I think this market fundamentally cannot resolve as stated because Chatbot Arena includes closed LLMs which do not disclose the details of how they have been finetuned. GPT-4 and Bard might already be using DPO and we wouldn't know about it. To my chagrin, some people (for example Nous Research) categorize DPO as a variant of RLHF, so there's plausible deniability whenever OpenAI or Google refer to their finetuning as "RLHF"

@NoraBelrose that's a good point. If there's uncertainty about close source models at the end of the year I might resolve N/A or push the deadline in the hope that the information will surface later.

Mixtral-Instruct was trained with DPO