jointly built from data, objectives, and training procedures; parameter count is its area, not proof of better transport.
模型讀入大量異質資料與訓練、調適訊號，產生可供多種任務使用的表徵與生成行為；核心目標是在不同情境中保留可調適的能力。 Input: heterogeneous data plus training and adaptation signals. Output: representations and generative behaviors usable across tasks. Objective: provide adaptable capability across contexts.
- 01 資料取樣與 tokenization 決定模型能看見的訊號。 Data sampling and tokenization determine which signals the model can observe.
- 02 預訓練目標與架構把訊號編入可重用的表徵。 The pretraining objective and architecture compress signals into reusable representations.
- 03 提示、檢索或後訓練方式，會影響哪些能力能在評測中表現出來。 Prompts, retrieval, or post-training determine which capabilities are elicited during evaluation.
只看到一個訓練節點的分數跳升，就宣稱能力『突然出現』，卻沒有比較多個規模、重複實驗或替代指標。 Calling a one-checkpoint performance jump emergence without multiple scales, replications, or alternative metrics.
閱讀基礎模型研究時，分欄記錄模型規模、資料、訓練目標、調適方法與評測方式。 When reading foundation-model work, record scale, data, objective, adaptation, and evaluation in separate columns.
不用回看頁面的自我檢查Check your recall without looking back
三個回想題Three recall questions
先口頭回答，再展開解析。Answer aloud before opening each explanation.
01 為什麼參數量不是訓練算力？ Why is parameter count not training compute?
算力還取決於 token 數、架構、序列長度與最佳化流程。 Compute also depends on token count, architecture, sequence length, and optimization procedure.
02 什麼是能力引出（capability elicitation）？ What is capability elicitation?
用提示、工具或調適方法，讓既有表徵能在特定任務中被觀察。 It uses prompts, tools, or adaptation to make existing representations observable on a task.
03 宣稱能力突然出現，至少要檢查什麼？ What should an emergence claim check at minimum?
多個規模點、隨機種子造成的變異、指標曲線形狀，以及資料與後訓練是否同時改變。 Multiple scale points, seed variance, metric shape, and whether data or post-training changed simultaneously.
把工具帶回你的研究Take the tool back to your research
最後一個任務One final task
挑一篇基礎模型論文，只根據可取得的書目資料與摘要，分欄記下模型規模、資料、訓練目標、調適方法與評測；找不到的資訊明確標成未知。 Choose a foundation-model paper and, using only available metadata or abstract, fill five columns—scale, data, objective, adaptation, and evaluation—marking missing fields as unknown.