Diverse, not Short: A Length-Controlled Data Selection Strategy for Improving Response Diversity of Language Models

TL;DR

This paper introduces Diverse-NS, a length-controlled data selection strategy that enhances response diversity in language models without sacrificing output length, leading to more expressive and creative responses.

Contribution

The paper presents a novel length-controlled data filtering method that improves response diversity in language models, addressing length bias in existing diversity metrics and reward models.

Findings

Diverse-NS significantly increases lexical and semantic diversity in responses.

The method achieves these improvements with only 3,000 preference pairs.

Smaller models like Olmo-2-7B can effectively teach larger models diversity.

Abstract

Diverse language model responses are crucial for creative generation, open-ended tasks, and self-improvement training. We show that common diversity metrics, and even reward models used for preference optimization, systematically bias models toward shorter outputs, limiting expressiveness. To address this, we introduce Diverse, not Short (Diverse-NS), a length-controlled data selection strategy that improves response diversity while maintaining length parity. By generating and filtering preference data that balances diversity, quality, and length, Diverse-NS enables effective training using only 3,000 preference pairs. Applied to LLaMA-3.1-8B and the Olmo-2 family, Diverse-NS substantially enhances lexical and semantic diversity. We show consistent improvement in diversity with minor reduction or gains in response quality on four creative generation tasks: Divergent Associations, Persona Generation, Alternate Uses, and Creative Writing. Surprisingly, experiments with the Olmo-2 model family (7B, and 13B) show that smaller models like Olmo-2-7B can serve as effective "diversity teachers" for larger models. By explicitly addressing length bias, our method efficiently pushes models toward more diverse and expressive outputs.