SCOPE Compiles Natural-Language Policies Into Neurosymbolic Predicate Programs
Arnon Mazza and Elad Levi submitted the SCOPE paper to arXiv on October 22, 2026. SCOPE compiles a natural-language policy into an executable program of neural predicates, combining small models, deterministic extractors, and code that composes their outputs. The program and its training data are optimized jointly using adversarial counterexamples targeting deviations from the policy. In the authors’ evaluation on a benchmark of real-world policy tasks, SCOPE achieved the highest average accuracy across tasks and models compared with prompt optimization, task decomposition, and workflow-generation methods. They also report that programs optimized with the weakest executor transferred to other models without further optimization, produced more stable predictions than monolithic models, and generated verdicts interpretable through their predicate outputs.
The paper targets compound, application-specific policy enforcement for LLM agents. Its reported use of small models, stable predictions, and predicate-level explanations suggests a potentially more efficient and interpretable alternative to high-cost, less stable LLM-as-a-judge approaches, though these results come from the authors’ evaluations.
Computer Science > Artificial Intelligence
arXiv:2610.31062 (cs)
[Submitted on 22 Oct 2026]
Title:SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs
Authors:Arnon Mazza, Elad Levi
View a PDF of the paper titled SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs, by Arnon Mazza and 1 other authors