引導式測驗 · 中等 等級
AI 幻覺測驗
檢查流暢的捏造、來源真實性、檢索限制、引文驗證以及對有關事實可靠性的主張的基於證據的評估。
問題 1 的 8
為什麼僅下一個標記預測無法驗證生成答案中的事實主張?
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。
- 為什麼僅下一個標記預測無法驗證生成答案中的事實主張?
- 每個模型僅連接到不可靠的網站
- 所有使用者都會提出語法錯誤的問題
- 合理的延續並不等於證明陳述正確的證據
- 流利的回答自動證明了事實
- 為什麼幻覺特別危險?
- 它們總是以不正確的語法編寫,因此使用者會忽略它們
- 它們只出現在錯誤訊息中
- 它們導致模型崩潰
- 它們的表達流暢而自信,因此很容易遺漏錯誤
- 檢索增強生成(RAG)能否完全消除幻覺?
- 是的,RAG 保證提供事實答案
- 不,模型仍然可以矛盾或忽略嘈雜或衝突的檢索文本
- 是的,但僅限週末
- 否,因為 RAG 從未提供任何文件
- 哪個才是幻覺的真實例子?
- 該模型引用了不存在的法院案例
- 模型拒絕回答問題
- 模型提出澄清問題
- 模型返回空白響應
- Which workflow can help check a generated factual answer?
- Trust every citation because it looks formal
- Treat a low sampling temperature as proof of correctness
- Hide the evidence so the answer is easier to read
- Retrieve relevant material and verify whether its statements and citations support the claims
- How should a claim that a new model has eliminated hallucinations be assessed?
- Accept it solely because the announcement sounds confident
- Assume one successful example proves every application
- Use defined tasks, evidence, and measurement conditions rather than assume universal reliability
- Judge only whether the model produces polished prose
- A generated answer cites a real page, but the page does not support the associated factual statement. What remains unverified?
- The support for that particular claim
- Whether a URL can be opened
- Whether the model used fluent sentences
- Whether the citation contains text
- A supplied document says delivery takes five business days, but its generated summary says two. Which check reveals the error?
- Counting how many sentences the summary contains
- Comparing the summary's quantity with the source document
- Checking only whether the prose sounds confident
- Assuming summaries may replace any source number