Midjourney’s surprise: new research on making LLMs write more creatively
1 min read
Summary
Midjourney, the AI image generating company, has released a new research paper in collaboration with New York University (NYU) about training text-based large language models (LLMs) to improve creative output
Two new techniques were introduced, Diversified Direct Preference Optimization and Diversified Odds Ratio Preference Optimization, which are designed to increase the range of outputs while maintaining readability
The paper highlights that there is still scope for improvement in creative text-based LLMs, which to date have struggled to produce writing with multiple possible outputs and repetition of themes
The study used two base models, Meta’s Llama-3.1-8B and Mistral-7B-v0.