Sign InOpen Brain
arXivPaperNeeds Review

When LLM Decompilers Recompile More and Preserve Less

LLM decompilers can produce compilable C that passes shipped tests while changing behavior or hiding crashes. Validate reconstructed code against the original with shared fuzzed inputs.

arXiv · Sep 4, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Decompile-Diverge found that outputs passing every shipped test still differed from original binaries on **4.9% overall**, reaching **13% for one system**. Its oracle generates a driver, grows inputs from the reference, and compares both implementations on the same corpus.

Practical Implication

For security work, treat recompilation and fixture tests as entry checks, not semantic proof. Differential fuzzing should gate agent-generated decompilation before using it for vulnerability analysis or downstream automation.

Agent-Ready Context
Decompile-Diverge found that outputs passing every shipped test still differed from original binaries on **4.9% overall**, reaching **13% for one system**. Its oracle generates a driver, grows inputs from the reference, and compares both implementations on the same corpus.

For security work, treat recompilation and fixture tests as entry checks, not semantic proof. Differential fuzzing should gate agent-generated decompilation before using it for vulnerability analysis or downstream automation.

The strongest refinement LLM raised Ghidra's build rate from **75% to 90%** while matched behavior fell from **74% to 62%**. Up to one tenth of CVE-grounded functions lost the disclosed crash, but results cover the evaluated corpora and configurations.
Connected Context · Feed7 Judgment

This adds semantic equivalence as an independent acceptance gate for decompilation agents: recompilation and shipped tests can improve while fidelity to the binary worsens. It reinforces prior evidence that green functional checks miss consequential constraints, and supplies a reference-based oracle—shared fuzzed inputs and differential execution—suited to behavior drift and lost security-relevant crashes.

SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering AgentsBoth show that functional test success is an incomplete acceptance criterion; SWE-Gate finds review-constraint violations, while this Signal finds divergence from original binary behavior.Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code ModelsThe placebo-controlled repair result makes refinement gains suspect without causal validation; here, refinement raises buildability while reducing behavioral preservation.Clean Engineering, Unstable Measurement: A Preregistered Reliability Failure of Black-Box LLM Observers on Shared EndpointsThe unstable-judge study supports validating an evaluation instrument before gating releases; this Signal offers a deterministic reference comparison that avoids relying solely on black-box judgment.
Context Map
benchmarkcodingsecurity#agent-evals#benchmark-integrity#agent-reliability
Uncertainty
The strongest refinement LLM raised Ghidra's build rate from **75% to 90%** while matched behavior fell from **74% to 62%**. Up to one tenth of CVE-grounded functions lost the disclosed crash, but results cover the evaluated corpora and configurations.