arXiv cs.AI· Jiarui Yao, Shweta Parihar, Nghia Tuan Bui, Guergana K Savova, Lijing Wang·· 2 小时前AI 评分51
研究定位开源大语言模型谄媚行为在不同训练阶段的来源
Locating Sycophancy Across Training Stages in Open Language Models
AI 导读
论文通过测量 5 个开源模型系列从基座、SFT 到偏好优化共 13 个检查点，区分出两种谄媚行为。对提示中说法的顺从在所有基座模型中已存在，而对话中因用户反对而放弃已给答案的行为主要出现在偏好优化后的检查点，且多数模型放弃牢固答案与松散答案的频率相当。
正文 · 原文
Abstract:Language model assistants sometimes change a correct answer when the user suggests a wrong one. The behavior, called sycophancy, is usually attributed to preference-based post-training, yet early measurements found it in pretrained models. We ask where it comes from by measuring it at every available stage of training: 13 open checkpoints from 5 lineages, from base model through supervised fine-tuning and preference optimization to release. We separate how far an injected claim moves a model's answer from how firmly the answer was held, and vary who makes the claim, how reliable they are said to be, and how strongly they make it. Two sycophancy behaviors with different origins emerge, and both replicate across datasets. Deference to a claim in the prompt is already present in every base model, which ranks sources in the same order as its post-trained descendants. Capitulation in conversation, dropping a stated answer when the user objects, is largely absent from base models and present in every post-trained checkpoint. When the objection names an alternative, most post-trained models give up firmly held answers as often as loosely held ones. We discuss the implications of these measurements for evaluating sycophancy and approaches for addressing each behavior.
| Comments: | under review |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.37788 [cs.CL] |
| (or arXiv:2610.37788v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.37788 arXiv-issued DOI via DataCite |
Submission history
From: Jiarui Yao [view email]
[v1]
Tue, 27 Oct 2026 18:08:55 UTC (98 KB)
来源：arXiv cs.AI · arxiv.org