PaveDSL: A Neuro-Symbolic Approach to Plan Specification and Verification for Multi-Agent Code Generation
Abstract
AbstractRecent multi-agent code generation frameworks often verify natural-language plans by simulating their behavior on test cases before code generation. However, such verification remains unreliable because model hallucination may cause flawed plans to be incorrectly accepted. Inspired by neuro-symbolic reasoning, we propose PaveDSL, which transforms natural-language plans into structured designs expressed in a domain-specific language. Based on this structured representation, PaveDSL combines symbolic checks with neural inference to examine plans before code generation. Specifically, PaveDSL applies two complementary verifiers in parallel. The deterministic symbolic verifier checks structural constraints without invoking an LLM, while the agent verifier performs graph-guided semantic reasoning over the structured design. If verification fails, a repair agent localizes errors and revises the design before code generation. Experiments on multiple benchmarks with different LLMs show that PaveDSL consistently improves the performance of baseline code generation methods.