AI推理模型評測：犧牲知識換效能，小模型幻覺率82% | AI Reasoning Models Trade Knowledge for IQ, 82% Hallucinate

8月 17, 2026

By Kit 小克 | AI Tool Observer | 2026-08-17

🇹🇼 AI推理模型評測：犧牲知識換效能，小模型幻覺率82%

最近 Hacker News 上一篇分析引爆熱議：AI推理模型正在用「犧牲知識」換取推理能力，開發者 Walter van der Giessen 指出 GLM-5.2 在 AIME 2026 數學競賽解題率高達 99.2%，Qwen3.5 用更少參數也拿下 91.3%，但同一批模型在事實查核基準 SimpleQA 上表現慘烈——就連目前最強的 Gemini 2.5 Pro 也只有 53% 正確率，小模型甚至出現 80% 以上的幻覺率。這篇文章帶你看懂為什麼「越聰明的AI越會亂編」，以及日常使用該怎麼避雷。

為什麼推理能力上去了，事實卻更容易出錯？

原因出在「壓縮效率」不同。推理是可重複使用的「步驟」，一套解題邏輯能套用在無數題目上，模型只要學會流程就能舉一反三，壓縮成本很低。但事實知識剛好相反——每一條冷知識都是獨立資訊，研究估計模型平均每個參數大約只能存 2 位元(bit)的事實。換句話說，模型參數越小，能塞進去的「世界知識」就越少，訓練時又把資源優先分給推理能力，結果就是知識被自然「擠出去」。

小模型幻覺率飆82%代表什麼

GLM-5.2：400億啟用參數，AIME 2026解題率99.2%
Qwen3.5：170億啟用參數即拿下91.3%，但4B、9B版本在知識類問題上幻覺率高達80-82%
Gemini 2.5 Pro：目前SimpleQA事實準確率最高的模型，卻也只有53%

這代表現在許多主打「輕量高效」的AI模型，拿去寫程式、解數學題很強，但你若拿它來問「某個冷門歷史事件」或「小眾產品規格」，答錯甚至瞎掰的機率其實不低。

實際使用建議：把AI當推理引擎，不要當百科全書

作者提出的解法是「外掛知識庫」——與其期待模型把所有事實都背起來，不如讓它在需要時查資料（RAG、網頁搜尋、文件檢索），自己只負責邏輯推理與整合。這也是為什麼近期 Claude、GPT 等產品都在強化「工具呼叫」與「即時搜尋」功能，而不是一味把模型越練越大。對一般使用者來說，實用結論很簡單：

問數學、寫程式、抓邏輯錯誤——直接信任推理模型
問具體數字、日期、冷門事實——務必要求AI引用來源或自己開網頁搜尋交叉比對
選模型別只看跑分，SimpleQA分數低不代表模型笨，但代表你該多一道查證程序

來源：Models Are Getting Dumber on Purpose、Hacker News 討論串

好不好用，試了才知道

🇺🇸 AI Reasoning Models Trade Knowledge for IQ, 82% Hallucinate

A blog post trending on Hacker News (241 points, 139 comments) this week makes a blunt claim: frontier AI reasoning models are deliberately trading away factual knowledge to get better at reasoning. Developer Walter van der Giessen points out that GLM-5.2 hits 99.2% on the AIME 2026 math benchmark and Qwen3.5 scores 91.3% with far fewer active parameters — yet on the SimpleQA factuality benchmark, even the current leader Gemini 2.5 Pro only manages 53% accuracy, and smaller models show hallucination rates as high as 80-82% on knowledge questions.

Why Smarter Models Get Facts More Wrong

The tradeoff comes down to compression efficiency. Reasoning is a reusable procedure — once a model learns how to solve one algebra problem, the same logic generalizes to thousands of others, so it is cheap to store. Facts do not work that way. Research estimates a model can only reliably store around two bits of factual knowledge per parameter. Smaller, more efficient models simply do not have room for both frontier-level reasoning and a deep library of world knowledge — and labs are choosing reasoning.

What an 82% Hallucination Rate Actually Means

GLM-5.2: 40B active parameters, 99.2% on AIME 2026
Qwen3.5: 91.3% on AIME with just 17B active parameters, but its 4B/9B variants hallucinate 80-82% of the time on knowledge benchmarks
Gemini 2.5 Pro: the current SimpleQA leader, still tops out at only 53% factual accuracy

In practice, this means the same lightweight model that crushes your coding or math tasks may confidently make up an answer the moment you ask about an obscure historical date or a niche product spec.

The Practical Fix: Treat AI as a Reasoning Engine, Not an Encyclopedia

The proposed solution is not cramming more facts into weights — it is pairing models with external retrieval (RAG, web search, live documentation) so the model reasons over verified data instead of recalling it from memory. That is also why tool-calling and live search are becoming standard features across Claude, GPT, and other assistants. The takeaway for everyday use:

Trust reasoning models for math, code, and logic — that is where the parameters went
Never take a specific date, number, or obscure fact at face value — ask the model to cite a source or verify it with a web search yourself
A low SimpleQA score does not mean a model is "dumb" — it means you need one more verification step

Sources: Models Are Getting Dumber on Purpose, Hacker News discussion

好不好用，試了才知道

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士