Reproduce “Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wangshu Zhu, Qianwen Guo, Yushun Dong
The authors define Phantom Merge (PM) as claims unsupported by evidence for an agent’s chosen entity, even when supported elsewhere. They introduce AGR-slot to rank claims and FAC to correct flagged claims using chosen-entity evidence, or remove them. On held-out Shopping data, AGR-slot reaches 96.2% AUROC; FAC lowers trajectory-level PM from 51.4% to 6.2% in complete-intervention evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.