按年被提到
20232024 · 5 次2026
大模型与语言模型›大模型推理与思维链·arXiv:2307.02477 · cs.CL

Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks

推理还是背诵?通过反事实任务探究语言模型的能力与局限
Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyürek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, Yoon Kim · 2023

用反事实任务变体检验语言模型是在推理还是在背诵

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 8 人首次 2023-07-11 · 最近 2024-11-30发表于 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT)
8不同的人
7活跃月份
7讨论串
58被引(OpenAlex)

论文提出一个评测框架:把标准任务改成偏离默认假设的“反事实”变体,来区分语言模型的抽象推理能力与对预训练中见过任务的特化记忆。在 11 个任务上,模型在反事实变体上仍有不错表现,但相比默认条件性能一致且显著下降。这表明当前语言模型既有一定的抽象解题能力,也常依赖狭窄、难以迁移的解题流程,应更谨慎地解读其表现。

HN 为什么提它

HN 上的人引用它说明大模型在反事实或分布外的变体任务上会明显退化,用来争论 LLM 到底是在推理还是在背诵,也有人质疑它没有对人类做同样的测试。

实证研究难度 3 / 5约 45 分钟有时效性
读之前最好知道
1大语言模型基础
2提示评测与基准测试
3分布外泛化(out-of-distribution generalization)
4思维链(chain-of-thought)

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
推荐sgt101 · 2023-12

“an interesting pre-print is the "reasoning or reciting" one from Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks”

推荐这篇关于 LLM 是推理还是背诵的预印本,关键证据在图 7 和 5.6 节
Will scaling work?原评论 ↗
批评famouswaffles · 2024-08

“Your paper literally admits humans would also perform worse at counterfactuals.”

批评它的结论:论文自己承认人类在反事实任务上也会更差,却没测试人类
Markov chains are funnier than LLMs原评论 ↗
引用skepticATX · 2023-07

“Current LLMs fall over when facing counterfactuals: Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks”

引用它:当前 LLM 面对反事实任务会出错
AI Safety and the Age of Dislightenment原评论 ↗
引用mjburgess · 2024-06

“they will fail under relevant counter-factual permutation of the conventions of the language/domain/topic, etc. -- see, e.g., Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks”

引用它:语言或领域的约定被反事实改变后,LLM 会失败
The illusion of state in state-space models原评论 ↗
引用cowsaymoo · 2024-07

“this paper from 4 days ago shows that Claude can fall apart quickly in out of distribution tasks”

引用这篇几天前的论文:Claude 在分布外任务上很快崩溃
Tabloid: A clickbait headline programming language (2021)原评论 ↗
引用HAL3000 · 2024-08

“Your last linked paper has flawed methodology for what it attempts to demonstrate, as shown in this paper: Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks”

指出对方最后链接的论文方法论有缺陷,并以这篇论文为据
Markov chains are funnier than LLMs原评论 ↗
看全部 7 条 →
引用dl9999 · 2024-11

“I'd like to understand why there would be studies like this: Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks if it's all just prediction.”

拿它质疑:若一切只是预测,为何会有这类研究
DeepThought-8B: A small, capable reasoning model原评论 ↗
被提交为 story · 1 次