“I found their position paper on it super convincing Alice in Wonderland”
Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models
一道简单常识数学题,让 GPT-4 等顶尖模型推理崩溃
论文用一个人类很容易解决、用简洁自然语言表述的常识数学题(AIW 问题),展示包括 GPT-4 和 Claude 3 Opus 在内的所有最先进模型的泛化与基本推理出现剧烈崩溃:平均表现低,且在不改变结构和难度的模板微调下波动很大。模型还对错误答案高度自信,给出听起来合理的“解释式”编造,思维链和多步重新评估等常规干预也无效。作者呼吁重新评估现有基准对能力的声称,并建立能发现这类缺陷的标准化基准。
HN 上的人拿它当作“最先进 LLM 在简单题上也会推理崩溃、幻觉很容易诱发”的证据,用于反驳对基准分数和 LLM 推理能力的乐观解读。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Hallucinations still are trivial [1] to elicit.”
“research paper shows the reasoning breakdown in SOTA LLMs by asking a simple question”
“when it is probed further [1] it starts to break down in exactly the same way a crammer does”
“Edit: found two more papers”
“These results are very similar to the "Alice in Wonderland" problem [1, 2], which was already discussed a few months ago.”
看全部 7 条 →
“Did you link the wrong Arxiv paper? Alice in Wonderland does not look like a position paper nor does it share any authors with this Genie 2 work.”
- Simple tasks showing reasoning breakdown in state-of-the-art LLMs ↗tosh · 2024-06-05 · 375 分