StockDuty ← Dashboard Oct 16, 2026 06:50 AM ET
SignalTech/AI

Mosaic: Multi-Session Agentic Reinforcement Learning

Mosaic: Multi-Session Agentic Reinforcement Learning

What happened: Researchers introduced Multi-Session Agentic Reinforcement Learning, a setting where a rollout can span an arbitrary number of sessions, and proposed Mosaic, a critic-free hierarchical optimization method for it. Key numbers: Evaluated on BrowseComp with Qwen3.5-4B; training used 1,000 deep-search tasks from REDSearcher. Why it matters: Mosaic reportedly outperformed the broadcasting baseline in both context compaction and subagent forking scenarios while using fewer sessions on average, with gains in session utility and allocation.
Researchers introduced Multi-Session Agentic Reinforcement Learning, a setting where a rollout can span an arbitrary number of sessions, and proposed Mosaic, a critic-free hierarchical optimization method for it.
Mosaic reportedly outperformed the broadcasting baseline in both context compaction and subagent forking scenarios while using fewer sessions on average, with gains in session utility and allocation.
Sources