When LLM Decompilers Recompile More and Preserve Less: Decompile-Diverge Finds Behavioral Drifts

Forum topic · 小凯 · 2026-09-09

Summary

Researchers Chang Liu, Edward Raff, and Kristopher Micinski (arXiv:2609.05370) show that LLM-based decompilers can look better than traditional tools while preserving less. Current evaluation relies on recompilability and re-executability: whether decompiled output builds and passes shipped input/output tests. The paper demonstrates these metrics can reward wrong paths — a decompiled function may pass every shipped test yet diverge on other legitimate inputs, and a disclosed vulnerability may silently vanish from the recompiled code. The authors propose Decompile-Diverge, a behavioral comparison oracle that synthesizes per-function drivers, grows a fuzzing corpus from reference implementations, and reruns decompiled code on identical inputs to detect behavioral changes without fixed or hand-written tests. Across eight systems in nine configurations, candidates passing all shipped tests still diverged on their corpus in 4.9% of cases overall, up to 13% for a single system. On 300 real GitHub library functions and 287 CVE-related functions, recompilability and behavioral fidelity can separate: the strongest refined LLM raised Ghidra's build rate from 75% to 90% while match rate fell from 74% to 62%, and up to one in ten disclosed vulnerabilities showed crash-absence in its output. Source-level analysis attributes divergence to introduced fields, types, callees, and guards replacing visible unknowns.

Paper Overview

Field: ML
Authors: Chang Liu, Edward Raff, Kristopher Micinski
Published: 2026-09-04
arXiv: 2609.05370

Background

Decompilation recovers high-level source from compiled machine code and underpins security tasks like vulnerability detection and malware analysis. Traditional decompilers (Ghidra, Hex-Rays) expose unresolvable content as visible placeholders and often emit pseudocode that will not compile or execute. LLM-based decompilers produce clean, idiomatic C, and are now judged almost entirely by recompilability and re-executability: whether output builds and passes shipped input/output tests.

Problem

These metrics can reward the wrong path:

A function may recompile and pass every shipped test yet diverge on other legitimate inputs.
A disclosed vulnerability may disappear from the recompiled code with no visible trace of the crash.

No existing test suite catches these failures.

Decompile-Diverge

The authors propose Decompile-Diverge, a behavioral comparison oracle that requires no fixed or hand-written tests:

1. Synthesize a driver for each function. 2. Grow a fuzzing corpus from the reference implementation. 3. Rerun the decompiled code on the same inputs to detect behavioral divergence.

Findings

Across eight systems in nine configurations, candidates that passed all shipped tests still diverged from the original on the authors' corpus in 4.9% of cases overall, up to 13% on a single system.
On 300 real GitHub library functions and 287 CVE-related functions, recompilability and behavioral fidelity can diverge: the strongest refined LLM raised Ghidra's build rate from 75% to 90%, while match rate dropped from 74% to 62%.
Up to one in ten disclosed vulnerabilities exhibited crash-absence in the LLM output.
Source-level analysis attributes divergence to introduced fields, types, callees, and guards that replace the visible unknowns left by traditional tools.

Takeaway

Recompilability alone is an insufficient — and potentially misleading — metric for LLM decompilers. Behavioral comparison against reference implementations, as in Decompile-Diverge, reveals silent semantic loss and vanished vulnerabilities that shipped tests miss.

---

*Auto-collected on 2026-09-09.*