Training agents to compress their own context for longer reasoning horizons
Researchers propose CompactionRL, a reinforcement learning framework that trains language model agents to handle tasks exceeding typical context windows by learning to compress interaction history on the fly. The method jointly optimizes task performance and summary quality through token-level normalization and cross-trajectory advantage estimation, enabling agents to maintain coherence across extended rollouts. This addresses a fundamental scaling bottleneck for agentic LLMs: as task horizons grow, trajectory data quickly exhausts finite context, forcing either task truncation or architectural redesign. CompactionRL's approach to teach models when and how to summarize their own state could reshape how long-horizon reasoning systems are trained, particularly for real-world applications requiring sustained multi-step planning.
Modelwire context
ExplainerThe genuinely underappreciated detail here is that CompactionRL doesn't just handle long contexts at inference time, it trains the compression behavior itself through reinforcement learning, meaning the model learns a policy for when to summarize, not just how. That distinction matters because most existing approaches treat context management as a post-hoc engineering patch rather than a learned capability.
This connects directly to the asynchronous RLHF work covered in 'Staleness-Learning Rate Scaling Laws for Asynchronous RLHF' (July 1), which showed how trajectory data quality degrades under real training conditions. CompactionRL is essentially attacking a complementary bottleneck: not staleness across parallel workers, but coherence loss across sequential steps within a single rollout. Both papers are circling the same core problem, which is that RL training pipelines for LLMs break down as the gap between data generation and policy optimization widens, whether in time or in token count. The MAGNET story on long-form narrative generation also touched this pressure point, noting that structured agent approaches outperform raw prompting precisely when task horizons extend.
The critical test is whether CompactionRL's cross-trajectory advantage estimation holds up on tasks with genuinely sparse rewards across hundreds of steps, such as competitive coding or long-horizon web navigation benchmarks. If published ablations show performance degrading sharply when reward signals are delayed beyond a certain horizon length, the method's practical range is narrower than the framing suggests.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsCompactionRL · LLM agents · reinforcement learning
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.