用因果干预方法减少大模型偏见，仅用100样本就提升泛化能力。
Mitigating Spurious Correlations in LLMs via Causality-Aware Post-Training
- 将预测分解为事件估计与干预，从根源上削弱预训练偏见。
- 30亿参数模型仅用100个样本，在ID和OOD任务上超越传统微调与更大模型。
- 适合关注模型鲁棒性、小样本优化的开发者与研究者。
尽管大语言模型在语言建模中表现出色，但近期研究表明其在分布外（OOD）样本上表现不佳，主要源于预训练阶段习得的虚假相关性。本文提出因果感知后训练（CAPT），通过将有偏预测分解为无偏的事件估计与事件干预两个步骤，消除预训练偏见而不引入额外微调偏差，从而提升模型泛化能力。在形式因果推理基准CLadder和逻辑推理数据集PrOntoQA上的实验表明，仅使用100个分布内（ID）微调样本，经过CAPT微调的30亿参数模型在ID和OOD任务上均优于传统监督微调（SFT）及更大的语言模型，验证了CAPT的有效性与样本高效性。
原文摘要 · Abstract (English)
While large language models (LLMs) have demonstrated remarkable capabilities in language modeling, recent studies reveal that they often fail on out-of-distribution (OOD) samples due to spurious correlations acquired during pre-training. Here, we aim to mitigate such spurious correlations through causality-aware post-training (CAPT). By decomposing a biased prediction into two unbiased steps, known as \textit{event estimation} and \textit{event intervention}, we reduce LLMs' pre-training biases without incurring additional fine-tuning biases, thus enhancing the model's generalization ability. Experiments on the formal causal inference benchmark CLadder and the logical reasoning dataset PrOntoQA show that 3B-scale language models fine-tuned with CAPT can outperform both traditional SFT and larger LLMs on in-distribution (ID) and OOD tasks using only 100 ID fine-tuning samples, demonstrating the effectiveness and sample efficiency of CAPT.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。