# 研究：LLM 高分却缺基础，Qwen3-80B 一致性仅 48%

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-12 13:19
- AIHOT 分数：46
- AIHOT 链接：https://aihot.news/items/cmtxykjfz08myrous4cf9294x
- 原文链接：https://x.com/rohanpaul_ai/status/2098642587011915981

## AI 摘要

一项研究对比 8 个 LLM 与超 18000 名人类学习者，发现人类答对难题时 72.7% 也能答对其所有前置基础题，整体得分 79.6%；Qwen3-80B-Instruct 得分更高达 92.5%，但仅 48.16% 的正确答案满足同样的前置一致性。论文指出高准确率可能掩盖知识断层，建议用难易关联题组而非孤立基准题评估推理模型。

## 正文

Humans usually need the foundations before the advanced skill; we need to know the basics before they know the harder thing

But LLMs can get the advanced answer right while missing the foundations underneath it.

The paper compares 8 LLMs with more than 18,000 human learners and asks: if you can solve a harder problem, can you also solve the easier skills underneath it?

Humans were much more consistent.

They scored 79.6% overall, and 72.7% of their correct answers also had every tested prerequisite correct.

QWEN3-80B-INSTRUCT scored higher at 92.5%, but reached that same consistency on only 48.16% of its correct answers.

In a separate test, prerequisite examples were not consistently better than same-skill or similar examples.

Even standard reasoning judges largely missed this pattern.

So high accuracy can hide disconnected pockets of knowledge.

So the paper recommends: evaluate reasoning models with connected sets of easy and hard problems, not isolated benchmark questions alone.