AgentDropoutV2: Optimizing Information Flow in Multi-Agent Systems via Test-Time Rectify-or-Reject Pruning
Yutong Wang, Siyuan Xiong, Xuebo Liu, Wenkang Zhou, Liang Ding, Miao Zhang, Min Zhang · 2026
Read AgentDropoutV2: Optimizing Information Flow in Multi-Agent Systems via Test-Time Rectify-or-Reject Pruning by Yutong Wang, Siyuan Xiong, Xuebo Liu, Wenkang Zhou, Liang Ding, Miao Zhang, Min Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-23258.txt.
Description
Yutong Wang and six co-authors present MASRubric, a system for auditing information flow in multi-agent systems before errors spread. Such systems excel at complex reasoning, but downstream agents build on an intermediate agent's mistakes. Checking intermediate messages needs an explicit standard, yet rubrics are usually written by experts or against a reference answer, neither on hand for unseen messages.
MASRubric instead distills failed trajectories offline into a reusable bank of pitfall criteria, each naming a repeat error's root mistake, the cases where it arises, and the check that exposes it. Online, relevant criteria are retrieved and checked message by message, and the pass rate decides whether a message is broadcast, returned for revision with diagnostic feedback, or withheld. Tests show accuracy gains up to 2.83 points on math tasks and 1.74 points on code tasks, and the bank transfers to a stronger backbone system without re-mining.
Catalog record
| CREATOR | Yutong Wang, Siyuan Xiong, Xuebo Liu, Wenkang Zhou, Liang Ding, Miao Zhang, Min Zhang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-23258