Dual-objective Language Models: Training Efficiency Without Overfitting

TL;DR

This paper introduces a dual-objective training method for language models that combines autoregressive and masked-diffusion objectives, improving training efficiency and reducing overfitting without changing model architecture.

Contribution

It demonstrates that combining both objectives yields better performance and overfitting resilience across various data repetition levels without architectural modifications.

Findings

Dual-objective training outperforms single-objective models.

Optimal balance between objectives is consistent across settings.

Combining objectives enhances training efficiency and robustness.

Abstract

This paper combines autoregressive and masked-diffusion training objectives without any architectural modifications, resulting in flexible language models that outperform single-objective models. Autoregressive modeling has been a popular approach, partly because of its training efficiency; however, that comes at the cost of sensitivity to overfitting. On the other hand, masked-diffusion models are less efficient to train while being more resilient to overfitting. In this work, we demonstrate that dual-objective training achieves the best of both worlds. To derive the optimal balance between both objectives, we train and evaluate 50 language models under varying levels of data repetition. We show that it is optimal to combine both objectives under all evaluated settings and that the optimal balance is similar whether targeting autoregressive or masked-diffusion downstream performance.