The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management

We challenge the assumption that sophisticated LLM-based summarization is necessary for managing long agent contexts. Through systematic evaluation, we demonstrate that simple observation masking — a sliding window based approach that selectively drops parts of the context — performs on par with LLM summarization while being substantially cheaper to run. The findings suggest that the field may be over-indexing on complex solutions for context management.

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