Title: ACON: Enhancing Long-Horizon LLM Agent Performance via Context Compression Optimization

Abstract:

As large language models (LLMs) are increasingly utilized as autonomous agents within dynamic, real-world settings, their effectiveness hinges on the ability to maintain accurate logs of both actions taken and observations made. However, the uncontrolled expansion of context windows in extended agentic workflows creates two significant challenges: excessive memory consumption during inference and a decline in reasoning capabilities caused by the presence of superfluous data. Current compression techniques are insufficient for these demands, frequently depending on fragile heuristics or necessitating parameter updates that are unfeasible for proprietary or massive-scale LLMs.

To address these limitations, we present Agent Context Optimization (ACON), a comprehensive framework designed to optimally condense both historical records and observations into compact, high-value representations. Unlike previous approaches, ACON operates through optimization within the natural language space. It systematically refines compression directives by analyzing agent failures, thereby safeguarding essential state information without requiring model fine-tuning. To further reduce computational burdens, we distill these optimized compressors into more efficient, smaller-scale models.

Evaluations conducted on AppWorld, OfficeBench, and Multi-objective QA benchmarks reveal that ACON decreases peak token consumption by 26% to 54% while simultaneously boosting task success rates compared to existing compression methods. Notably, the framework allows smaller LLMs to operate effectively as long-horizon agents, delivering performance gains of up to 46% by reducing context distraction. The source code is publicly accessible at https://github.com/microsoft/acon.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC