ExplainFuzz: Explainable and Constraint-Conditioned Test Generation with Probabilistic Circuits

A research team from UCLA and UC Berkeley has introduced ExplainFuzz, a novel AI-powered framework for automated software testing that addresses critical limitations of current approaches. Unlike traditional grammar-based fuzzers or Large Language Models (LLMs) that often produce ill-formed inputs and lack transparency, ExplainFuzz leverages Probabilistic Circuits (PCs) to learn structured distributions over test inputs interpretably. Starting from a Context-Free Grammar (CFG), the system compiles a grammar-aware PC and trains it on existing inputs, enabling it to generate new, realistic test cases through controlled sampling.

ExplainFuzz's key innovation lies in its native conditioning capability, allowing users to incorporate specific test constraints—like requiring a SQL query to contain 'GROUP BY'—and generate diverse inputs that satisfy both grammar rules and user requirements. This approach moves beyond mutational fuzzing, which is limited to exploring local neighborhoods of seed inputs, to instead learn the broader input distribution. The results are striking: compared to grammar-aware mutational fuzzing, ExplainFuzz increased bug-triggering rates from 35% to 63% for SQL testing and from 10% to 100% for XML testing, while also achieving significant perplexity reduction over probabilistic CFGs and LLMs.

This research demonstrates that PCs can serve as a powerful foundation for grammar-aware, controllable test generation that captures context-sensitive probabilistic dependencies. The framework's explainability component allows developers to understand why specific test inputs were generated, bridging the gap between automated testing and human debugging workflows. By producing more coherent and realistic inputs that better reflect actual data distributions, ExplainFuzz represents a significant advancement toward more effective and trustworthy automated software testing pipelines.