Diverse, not Short: A Length-Controlled Data Selection Strategy for Improving Response Diversity of Language Models
Vijeta Deshpande, Debasmita Ghose, John D. Patterson, Roger Beaty, Anna Rumshisky
TL;DR
This work addresses the problem that alignment and preference optimization can reduce language model output diversity by introducing Diverse-NS, a length-controlled data selection pipeline that constructs diverse yet length-parallel preference data. The method generates two continuations per prompt via sequential prompting, filters pairs using diversity and quality signals while enforcing near-equal length, and trains with Direct Preference Optimization using only about $3{,}000$ pairs. Empirical results across Llama-3.1-8B and Olmo-2 families show notable lexical and semantic diversity gains across four creative tasks (DAT, PGT, AUT, CWT) with generally preserved or modestly improved quality; smaller models can even impart diversity benefits to larger ones (small-to-large transfer). To enable fair comparisons, the authors introduce the Δ Diversity Decile ($\Delta DD$) metric, a length-aware evaluation that mitigates length bias in diversity assessments, and demonstrate that lightweight proxies (TTR, MAAS) can substitute for expensive metrics with minimal loss. The work thus provides a scalable, compute-conscious pathway to more expressive, diverse outputs in aligned LLMs, with practical implications for creative generation and model improvement.
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.