MAVEN scaffold boosts open-source AI tool-calling accuracy by 23%

MAVEN addresses a core challenge in agentic AI: generalization across tool-calling environments. Existing large language models often excel on individual benchmarks but fail when asked to compose reasoning strategies, preserve intermediate states, or coordinate multiple tools. MAVEN (Modular Agentic Verification and Execution Network) introduces a lightweight symbolic scaffold that decomposes tasks into structured steps, orchestrates tools adaptively, and performs intermediate verification. This approach requires no additional training, making it a drop-in upgrade for open-weight models like GPT-OSS-120b.

The research team evaluated MAVEN on established benchmarks (BFCL v3, TauBench, Tau2Bench, AceBench) and created MAVEN-Bench, a stress-test benchmark for multi-step mathematical and physical reasoning with adversarial task composition. On MAVEN-Bench, MAVEN lifted its base model from 48% to 71% accuracy—a 23-point gain. Despite using an open-weight backbone, MAVEN matched or exceeded proprietary frontier baselines while operating at an estimated cost ratio of roughly 1/10. These results suggest that lightweight verification scaffolds, rather than brute-force scaling, can enable reliable, cost-effective agentic reasoning in real-world deployments.