PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments

Physical intelligence requires agents to autonomously adapt their mechanical designs when environmental conditions shift. While self-evolving LLMs show promise, existing benchmarks isolate the required skills: physical reasoning benchmarks test passive understanding without design, design benchmarks assume fixed physics, and environment-varying suites omit self-evolution. We introduce PACE-Bench (**P**hysics **A**daptation via **C**ode **E**volution), where agents receive a code-driven design that works in a source environment but fails after physical parameters shift, and must revise it using only sandbox feedback within 20 attempts. Across 36 tasks $\times$ 4 mutation stages (144 task pairs) spanning six physics domains, we evaluate 3 model scales over 10 self-evolving methods from four paradigms. Overall performance is low: the strongest configuration reaches only 35.9\% Pass@2. Three findings emerge. First, simulator-grounded feedback is essential: unverified self-revision compounds errors rather than correcting them. Second, premature convergence is the dominant failure mode, and all methods face an exploration-exploitation tradeoff. Third, adaptive design, not physical inference, is the primary bottleneck: even disclosing all physical changes does not improve performance. PACE-Bench fills a critical gap by evaluating physical adaptation under dynamic environments, providing foundation for measuring progress in code-driven physical intelligence.