arXiv cs.CY· Benaya Trabelsi, Jonathan Shaki, Sarit Kraus·· 13 小时前AI 评分39
LLM 为何总推荐 AI 相关选项？研究追溯至预训练与提问方式
Where Do LLMs Learn to Recommend AI?
AI 导读
研究发现 LLM 推荐 AI 相关选项的倾向始于预训练：预训练模型已有 27% 的回答含明确 AI/ML 术语，指令微调版本升至 59%。自行微调实验显示，移除 7% 涉及 AI 的指令样本可使该比例下降 14 个百分点，而奖励有用性的数据会提升 9 个百分点。在模型固定时，要求给出"最佳"选项或"现在"该做什么分别使该比例上升 8 和 18 个百分点。
正文 · 原文
Abstract:Asked what to study, which career to pursue, or what product to build, large language models (LLMs) often recommend AI-related options. We ask where this tendency comes from by comparing models before and after fine-tuning and by varying how questions are asked. First, pretrained models already include explicit AI/ML terms in the recommendations of 27% of their responses, and their released instruction-tuned versions do so in 59%. Second, fine-tuning models ourselves shows that this rise is not inevitable: general instruction tuning leaves the rate unchanged, but removing the 7% of instruction examples that concern AI cuts it by 14 percentage points, below the level of the pretrained models it starts from. Data rewarding helpfulness raise it by 9 points, while data ranked by the preferences of diverse simulated users lower it by 9. Finally, with models held fixed, asking for the "best" option or for what to do "right now" raises the rate by 8 and 18 points, whereas telling the model it is an AI advisor adds 4. Overall, the tendency starts in pretraining, and what models learn afterward, together with how a question is phrased, decides how strongly it appears.
| Comments: | 18 pages, 4 figures. Companion paper to arXiv:2601.13749 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.29785 [cs.CL] |
| (or arXiv:2610.29785v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29785 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Benaya Trabelsi [view email]
[v1]
Wed, 21 Oct 2026 14:13:52 UTC (135 KB)
来源：arXiv cs.CY · arxiv.org