---
format: "aidr-story-markdown/v1"
id: "ece011d9c6135f9db76a03643814a5a81b89cb09aa874e8198e3998f24c111d5"
canonical_url: "https://aidr.today/ece011d9?lang=en"
title: "When LLM Decompilers Recompile More and Preserve Less"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-12T02:44:55.000Z"
category: "Research"
topics: ["llm","coding","benchmark","security"]
source_urls: ["https://arxiv.org/abs/2609.05370","https://lobste.rs/s/uzfa4k/when_llm_decompilers_recompile_more"]
summary: "LLM-based decompilers may produce code that recompiles and passes shipped tests but diverges from the original behavior on other inputs or removes disclosed vulnerabilities. To address this, Decompile-Diverge was proposed as a behavioral comparison oracle that synthesizes drivers and grows fuzzing corpora to detect these changes. Testing across eight systems showed that candidates passing all shipped tests still diverged from the original in up to 13% of cases for a single system."
---

# When LLM Decompilers Recompile More and Preserve Less

> [Open the canonical story](<https://aidr.today/ece011d9?lang=en>)

**Published:** 2026-10-12T02:44:55.000Z
**Category:** Research
**Topics:** llm, coding, benchmark, security

## Summary

LLM\-based decompilers may produce code that recompiles and passes shipped tests but diverges from the original behavior on other inputs or removes disclosed vulnerabilities\. To address this, Decompile\-Diverge was proposed as a behavioral comparison oracle that synthesizes drivers and grows fuzzing corpora to detect these changes\. Testing across eight systems showed that candidates passing all shipped tests still diverged from the original in up to 13% of cases for a single system\.

## Sources

- [Story source](<https://arxiv.org/abs/2609.05370>)
- [Discussion](<https://lobste.rs/s/uzfa4k/when_llm_decompilers_recompile_more>)