Reproduce “Optimizing Diversity and Quality through Base-Aligned Model Collaboration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yichen Wang, Chenghao Yang, Tenghao Huang, Muhao Chen, Jonathan May, Mina Lee
An inference-time framework called Base-Aligned Model Collaboration (BACO) was developed to dynamically combine a base large language model with its aligned counterpart, optimizing both output diversity and quality. The framework achieved a 0.142 improvement in Coverage and 32.7% Dominance on the diversity-quality Pareto frontier, with human evaluations noting higher perceived quality, diversity, and creativity in generated texts.
Key insights
- Dynamic, token-level collaboration between a base model (diversity-rich) and an aligned model (quality-rich) effectively optimizes the diversity-quality trade-off, outperforming either model in isolation.
- Router strategies based on a combination of logit uncertainty (e.g., max probability) and content-based rules (e.g., punctuation/function words) are most effective for achieving both semantic diversity and output coherence.
- The framework provides controllable adjustment of the diversity-quality balance through a simple threshold parameter, enabling fine-tuned output generation for diverse open-ended tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.