Inference-time preference elicitation in multi-agent LLMs
Bringing ideas from social choice to how multi-agent LLM systems reach collective answers. Ongoing work.
Independent research, ongoing
Multi-agent LLM systems usually combine their agents’ answers by simple majority vote. I am exploring how eliciting more information from the agents at inference time, and aggregating it with ideas from social choice, can help these systems reach better collective answers.
Ongoing work. More details coming soon.