EvoArena
An AI agent evaluation benchmark proposed by the National University of Singapore, focused on assessing agents' continuous adaptation capabilities in dynamically evolving environments. It encompasses three scenarios: Terminal, Software Engineering (SWE), and personalized dialogue.
Timeline (last 90 days)
EvoArena was proposed by a team led by PhD student Xu Xundong from the National University of Singapore, and is used to evaluate the continuous adaptation ability of AI agents in dynamically changing environments.
An agent's behavior is the result of the joint action of the model, tools, memory, and environmental state; even if the model, tools, memory, and task instructions remain unchanged, a change in the environment will cause behavioral changes.
EvoArena focuses on Environment Evolution, which is distinct from Dynamic Task generation and Agent Evolution.
EvoArena constructs three types of dynamic environment scenarios: Terminal-Bench Evolve (workflow scenario), SWE-Chain Evolve (software engineering scenario), and Persona-Mem Evolve (social intelligence scenario).
In EvoArena testing, GPT-5.5, which achieved the highest Step Level score, had an accuracy of just over 50%, while Gemini-3.1 Pro and GLM-5.1, which performed best at the Chain Level, reached only approximately 28%–29%.
All Facts (5)
EvoArena was proposed by a team led by PhD student Xu Xundong from the National University of Singapore, and is used to evaluate the continuous adaptation ability of AI agents in dynamically changing environments.
50%UnverifiedAn agent's behavior is the result of the joint action of the model, tools, memory, and environmental state; even if the model, tools, memory, and task instructions remain unchanged, a change in the environment will cause behavioral changes.
50%UnverifiedEvoArena focuses on Environment Evolution, which is distinct from Dynamic Task generation and Agent Evolution.
50%UnverifiedEvoArena constructs three types of dynamic environment scenarios: Terminal-Bench Evolve (workflow scenario), SWE-Chain Evolve (software engineering scenario), and Persona-Mem Evolve (social intelligence scenario).
50%UnverifiedIn EvoArena testing, GPT-5.5, which achieved the highest Step Level score, had an accuracy of just over 50%, while Gemini-3.1 Pro and GLM-5.1, which performed best at the Chain Level, reached only approximately 28%–29%.
50%