A Continuum of Generation Tasks for Investigating Length Bias and Degenerate Repetition

TL;DR

This paper introduces a new experimental framework to study how task constrainedness affects language model behaviors like length bias and repetition, revealing that these issues are linked to the mode of the distribution rather than overall entropy.

Contribution

It provides a novel method to vary task constrainedness continuously, clarifying the relationship between task constraints and model degenerate behaviors.

Findings

Repetition decreases smoothly with increased constrainedness.

Length bias also decreases with constrainedness, but not due to constrainedness.

Problems mainly affect the mode, not the entire distribution.

Abstract

Language models suffer from various degenerate behaviors. These differ between tasks: machine translation (MT) exhibits length bias, while tasks like story generation exhibit excessive repetition. Recent work has attributed the difference to task constrainedness, but evidence for this claim has always involved many confounding variables. To study this question directly, we introduce a new experimental framework that allows us to smoothly vary task constrainedness, from MT at one end to fully open-ended generation at the other, while keeping all other aspects fixed. We find that: (1) repetition decreases smoothly with constrainedness, explaining the difference in repetition across tasks; (2) length bias surprisingly also decreases with constrainedness, suggesting some other cause for the difference in length bias; (3) across the board, these problems affect the mode, not the whole distribution; (4) the differences cannot be attributed to a change in the entropy of the distribution, since another method of changing the entropy, label smoothing, does not produce the same effect.