The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management
Published in DL4C Workshop at NeurIPS 2025, 2025
We show that simple sliding window based observation masking matches the performance of LLM-based summarization for managing agent context, at a fraction of the computational cost.
Recommended citation: Lindenbauer, T., Slinko, I., Felder, L., Bogomolov, E., & Zharov, Y. (2025). "The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management." Deep Learning for Code Workshop (DL4C) at NeurIPS 2025.
Download Paper