arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.04735cs.LGcs.AIcs.CLcs.CYcs.IR
收藏

从数据到行为:在训练前预测未预料的模型行为

From Data to Behavior: Predicting Unintended Model Behaviors Before Training

Zhejiang University(浙江大学);National University of Singapore(新加坡国立大学)

Mengru Wang, Zhenqian Xu, Junfeng Fang, Yunzhi Yao, Shumin Deng, Huajun Chen, Ningyu Zhang

AI总结:

提出MDF方法,通过数据均值表示预测模型预训练阶段的潜在偏见和安全风险,实现高效检测。

摘要:

大型语言模型(LLMs)即使在没有显式提示或恶意内容的情况下,也可以从看似无害的训练数据中获得未预料的偏见。现有方法在微调前难以检测此类风险,使事后评估成本高昂且低效。为解决这一挑战,我们引入了Data2Behavior任务,用于在训练前预测未预料的模型行为。我们还提出了一种轻量级方法Manipulating Data Features (MDF),该方法通过数据的均值表示总结候选数据,并将其注入到基础模型的前向传递中,使数据中的潜在统计信号能够塑造模型激活,并揭示潜在的偏见和安全风险,而无需更新任何参数。MDF仅消耗约微调所需GPU资源的20%即可实现可靠的预测。在Qwen3-14B、Qwen2.5-32B-Instruct和Gemma-3-12b-it上的实验确认,MDF能够预测未预料的行为,并提供预训练漏洞的洞察。

英文摘要:

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑