Back to Project
M

Mosaic

Project Oct 16, 2026

A research method that trains agents across context compaction and subagent sessions using session-aware reinforcement learning.

Visit primary link 2 sources attached
Mosaic authors' training pipeline diagram showing successful and failed trajectory groups, session-specific normalization and a policy update.
Figure 2 from the Mosaic authors' paper: the proposed multi-session reinforcement learning pipeline. This is a method diagram, not an independently reproduced benchmark. Resized variants preserve the complete image. Source: Yilong Xu and Mosaic coauthors
ai-agentsagent-ecosystemverification

What happened lately

Mosaic timeline

Mosaic is a multi-session agentic reinforcement learning method introduced in an October 15, 2026 preprint. It adjusts learning signals using session counts and estimates the contribution of individual sessions, including context compaction and subagent work.

The authors evaluate a Qwen3.5-4B agent on BrowseComp with Serper search and Jina Reader tools. Reported gains concern this experimental setup; the project does not establish a general improvement to every coding agent or a drop-in production extension.