Abstract:Large language models (LLMs) with a large number of parameters achieve strong performance but are often prohibitively expensive to deploy. Recent work explores using teams of smaller, more efficient LLMs that collectively match or even outperform a single large model. However, jointly updating multiple agents introduces compounding distribution shifts, making coordination and stability during training difficult. We address this by introducing Sequential Agent Tuning (SAT), a coordinator-free training paradigm. SAT represents the team as a factorized policy and employs block-coordinate updates over agents, enabling scalable, decentralized training without a central controller. Specifically, we develop a sequence-aware, on-policy advantage estimator that conditions on the evolving team policy, coupled with per-agent KL trust regions that isolate occupancy drift. Theoretically, this framework provides two critical guarantees. First, it ensures monotonic improvement, stabilizing the training process. Second, it establishes provable plug-and-play invariance: any agent can be upgraded to a stronger model without retraining the rest of the team, with a formal guarantee that the performance bound improves. Empirically, a team of three 4B agents (12B total) trained with SAT surpasses the much larger Qwen3-32B on AIME24/25 benchmarks by 3.9\% on average. We validate our plug-and-play theory by swapping in two 8B agents, which boosts the composite score by 10.4\%. We provide code and appendix of proof at this https URL
| Comments: | Published at AAMAS 2026 |
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2605.05216 [cs.LG] |
| (or arXiv:2605.05216v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.05216 arXiv-issued DOI via DataCite |
From: Yi Xie [view email]
[v1]
Fri, 17 Apr 2026 01:45:30 UTC (100 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。