PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments
Summary
PACE-Bench introduces a simulator-grounded benchmark for evaluating self-evolving agents on physics adaptation tasks involving iterative code redesign after environmental mutations, revealing that mechanism redesign is a major bottleneck compared to parameter inference.
View Cached Full Text
Cached at: 08/18/26, 03:50 AM
Paper page - PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments
Source: https://huggingface.co/papers/2608.14441
Abstract
PACE-Bench evaluates self-evolving agents on physics adaptation tasks requiring iterative code redesign after environmental mutations, revealing that simulator-grounded reflection outperforms unverified self-revision but mechanism redesign remains a major bottleneck.
Self-evolving agentsimprove future behavior from interaction experience, yet existing evaluations typically optimize under fixed execution conditions and do not test recovery after those conditions change. To address this gap, we introducePACE-Bench(Physics Adaptation via Code Evolution), asimulator-grounded benchmarkof 144 source-to-target adaptation pairs across six physics domains. Each pair links a source environment to a mutated target environment with the same goal and interface. Acode-driven designthat succeeds in the source fails in the target, where agents must iteratively adapt it into a working target design usingdiagnostic sandbox feedbackwithin a limited attempt budget. We compare ten self-evolving methods from four paradigms. The benchmark remains far from saturated:Reflexion+ Qwen3-14B succeeds on only 35.9\% of full-benchmark pairs, while GPT-5.5 solves 66.7\% of the Statics subset under the full budget. Together, these results show that simulator-grounded reflection is more reliable than unverified self-revision, whilememory anchorsagents to early designs and broadtree searchexplores without converging. Even revealing exact physical changes does not raise the performance ceiling, pointing tomechanism redesignrather than parameter inference as the central bottleneck. Data and code are available at https://github.com/thunlp/PACE-Bench.
View arXiv pageView PDFProject pageGitHubAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2608.14441 in a model README.md to link it from this page.
Datasets citing this paper1
#### YuhaoZhan/PACE-Bench Viewer• Updated1 day ago • 1 • 60 • 1
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2608.14441 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Collider-Bench: Benchmarking AI Agents with Particle Physics Analysis Reproduction
Collider-Bench is a new benchmark that evaluates LLM agents on reproducing particle physics analyses from the Large Hadron Collider using only public papers and open software, requiring physical reasoning to fill missing implementation details.
PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents
The paper introduces PACEvolve++, a reinforcement learning framework that improves test-time policy adaptation for evolutionary search agents by decoupling hypothesis generation from execution.
BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution
BenchEvolver is an evolutionary framework that automatically generates harder coding problems from existing ones, creating challenging benchmarks that maintain validity and diversity while enabling model self-improvement and enhanced training performance.
PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
Introduces PAST-Bench, a benchmark for evaluating whether personal AI agents improve from retained experience across sessions, and Hermes+, an extension with targeted interventions. Finds improvement is real but uneven across capabilities and models.
Deep Research in Physical Sciences: A Multi-Agent Framework and Comprehensive Benchmark
This paper introduces PhySciBench, a benchmark of 200 expert-curated questions for physical sciences, and DelveAgent, a multi-agent framework that improves accuracy and reduces inference costs compared to baselines like Gemini Deep Research.