SafeRun: Enabling Determinism in LLM Planning for Running

Researchers have introduced SafeRun, a framework that forces large language models to obey strict safety rules when generating running plans, achieving a 100% safety score in benchmark tests while standard LLM-based approaches fell short [1]. Large language models, or LLMs, are machine learning systems trained on vast text corpora to perform natural language tasks such as generation [7]. Their probabilistic nature makes them unreliable in domains that demand determinism, including running planning where a violation of physiological constraints can create safety risks [1]. SafeRun addresses this by decoupling the planning process. An LLM handles the soft interpretation of a runner’s natural-language request, while a deterministic solver enforces hard constraints, ensuring that every output plan respects predefined safety boundaries [1]. The architecture preserves the flexibility of natural-language interaction without allowing the model to override critical limits. To measure performance, the team built a benchmark grounded in realistic physiological and safety constraints and tested it across five LLMs [1]. The SafeRun framework recorded a 100% safety score. By comparison, the PE approach averaged 79.1% and the CodeAct method averaged 97.6% on the same metric [1]. The benchmark has been made publicly available on Hugging Face, a platform that hosts models, datasets, and interactive demos for machine learning research [1][3]. Hugging Face’s infrastructure allows authors to link papers directly to associated artifacts such as models and datasets. When a repository’s README includes an arXiv identifier, the Hub extracts it and creates a dedicated paper page, which can also surface community-built demos [3][4]. Since November 2022, arXiv has integrated Hugging Face Spaces so that readers can launch interactive demos from a paper’s abstract page without writing code [4][5]. The SafeRun release arrives as the broader LLM landscape continues to diversify. Chinese firm DeepSeek, founded in July 2023, drew attention in early 2025 when its DeepSeek-R1 model matched the performance of leading Western systems while reportedly training its V3 model for US$6 million, a fraction of the cost of comparable models [6]. Alibaba Cloud’s Qwen family, distributed under open-source licenses such as Apache 2.0, has also expanded the pool of available LLMs [8]. Against this backdrop, SafeRun targets a specific failure mode—safety-constraint violation—that remains relevant regardless of which underlying LLM is used. A separate scoping review of quantum circuit generation systems, published on arXiv in early 2026, found that while all thirteen reviewed systems addressed syntactic validity, none reported end-to-end evaluation on quantum hardware, leaving a gap between generated artifacts and practical deployment [2]. SafeRun’s authors, by contrast, provide a concrete benchmark with quantitative safety scores, though the paper does not describe deployment in live coaching or wearable-device settings [1].