EUVIMEDEUVIMED entdecken
Crossref

When AI Agents See Each Other’s Answers, Their Answers Grow More Alike: A Controlled Study of Multi-Agent LLM Revision

Hideki Muto, Tetsuro Ogi, Takahiro Yakoh

International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)

DOI: 10.14569/ijacsa.2026.0171005 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Hideki Muto und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

AI systems can offer several answers when a question has no single correct solution. For people seeking alternatives, those answers need to differ. Multi-agent systems often let agents read one another’s answers during revision. Can agents keep their answers different when they read one another’s work? Here we show that reading other answers made the final answers more alike, even when agents were told to keep their answers separate. We copied the same five independent answers into conditions where agents could read zero, one or four other answers during four revision rounds. We tested 20 open-ended tasks with GPT-4o-mini and Qwen3-30B-A3B, with three repetitions per task and model. We compared the mean embedding distance among final answers at the task level. Relative to no access, distance fell by 8% and 15% with one and four other answers for GPT-4o-mini, and by 4% and 9% for Qwen. All four planned tests remained significant after adjustment for multiple tests. A second text encoder showed the same pattern for complete answers. In a post-hoc check, agents moved more toward the randomly selected answer they read than toward unread answers. Showing other answers also lengthened the input; this study cannot isolate their content effect or determine whether distinct ideas were lost. Qwen’s one-answer difference was near zero when only the first 128 tokens of each answer were measured. Access to other answers is therefore a consequential design setting. Systems intended to offer alternatives should assess both individual answers and the differences among them.

Zitieren

APA

Muto, H., Ogi, T., & Yakoh, T. (2026). When AI Agents See Each Other’s Answers, Their Answers Grow More Alike: A Controlled Study of Multi-Agent LLM Revision. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171005

Vancouver

Muto H, Ogi T, Yakoh T. When AI Agents See Each Other’s Answers, Their Answers Grow More Alike: A Controlled Study of Multi-Agent LLM Revision. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171005

BibTeX
@article{Muto2026,
  author = {Muto, Hideki and Ogi, Tetsuro and Yakoh, Takahiro},
  title = {When AI Agents See Each Other’s Answers, Their Answers Grow More Alike: A Controlled Study of Multi-Agent LLM Revision},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2026},
  volume = {17},
  number = {10},
  doi = {10.14569/ijacsa.2026.0171005}
}
Kostenlos starten →