Towards Automated Formal Verification of zkEVMs Using LLM-Guided Constraint Synthesis

2026-06-27 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Blockchain & Distributed Ledger Technology · Depth: Expert, extended

Summary

VeriSynth is a novel LLM-guided framework designed for automated formal verification of Zero-Knowledge Ethereum Virtual Machines (zkEVMs) by synthesizing executable Python/Z3 verification models directly from Rust source code. It addresses the bottleneck of manual specification in zkEVM verification, where subtle bugs like incorrect gas accounting can defeat cryptographic guarantees. VeriSynth employs a hybrid paradigm, using an LLM as a formalization frontend to translate code into symbolic constraints, with an SMT solver acting as the correctness arbiter. The framework integrates semantic decomposition, retrieval-grounded prompting, and verification-guided auto-repair. Evaluated on a 95-sample zkEVM verification benchmark, VeriSynth achieved a 91.6% bug detection rate, significantly outperforming LLM-only (46.3%) and conversational LLM (61.1%) baselines. It also detected 32 more bugs than Scroll's handwritten mutation-testing suite (57.9%), demonstrating its practical value. Each sample consumed approximately 100K tokens, totaling about 9.5M tokens for the benchmark.

Key takeaway

For security engineers and researchers developing or auditing zkEVM implementations, VeriSynth offers a robust approach to detect subtle, implementation-level semantic bugs that manual testing often misses. You should consider integrating LLM-guided constraint synthesis into your verification pipeline to complement existing mutation tests. This can significantly improve bug detection coverage, especially for complex opcode behaviors, by leveraging automated model generation and solver-backed validation to catch inconsistencies before they impact cryptographic guarantees.

Key insights

VeriSynth automates zkEVM formal verification by synthesizing SMT models from Rust code, guided by LLMs and validated by solvers.

Principles

• LLMs formalize code, SMT solvers validate correctness.
• Decompose complex state transitions into modular units.
• Ground LLM generation with previously verified examples.

Method

VeriSynth decomposes Rust code into semantic units, retrieves verified examples for grounding, synthesizes Python/Z3 models using an LLM with opcode-aware prompts, and iteratively auto-repairs models based on compiler and solver feedback.

In practice

• Represent EVM words as fixed-width bit-vectors in SMT models.
• Translate source-level branches into guarded constraints or Z3 If expressions.
• Apply frame conditions to ensure unchanged state components are constrained.

Topics

• zkEVM Verification
• Formal Verification
• LLM-Guided Synthesis
• SMT Solvers
• Rust Programming
• Opcode Semantics

Best for: AI Scientist, Research Scientist, Software Engineer

Related on AIssential

• Snyk VulnBench JS 1.0: Can LLMs Find the Same Bugs Twice? — LLM security review repeatability varies significantly, with reference-matched findings stable but additional reports highly inconsistent.
• VERITAS: Verifier-Guided Proof Search for Zero-Shot Formal Theorem Proving — VERITAS improves LLM theorem proving by integrating all verifier signals into a two-phase, feedback-driven search.
• Directed Symbolic Execution for Vulnerability Discovery: An LLM-Guided Approach in KLEE — LLMs can guide symbolic execution to overcome path explosion, improving vulnerability discovery by prioritizing vulnerable code and escaping cycles.
• Harnessing Code Agents for Automatic Software Verification — A general LLM code agent with a verification harness can fully automate formal software proof generation.
• An Empirical Study of LLM-Generated Specifications for VeriFast — LLMs preserve functional behavior in VeriFast specifications but achieve modest verification success due to domain-specific knowledge gaps.
• Open Problems Solved by LLMs? A Survey of Verifiable Mathematical Discovery — LLMs, integrated with search and formal verification, are increasingly solving open mathematical problems.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.