MAVEN — symbolic reasoning scaffold for agentic tool calling
New arXiv paper tackles generalization in LLM-based agent tool use across domains. LLMs excel on individual benchmarks but struggle to compose reasoning, preserve state, and coordinate tools reliably.
MAVEN (Modular Agentic Verification and Execution Network) is a lightweight symbolic scaffold that decomposes tasks structurally, orchestrates tool calls adaptively, and verifies intermediate results.
Tested on BFCL v3, TauBench, Tau2Bench, AceBench, plus a new MAVEN-Bench stress-test for multi-step math and physics reasoning.