Many Ethereum smart contracts are deployed without verified source code, making auditing, security analysis, and research difficult in practice. Decompilation aims to bridge this gap by recovering readable, high-level code from EVM bytecode, but doing so reliably remains challenging.

This talk shares our experience using large language models for smart contract decompilation, based on building and evaluating multiple approaches. I will discuss different ways LLMs can be applied to decompilation and the practical trade-offs involved. A central focus of the talk is how to evaluate decompilers in a meaningful way. I will explain why traditional metrics are insufficient, what properties are important in practice.

The goal of this session is to provide a clear, experience-based view of what LLM-based decompilation can achieve today, where it still falls short, and how careful benchmarking helps guide both research and real-world tooling.