Autonomous Test Case Generation From API Specifications Using AI: A Hybrid LLM and Reinforcement Learning Framework

Abstract

The exponential growth of REST API surfaces in microservice architectures has outpaced the capacity of manually authored test suites to provide meaningful specification-level coverage. Existing AI-driven test generation tools address this gap in isolation: large language model (LLM)-based approaches generate syntactically valid inputs but treat each API endpoint independently, discarding the inter-endpoint dependency semantics encoded in OpenAPI specifications; reinforcement learning (RL)-based approaches optimize test execution scheduling but presuppose that test cases already exist and leave the oracle construction problem unresolved. This article presents a hybrid autonomous test generation framework with four integrated components: a Spec-Semantic Graph (SSG) that encodes endpoint dependency relationships as a weighted directed graph; a Contextual Boundary Inference Engine (CBIE) that uses few-shot LLM prompting to extract semantic boundary conditions from natural-language description fields in OpenAPI specifications; a Coverage-Aware RL Test Scheduler (CARTS) that prioritizes test execution using a multi-dimensional coverage state vector; and Differential Oracle Synthesis (DOS) that constructs pass/fail verdicts without manually authored expected-output specifications. When evaluated on three open-source REST APIs with varying levels of specification completeness, the framework achieves higher endpoint coverage than any individual LLM-only or search-based baseline, reduces oracle false-positive rates, and incurs a one-time CI pipeline overhead that amortizes across specification versions rather than accumulating on every run. Together, the four components realize a specification-native quality assurance pipeline that derives test intelligence directly from the OpenAPI document without manual annotation.