Stage-Supervised Latent Reasoning for Single-Shot JavaScript Deobfuscation
Source
Rong Feng, Suman Saha
Author
Rong Feng, Suman Saha
Date
Key takeaways · AI-distilled
Intermediate outputs from a deterministic deobfuscation tool become training supervision for a Coconut-style latent reasoning modelA model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.Full definition →, instead of treating deobfuscation as one-step translation.
Among syntactically valid outputs, the model reached 80% semantic correctness on JsDeObsBench, which the authors read as better faithfulness than direct fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → or zero-shot baselines.
This is a work-in-progress paper with preliminary results; half of the model's outputs still fail syntactic validity.
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
Why it matters
Deobfuscation is a real bottleneck in malware and security analysis; this method roughly doubles syntactic validity over direct fine-tuning in a single-shot inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → setup.
articleFunction-Level Execution Feedback for Code Preference OptimizationIdris Nechnech, Sehwan Kim, Jimin Seo, Yeongoon Kim, Minhae Oh, Sangwoo Hong, Jungwoo Lee
articleLatent Undertow: How Ordinary Typos Break ProbesElad David, Max Fomin, Amit LeVi
articleCode-as-Auditor: Executable Compliance Reasoning via Regulation-to-CodeJisoo Kim, Taeyoon Kwack, Jinwoo Jang, Woo Kyung Kim, Honguk Woo