TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination

Yi Xie, Siao Liu, Falong Fan, Yuanqi Yao, Siyang Cao, Yue Zhao, Bo Liu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:139507-139546, 2026.

Abstract

Multi-agent LLM systems can improve reasoning and tool use, yet recent evidence shows their gains are often unstable and sensitive to interaction design. A promising direction is to train collaboration, but team post-training introduces a moving-target effect: when agents interact through a shared context, updating one agent shifts the context distribution faced by the others, which can regress coordination under naive sequential updates. We propose TeamTR, a trust-region framework for fine-tuning heterogeneous LLM teams that explicitly controls this occupancy shift. TeamTR evaluates each agent update on rollouts from the intermediate team induced by partially applied updates, and enforces per-agent trust regions via a token-decomposed reverse KL that is directly monitorable from those rollouts. This yields population-level per-update and per-stage improvement lower bounds whose functional form applies to any realized update order, and motivates a practical certificate proxy computed from logged surrogates and KL terms. We instantiate TeamTR for router-based text handoff with sequence-level returns and bounded group-normalized advantages, and show empirically that it mitigates coordination regressions, improves training stability across heterogeneous teams, and supports modular component replacement via a trust-region alignment step.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-xie26o, title = {{T}eam{TR}: Trust-Region Fine-Tuning for Multi-Agent {LLM} Coordination}, author = {Xie, Yi and Liu, Siao and Fan, Falong and Yao, Yuanqi and Cao, Siyang and Zhao, Yue and Liu, Bo}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {139507--139546}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/xie26o/xie26o.pdf}, url = {https://proceedings.mlr.press/v306/xie26o.html}, abstract = {Multi-agent LLM systems can improve reasoning and tool use, yet recent evidence shows their gains are often unstable and sensitive to interaction design. A promising direction is to train collaboration, but team post-training introduces a moving-target effect: when agents interact through a shared context, updating one agent shifts the context distribution faced by the others, which can regress coordination under naive sequential updates. We propose TeamTR, a trust-region framework for fine-tuning heterogeneous LLM teams that explicitly controls this occupancy shift. TeamTR evaluates each agent update on rollouts from the intermediate team induced by partially applied updates, and enforces per-agent trust regions via a token-decomposed reverse KL that is directly monitorable from those rollouts. This yields population-level per-update and per-stage improvement lower bounds whose functional form applies to any realized update order, and motivates a practical certificate proxy computed from logged surrogates and KL terms. We instantiate TeamTR for router-based text handoff with sequence-level returns and bounded group-normalized advantages, and show empirically that it mitigates coordination regressions, improves training stability across heterogeneous teams, and supports modular component replacement via a trust-region alignment step.} }
Endnote
%0 Conference Paper %T TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination %A Yi Xie %A Siao Liu %A Falong Fan %A Yuanqi Yao %A Siyang Cao %A Yue Zhao %A Bo Liu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-xie26o %I PMLR %P 139507--139546 %U https://proceedings.mlr.press/v306/xie26o.html %V 306 %X Multi-agent LLM systems can improve reasoning and tool use, yet recent evidence shows their gains are often unstable and sensitive to interaction design. A promising direction is to train collaboration, but team post-training introduces a moving-target effect: when agents interact through a shared context, updating one agent shifts the context distribution faced by the others, which can regress coordination under naive sequential updates. We propose TeamTR, a trust-region framework for fine-tuning heterogeneous LLM teams that explicitly controls this occupancy shift. TeamTR evaluates each agent update on rollouts from the intermediate team induced by partially applied updates, and enforces per-agent trust regions via a token-decomposed reverse KL that is directly monitorable from those rollouts. This yields population-level per-update and per-stage improvement lower bounds whose functional form applies to any realized update order, and motivates a practical certificate proxy computed from logged surrogates and KL terms. We instantiate TeamTR for router-based text handoff with sequence-level returns and bounded group-normalized advantages, and show empirically that it mitigates coordination regressions, improves training stability across heterogeneous teams, and supports modular component replacement via a trust-region alignment step.
APA
Xie, Y., Liu, S., Fan, F., Yao, Y., Cao, S., Zhao, Y. & Liu, B.. (2026). TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:139507-139546 Available from https://proceedings.mlr.press/v306/xie26o.html.

Related Material