Abstract
While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.
Abstract (translated)
尽管现实世界本质上是随机的,但大语言模型(LLMs)主要通过对固定基准答案的单轮推理进行评估。本研究将视角转向分布对齐:评估LLMs在重复提示下能否生成符合目标分布的输出,例如反映现实统计规律或均匀分布。我们通过职业语境中的性别、种族和情感属性来构建分布对齐任务。实证分析表明,现成LLMs及标准对齐技术(包括提示工程和直接偏好优化)均无法可靠控制输出分布。为弥合这一差距,我们提出一种新颖的微调框架,将转向令牌校准与语义对齐相结合。我们设计了混合目标函数,融合用于锚定潜在转向令牌概率质量的库尔巴克-莱布勒散度,以及用于将这些令牌与语义一致响应绑定的卡尼曼-特沃斯基优化。在六个不同数据集上的实验表明,我们的方法显著超越基线,在属性生成任务中实现了精准的分布控制。
URL
https://arxiv.org/abs/2604.05756