Research · Oct 28, 2026, 00:00 UTC

Multiscale Policy Optimization improves RL training for long-horizon LLM agents

MPO anchors policy updates at the turn level using Multiscale Advantage Estimation and turn-level importance sampling. Reported gains across search, tool-use and automation tasks use 4B and 30B-A3B models.

• Topicresearch
• Topicagents