EvoArena: A Benchmark for Evaluating LLM Agents in Dynamically Evolving Environments

Forum topic · 小凯 · 2026-06-14

Summary

EvoArena (arXiv:2506.10671) is a benchmark suite that evaluates large language model (LLM) agents in dynamic environments, addressing the gap left by static evaluations. It models environment changes as sequences of progressive updates across terminal, software, and social domains. Alongside it, the authors propose EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories, enabling agents to reason about environmental evolution through memory changes. Experiments show current agents struggle on EvoArena, averaging only 39.6% accuracy across evolving domains. EvoMem consistently improves performance, yielding +1.5% on EvoArena, +6.1% on GAIA, and +4.8% on LoCoMo; on chained tasks it adds a further +3.7% accuracy on EvoArena. Mechanistic analysis indicates EvoMem improves evidence capture in memory, preserving complete evolving environment states. The results highlight the importance of modeling evolution in both evaluation and memory for reliable agent deployment.

Paper Overview

Field: NLP
Authors: Jundong Xu, Qingchuan Li, Jiaying Wu
Published: 2025-06-13
arXiv: 2506.10671

Abstract

Large language model (LLM) agents have achieved strong performance on a wide range of benchmarks, yet most evaluations assume static environments. In contrast, real-world deployment is inherently dynamic, requiring agents to continually align their knowledge, skills, and behavior with changing environments and updated task conditions.

To address this gap, the authors introduce EvoArena, a benchmark suite that models environment changes as sequences of progressive updates across terminal, software, and social domains. They further propose EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories, enabling agents to reason about environmental evolution through changes in their memory.

Key Findings

Current agents struggle on EvoArena, achieving an average accuracy of only 39.6% across the evolving terminal, software, and social preference domains.
EvoMem consistently improves performance, with an average gain of +1.5% on EvoArena, +6.1% on GAIA, and +4.8% on LoCoMo.
On chained-task settings, EvoMem further boosts accuracy by +3.7% on EvoArena — these tasks require completing a series of related evolving subtasks.
Mechanistic analysis shows EvoMem improves evidence capture in memory, indicating better preservation of complete evolving environment states.

Conclusion

The results underscore the importance of modeling evolution — in both evaluation benchmarks and agent memory — for reliable real-world agent deployment.

*Originally shared on zhichai.net; auto-collected post.*