$ cat wiki/papers/2026/2610.22174-mosaic.md
Mosaic
TL;DR
Mosaic trains agents whose trajectories span multiple sessions, assigning credit to individual sessions instead of broadcasting one trajectory advantage everywhere. It connects Agentic Reinforcement Learning with Context Compaction. (source)
Authors & Org
Yilong Xu, Minrui Xu, Xinyu Zhou, Xiang Long, Haochen Tan, Zhijiang Guo. Affiliations are not established by the abstract record read here. First submitted 2026-10-15. (source)
Method
Context compaction and subagent forking break the assumption that an entire trajectory shares one continuously growing context. Mosaic uses critic-free hierarchical optimization: trajectory-level advantage estimation incorporates session usage, while session-level advantages reflect estimated marginal contributions. The setting permits an arbitrary number of sessions per trajectory. (source)
Results
On BrowseComp with Qwen3.5-4B, the authors report better performance than advantage broadcasting in both context-compaction and subagent-forking scenarios, with fewer sessions on average. The abstract gives no numerical success rate, improvement magnitude, token cost or uncertainty estimate. Fewer sessions alone does not establish lower total compute. (source)
Significance
The paper treats session allocation as a trainable part of Agents (LLM Agents), rather than merely an inference-time scaffold setting. Its comparison makes credit assignment across session boundaries an explicit training question. (source)
Open Questions
The abstract does not establish transfer beyond the named model and benchmark, how marginal contributions are estimated, or the effect of changing session budgets. Full paper and code were not reviewed. (source)
Cite
“Mosaic: Multi-Session Agentic Reinforcement Learning.” arXiv:2610.22174. This note uses the abstract and metadata in the snapshot.