arXiv 2610.21204 · cs.AI · 2026-10-15
Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally
Zhizhuo Kou, Jian Yang, Pengfei Hao, Junyu Luo, Jingshu Peng, Sirui Han, Yike Guo
The Hong Kong University of Science and Technology
在最新榜 #367 本站收录于 2026-10-25
一句话要点
研究发现推理模型想得越多，越把 some、or 读成纯逻辑义，丢掉“不是全部”“不是两者都”的日常含义，GPT-5 也一样；轨迹显示模型先想到含义再推翻，用早期读法作答可挽回。
Thinking longer makes reasoning models (GPT-5 included) read 'some' and 'or' literally, cancelling everyday implicatures; traces show models first entertain the enriched reading then cancel it, and answering from the early reading recovers it.
阅读 arXiv 原文 PDF Hugging Face 讨论 代码仓库
主题与学科
同主题：推理
- LEV: Thinking Fast and Deep with Looped Decision Models ▲337
- Thinking Inertia: LLMs Keep Thinking When Told Not To ▲241
- Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence ▲224
- Questioning the Questions: Sustaining Self-Evolution in Reasoning Models ▲85
- Improving Test-Time Scaling with Adaptive Looped Transformers ▲62
同主题：大语言模型
- Jeva: System One Models from First Principles 2026-10-13 · ▲371
- Post-Training Leaves Behavioral Shadows on Unrelated Decisions 2026-09-24 · ▲273
- WeVisOne-VL: Unified Visual Perception from Objects to Pixels 2026-10-22 · ▲247
- OmniEdu: Open Foundation Models for Learning and Teaching 2026-09-19 · ▲240
- SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes? 2026-10-19 · ▲170
数据说明
元数据来自 arXiv API（已核实）；点赞来自 Hugging Face 每日论文；引用来自 Semantic Scholar。最近刷新 2026-10-26。
要点由本站根据摘要自写，不转载摘要原文，请阅读原文。