SafeRun introduces a framework that combines Large Language Models with deterministic solvers to enable reliable planning in safety-critical domains like running training. The hybrid architecture separates LLM's natural language flexibility from hard constraint enforcement, achieving 100% safety compliance while maintaining instruction-following capabilities.
SafeRun addresses a fundamental challenge in deploying LLMs to domains where safety violations carry real consequences. Traditional LLMs generate outputs probabilistically, making them unsuitable for applications requiring deterministic behavior and strict rule adherence. By decoupling the interpretive layer from the constraint enforcement layer, SafeRun enables developers to leverage LLM capabilities without sacrificing safety guarantees.
The framework's innovation lies in its architectural separation: LLMs handle the nuanced, context-aware interpretation of user intent and natural language instructions, while a deterministic solver enforces hard constraints derived from physiological safety requirements and running best practices. This hybrid approach solves a persistent problem in AI deployment where pure neural approaches fail at strict compliance while rule-based systems lack flexibility.
The benchmark development is equally significant. Running planning presents concrete safety constraints—heart rate zones, recovery windows, distance progressions—making it an ideal validation domain. Results demonstrating 100% safety compliance across five LLMs establish that this architectural pattern works reliably in practice. The 79.1% average safety score from existing approaches illustrates how common failures are in unstructured LLM planning.
This work has implications beyond fitness applications. Any domain requiring both natural language understanding and deterministic safety compliance—medical advice systems, autonomous vehicle planning, industrial automation—could benefit from similar decoupled architectures. The framework suggests a scalable pattern for integrating probabilistic AI with formal verification, potentially addressing adoption barriers in regulated industries where LLM unreliability currently limits deployment.
- →SafeRun achieves 100% safety compliance in LLM-based planning by separating natural language interpretation from hard constraint enforcement
- →A new benchmark for running planning with physiological constraints provides reproducible validation across five different LLMs
- →The decoupled architecture pattern could extend to other safety-critical domains requiring both flexibility and strict rule adherence
- →Existing approaches averaged only 79.1% safety compliance, showing significant prior limitations in deterministic LLM planning
- →The publicly available benchmark enables further research into hybrid AI systems combining neural and symbolic reasoning