EvoArena and EvoMem: A New Approach to Maintaining Robustness of LLM Agents in Dynamic Environments

This article introduces the EvoArena benchmark suite and EvoMem memory paradigm, aiming to address the robustness issue of LLM agents in dynamically changing environments. EvoArena simulates real-world environmental evolution (such as terminal command changes, software API updates, and shifts in social preferences), while EvoMem tracks environmental changes through a patch-based memory structure and preserves the evolution history. Experiments show that both can effectively improve the performance of agents in both dynamic and static environments.

Existing LLM agents perform well in static benchmarks, but real-world environments are dynamic (software version updates, API interface changes, social preference evolution). Current evaluation methods ignore the impact of environmental evolution, leading to agents relying on static memory being prone to failure in actual deployment, especially in long-term operation scenarios.

EvoArena is a benchmark suite designed for dynamic environments, covering three core domains: terminal environments (command line syntax evolution), software environments (API/interface changes), and social preference environments (user preference adjustments). Its "chain task" design requires agents to complete a series of dependent evolutionary subtasks. Experiments show that mainstream agents have an average accuracy of only 39.6% on EvoArena, exposing the shortcomings of existing methods.

EvoMem uses a patch-based memory structure; instead of directly overwriting old memories, it records patches of environmental changes. Core mechanisms include: memory version control (drawing on software version management), differential encoding (efficiently storing version differences), and selective retrieval (obtaining historical memories on demand). This allows agents to trace the evolution of the environment, reason about the impact of changes, and retain complete evidence.

In EvoArena tests, EvoMem brings significant improvements: overall performance increases by 1.5%, and chain task accuracy increases by 3.7%. It also performs well in static benchmarks: a 6.1% improvement on the GAIA benchmark and a 4.8% improvement on the LoCoMo benchmark. Mechanism analysis shows that EvoMem improves evidence capture, state integrity, and reasoning chains.

Implications of this research for agent deployment: dynamic environment testing should be included in evaluations, memory architectures should be redesigned (introducing version control), and continuous learning strategies should be adopted. Future directions include expanding EvoArena to multimodal scenarios, combining meta-learning to accelerate adaptation, and researching human-machine collaboration to guide evolution.