New RL method trains AI agents to summarize past steps and keep working in long tasks

Original: CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents

Source: arxiv.org ↗