Learning Self-Orchestration in Decentralized Agent Swarms: Market-Based Post-Training (MPT)

Forum topic · 小凯 · 2026-10-17

Summary

This post introduces market-based post-training (MPT), a framework from Guowei Xu, Zhenting Qi, Amani Kiruga, Himabindu Lakkaraju, and Yilun Du (arXiv:2610.22322) for jointly learning action policies and self-orchestration in decentralized agent swarms. Unlike existing multi-agent systems that train agents within fixed orchestration structures—separating individual capability learning from collective behavior organization—MPT lets agents express willingness to participate through bids. Bids determine who acts and how collective outcomes provide local training credit, enabling capabilities and participation patterns to co-adapt without a central coordinating agent. Evaluated on embodied crafting, robot navigation, and deep research tasks, MPT outperforms monolithic post-training, and its learned self-orchestration beats alternative coordination strategies. Trained swarms adapt faster to related tasks and perform better at larger inference-time swarm sizes. Code and trained models are available on GitHub.

Paper Overview

Field: NLP Authors: Guowei Xu, Zhenting Qi, Amani Kiruga, Himabindu Lakkaraju, Yilun Du Published: 2026-10-15 arXiv: 2610.22322

English Translation of Summary

The capabilities of an agent swarm depend on both its individual members and how they coordinate. Yet existing multi-agent systems often train agents within fixed orchestration structures, separating the learning of individual capabilities from the organization of collective behavior. We introduce market-based post-training (MPT), a framework for jointly learning action policies and self-orchestration in decentralized agent swarms. Agents express their willingness to participate through bids, which determine who acts and how collective outcomes provide local training credit. Learning both action and bidding behavior allows individual capabilities and patterns of participation to adapt together without a central coordinating agent. We evaluate MPT on embodied crafting, robot navigation, and deep research, showing it outperforms monolithic post-training, and that learned self-orchestration outperforms alternative coordination strategies. Trained swarms adapt faster to related tasks and perform better at larger inference-time swarm sizes.

Resources

Paper: https://arxiv.org/abs/2610.22322
Code and trained models: https://github.com/Embodied-Minds-Lab/MPT

---

*Auto-collected on 2026-10-17.*