Proposed multi-agent training targets repetitive LLM answers

The proposal’s authors say LLMs give strikingly similar answers to the same prompt, even across different model families. They propose multi-agent reinforcement learning during post-training, with different roles within one model learning to diversify responses only if those responses meet a quality threshold. The authors report improved quality and diversity over several competing diversity-enhancing methods across four benchmarks spanning scientific ideation and creative writing.