EUVIMEDEUVIMED entdecken
Crossref

Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning

Wei Yang, Jesse Thomason

Proceedings of the AAAI Conference on Artificial Intelligence · 14.03.2026 · 40 (35) · 29820-29828

DOI: 10.1609/aaai.v40i35.40228 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 14.03.2026. Verfasst von Wei Yang und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Multi-agent systems of large language models (LLMs) show promise for complex reasoning, but their effectiveness is often limited by fixed collaboration protocols. These frameworks typically focus on macro-level orchestration while overlooking agents’ internal deliberative capabilities. This critical meta-cognitive blindspot treats agents as passive executors unable to adapt their strategy based on internal cognitive states like uncertainty or confidence. We introduce the Meta-Policy Deliberation Framework (MPDF), where agents learn a decentralized policy over a set of high-level meta-cognitive actions: Persist, Refine, and Concede. To overcome the instability of traditional policy gradients in this setting, we develop SoftRankPO, a novel reinforcement learning algorithm. SoftRankPO stabilizes training by shaping advantages based on the rank of rewards mapped through smooth normal quantiles, making the learning process robust to reward variance. Experiments show that MPDF with SoftRankPO achieves a 4-5% absolute gain in average accuracy across six mathematical and general reasoning benchmarks compared to state-of-the-art heuristic and learning-based multi-agent reasoning algorithms. Our work presents a paradigm for learning adaptive, meta-cognitive policies for multi-agent LLM systems, shifting the focus from designing fixed protocols to learning dynamic, deliberative strategies.

Zitieren

APA

Yang, W., & Thomason, J. (2026). Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(35), 29820-29828. https://doi.org/10.1609/aaai.v40i35.40228

Vancouver

Yang W, Thomason J. Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(35):29820-29828. doi:10.1609/aaai.v40i35.40228

BibTeX
@article{Yang2026,
  author = {Yang, Wei and Thomason, Jesse},
  title = {Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year = {2026},
  volume = {40},
  number = {35},
  pages = {29820--29828},
  doi = {10.1609/aaai.v40i35.40228}
}
Kostenlos starten →