CMU Proposes ARISE: Natural Language Agent Subroutines Outperform Code Tools, Self-Reported +6.1% Accuracy
A CMU preprint introduces ARISE, where LLMs execute natural language subroutines instead of using injected skills or Python tools, reporting a 6.1 percentage point gain over top baselines in five reasoning tasks.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Researchers at Carnegie Mellon University updated their preprint on October 14, 2026, introducing ARISE (Agent Routine Induction from Successful Experience). The method identifies recurring procedures in successful agent trajectories and synthesizes them into a reusable library of 'agent subroutines.' Unlike prior approaches that inject knowledge as skills into the context or rewrite it as Python tools, ARISE has large language models directly interpret and execute these natural language specifications.
In evaluations across five tasks spanning narrative deduction, rule application, and constraint planning, the authors report that ARISE outperforms chain-of-thought prompting in all ten model-task settings. Compared to state-of-the-art experience induction methods, it scores 6.1 percentage points higher; compared to code-skill induction, it scores 11.5 points higher. Controlled experiments show that changing only the execution interface—having an LLM execute the subroutine via a separate call rather than placing the same text in context as a skill—improves average accuracy by up to 15.6 percentage points. Rewriting the same procedures as Python tools resulted in lower accuracy in nine of ten settings.
These results are self-reported by the authors, with no details provided on hardware configuration or statistical significance tests, and are limited to specific logical reasoning tasks. There are currently no third-party independent reproductions, and generalizability to complex open environments remains to be observed.