Chiasmus: A Neurosymbolic System Giving LLMs Formal Reasoning for Code Analysis
Key Takeaways
- ▸LLM coding assistants lack the ability to perform exhaustive structural analysis of code, relying instead on ad hoc grep and pattern matching that fails for transitive questions like reachability and impact analysis
- ▸Chiasmus combines tree-sitter parsing, Prolog logic programming, and Z3 constraint solving to give LLMs access to formal reasoning engines that can answer structural code questions with certainty
- ▸The neurosymbolic approach divides labor between LLMs (perception, natural language understanding) and symbolic solvers (logical inference, graph traversal, constraint satisfaction), enabling more reliable code analysis while reducing token consumption
Summary
Chiasmus is a new Model Context Protocol (MCP) server designed to address a fundamental weakness in LLM-based coding assistants: their inability to perform exhaustive structural analysis of code. While LLMs excel at writing code, they struggle with complex reasoning tasks that require understanding code structure, such as determining whether user input can reach a specific SQL query through any call chain or identifying all dead code in a module. Traditional approaches relying on grep and pattern matching quickly fail for transitive questions that demand graph traversal and logical inference.
The system bridges the gap between neural and symbolic AI by combining tree-sitter parsing, Prolog logic programming, and Z3 constraint solving. Source files are parsed into abstract syntax trees (ASTs) that capture structural relationships like method definitions and call chains, then converted into Prolog facts. This provides LLMs with a formal knowledge base that can answer code structure questions with certainty while consuming a fraction of the tokens required by grepping approaches. Rather than making multiple tool calls to trace through source files, the LLM can now query a symbolic reasoning engine that can exhaustively traverse code graphs and perform constraint satisfaction.
Chiasmus is grounded in the neurosymbolic AI paradigm, which combines the perception and language understanding capabilities of neural networks with the reasoning and verification abilities of symbolic knowledge systems. The project currently supports Python, Go, TypeScript, JavaScript, and Clojure, with extensibility for other languages through custom adapters.
- The system supports multiple languages including Python, Go, TypeScript, JavaScript, and Clojure, with plans for broader extensibility
Editorial Opinion
Chiasmus represents an important step toward more reliable AI-assisted code analysis by recognizing the complementary strengths of neural and symbolic systems. While LLMs have revolutionized code generation, this work correctly identifies that pattern-matching approaches are fundamentally inadequate for structural reasoning tasks that require logical certainty. By embedding formal reasoning engines within an LLM's tool ecosystem, Chiasmus offers a practical path to more trustworthy AI coding assistants—particularly valuable for security-critical questions like data flow analysis where probabilistic answers are insufficient.
