Adaptive Context Management Boosts LLM Agent Performance
Researchers introduce AdaCoM, a system that trains an external LLM to manage context for frozen agents via reinforcement learning, improving performance on long-horizon tasks like web search and deep research. AdaCoM prunes stale content while preserving task constraints, revealing a Fidelity-Reliability Trade-off.
Key facts
- AdaCoM trains an external LLM to manage context of a frozen agent.
- Uses flexible modification actions and end-to-end reinforcement learning.
- Tested on web search and deep research benchmarks.
- Improves performance by preserving task constraints and progress.
- Prunes stale content to avoid long-context degradation.
- Reveals a Fidelity-Reliability Trade-off.
- Prior methods require training the agent itself, impractical for closed-source agents.
- Different agents may require different context management strategies.
Entities
Institutions
- arXiv