PACE-Bench: Code Evolution and Adaptation Benchmark in Dynamic Physical Environments

Background and Context

Current research on self-evolving agents primarily focuses on how systems improve future behavior through interaction experience. However, existing evaluation frameworks are typically confined to fixed execution conditions, failing to effectively test an agent's recovery and adaptability after sudden environmental changes. This gap makes it difficult to determine whether agents possess genuine robustness against dynamic shifts. To address this, the paper introduces PACE-Bench (Physics Adaptation via Code Evolution), a simulator-grounded benchmark designed to evaluate adaptability in dynamic physical environments. The benchmark is meticulously constructed with 144 source-to-target adaptation pairs spanning six distinct physics domains. Each pair links a source environment with a mutated target environment, maintaining identical goals and interfaces. This design ensures task consistency while introducing environmental parameter changes, forcing agents to move beyond simple strategy reuse and engage in substantive adaptation.

The core challenge lies in the fact that code-driven designs successful in the source environment often fail in the target. Agents must iteratively modify code within a limited trial budget, utilizing feedback from a diagnostic sandbox to achieve feasibility in the mutated environment. This setup closely mirrors real-world scenarios where environmental dynamics change, imposing higher demands on adaptive capabilities. By restricting the trial budget, the benchmark simulates resource-constrained real-world conditions, compelling agents to make efficient decisions within finite timeframes. This approach shifts the focus from static performance to dynamic resilience, providing a more rigorous test of an agent's ability to learn and reason from errors rather than merely memorizing or imitating previous successes.

Deep Analysis

Technically, PACE-Bench adopts a code-driven design paradigm, emphasizing adaptation through code modification rather than mere parameter adjustment. Agents interact with a diagnostic sandbox to receive feedback on design failure causes, enabling iterative code optimization. This mechanism requires agents to possess the ability to learn and reason from errors, moving beyond simple memory or imitation. The study compares ten self-evolution methods from four distinct paradigms, ranging from simple self-reflection to complex tree search strategies. These methods include Reflexion, memory-based approaches, and broad tree search. The significant performance differences among these methods reveal the strengths and weaknesses of each paradigm in handling dynamic environmental changes. For instance, Reflexion relies on an agent's reflection on past failures, while tree search methods explore multiple possible code modification paths to find solutions.

Experimental results indicate that PACE-Bench is far from saturation, suggesting substantial room for improvement in the field. Specifically, Reflexion combined with the Qwen3-14B model solved only 35.9% of adaptation pairs in the full benchmark. In contrast, GPT-5.5 achieved a success rate of 66.7% on the Statics subset within the full budget. These figures demonstrate that even state-of-the-art models face significant challenges when confronted with dynamic environmental changes. The diversity of methods compared not only enriches the benchmark's dimensions but also provides critical insights into the operational mechanisms of self-evolving agents. The limited trial budget further simulates real-world resource constraints, forcing agents to prioritize efficiency and accuracy in their decision-making processes.

Industry Impact

Ablation studies further reveal the impact of different mechanisms on performance. Simulator-based reflection proves more reliable than unverified self-correction, as it provides specific feedback information that helps agents accurately locate problems. In contrast, memory mechanisms tend to lock agents into early designs, making it difficult to escape local optima after environmental changes. While broad tree search explores more possibilities, it often lacks convergence, failing to find optimal solutions within limited budgets. A key finding is that even revealing exact physical changes to the agent does not significantly improve performance ceilings. This result points to a deeper issue: mechanism redesign, rather than simple parameter inference, is the core bottleneck for agent adaptation in dynamic environments. This implies that agents require higher-level abstraction and reconstruction capabilities, not just adjustments to existing parameters.

The introduction of PACE-Bench holds profound significance for both the open-source community and industrial deployment. It provides a standardized evaluation platform, enabling researchers to fairly compare different self-evolution methods in dynamic environments and drive progress in the field. The identified bottleneck—mechanism redesign over parameter inference—offers a clear direction for future research. In industrial applications, this benchmark helps assess agent robustness in real dynamic environments, such as robotic control and autonomous driving, providing a reliable basis for actual deployment. The public availability of PACE-Bench data and code fosters collaboration and innovation within the open-source community. By continuously improving and expanding the benchmark, researchers can gain a deeper understanding of the limitations of self-evolving agents and develop more powerful and flexible adaptation mechanisms.

Outlook

Future research may need to explore more advanced abstraction mechanisms, such as modular design and dynamic architecture adjustment, to enhance agent adaptability. The findings from PACE-Bench suggest that current methods are insufficient for handling complex dynamic changes, highlighting the need for new paradigms in agent design. The ability to perform mechanism redesign rather than relying on parameter inference will be crucial for developing agents that can operate effectively in unpredictable environments. This shift in focus could lead to the creation of more general and robust artificial intelligence systems capable of navigating complex, dynamic worlds.

As the field progresses, the integration of simulator-based feedback with advanced reasoning capabilities will likely become standard practice. The benchmark's emphasis on code evolution provides a tangible pathway for improving agent performance, moving beyond theoretical models to practical implementations. By addressing the core bottleneck of mechanism redesign, researchers can unlock new levels of adaptability and resilience in self-evolving agents. This work lays the foundation for building intelligent systems that can respond to the complexities of the real world, ensuring that future AI applications are not only efficient but also robust and adaptable to changing conditions.