PaceMaker Agents Flag User Requests That Clash With Context
TL;DR
- PACE, accepted to Findings of EMNLP 2026, pairs user requests with persona-based knowledge base facts to test whether assistants can refuse ill-fitting asks.
- Its companion multi-agent system PaceMaker chains query reformulation, multi-hop graph traversal, and conflict-aware filtering to pull evidence against a request.
- The abstract claims PaceMaker beats prior approaches on both evidence retrieval quality and conflict decision accuracy, without publishing per-metric figures.
A new paper argues personalized assistants should refuse some requests not because the requests are forbidden, but because the user's own circumstances make them wrong.
"Personalized assistants should not only comply with user requests but also assess whether those requests are appropriate given the user's current circumstances," the authors write in the abstract of PACE, accepted to Findings of EMNLP 2026.
The paper releases a dataset also called PACE, pairing user requests with persona-based knowledge base facts, and a companion multi-agent system named PaceMaker. In PaceMaker, "specialized agents coordinate across query reformulation, multi-hop graph traversal, and conflict-aware filtering to retrieve contextually decisive evidence." The abstract reports the system "consistently outperforms existing approaches" on evidence retrieval quality and conflict decision accuracy. It does not publish per-metric numbers. The multi-agent framing lands amid 409 agent stories in our tracker over the past 90 days.
The authors, Yoojin Kim, Jihyoung Jang and Hyounghun Kim, frame prior work as overfit to execution: assistants that follow an instruction well but never ask whether the instruction fits.
Originally reported by huggingface.co
Read the original article →Original headline: PACE Paper Builds PaceMaker Multi-Agent System to Surface Hidden Conflicts in User Requests