Uncertainty Aware Learning for Language Model Alignment

TL;DR

This paper introduces uncertainty-aware learning (UAL) for language model alignment, adaptively adjusting training based on sample uncertainty to improve performance across diverse tasks.

Contribution

The paper proposes a novel UAL method that incorporates sample uncertainty into training, enhancing alignment and efficiency of large language models.

Findings

UAL improves token clustering in feature space.

Significant performance gains on benchmarks, e.g., 10.62% on high-entropy tasks.

Consistent outperformance over standard fine-tuning methods.

Abstract

As instruction-tuned large language models (LLMs) evolve, aligning pretrained foundation models presents increasing challenges. Existing alignment strategies, which typically leverage diverse and high-quality data sources, often overlook the intrinsic uncertainty of tasks, learning all data samples equally. This may lead to suboptimal data efficiency and model performance. In response, we propose uncertainty-aware learning (UAL) to improve the model alignment of different task scenarios, by introducing the sample uncertainty (elicited from more capable LLMs). We implement UAL in a simple fashion -- adaptively setting the label smoothing value of training according to the uncertainty of individual samples. Analysis shows that our UAL indeed facilitates better token clustering in the feature space, validating our hypothesis. Extensive experiments on widely used benchmarks demonstrate that our UAL significantly and consistently outperforms standard supervised fine-tuning. Notably, LLMs aligned in a mixed scenario have achieved an average improvement of 10.62\% on high-entropy tasks (i.e., AlpacaEval leaderboard), and 1.81\% on complex low-entropy tasks (i.e., MetaMath and GSM8K).