Reproduce “PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents” on OpenResearch

About this paper

Zhuohan Gu, Qizheng Zhang, Omar Khattab, Samuel Madden

PEEK is a system that employs a "context map" as an agent-side cache for large language models (LLMs), designed to store reusable orientation knowledge about extensive, recurring external contexts. The system enables LLM agents to perform tasks with improved quality and efficiency, achieving performance gains of 6.3% to 34.0% on the OOLONG benchmark and 6.0% to 14.0% in solving rate on the CL-bench benchmark, while reducing computational costs by 1.4 to 5.8 times compared to state-of-the-art baselines.

Key insights

• Actively maintaining a curated, structured "context map" with reusable orientation knowledge substantially improves LLM agent performance in long-context tasks, filling a gap in existing context management strategies.
• The modular design, which separates the diagnosis of useful contextual information (Distiller) from the actual editing and updating of the context map (Cartographer), is crucial for preventing noise and duplication, leading to more effective cache management.
• The benefits of this agent-side caching approach generalize across different underlying LLMs (e.g., GPT-5-mini, GPT-5.5, Qwen3) and agent architectures (e.g., RLM, OpenAI Codex), highlighting its fundamental value.

Summary by alphaXiv.