Who May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions

Multimodal language systems often include a reviewer to refine the decisions of a primary agent, yet the reviewer can both correct mistakes and overturn correct choices. The authors frame the issue as one of institutional jurisdiction: given identical proposals, task information, model, and review output, when does the reviewer possess the authority to bind the decision? They contrast broad veto power with an evidence‑based gate that only permits replacement after generating a mechanically valid record of a violation of a binding constraint and a compliant repair. In experiments with two open‑source models, 96 spatial‑policy tasks, and a 96‑task transfer benchmark, the evidence filter markedly reduces the corruption of correct proposals while preserving most violation repairs. The balance between correction and corruption depends on the rule, the reviewer’s state changes, and the original proposal. In a prospective validation, the reviewer can choose to keep, replace, or escalade the decision; under this design, the Qwen model stops corrupting correct proposals, whereas Mistral continues to replace all 48 correct proposals and rejects repairs focused solely on well‑being that Mistral frequently attempts. The study shows that the reviewer’s behavior is governed by competence, action interface, and binding jurisdiction, rather than an independent model safeguard. This research is important because it clarifies how to design review mechanisms that minimize errors without introducing new biases, a key factor for the reliability of collaborative AI systems.

Read the original article on arXiv cs.AI

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.