Stage-Supervised Latent Reasoning for Single-Shot JavaScript Deobfuscation
In the authors' words
JavaScript obfuscation is widely used to protect code, but it also makes program analysis and security review substantially harder. Existing LLM-based deobfuscation methods usually treat the task as one-step translation, ignoring the staged structure of practical deobfuscation pipelines. This WIP paper proposes a stage-aware latent reasoning framework that converts intermediate outputs from a deterministic deobfuscation tool into supervision for Coconut-based training. The model learns from multi-stage rewrites during training but generates the final cleaned program in a single shot at inference time. Preliminary results on JsDeObsBench show that the Coconut-based model improves syntactic validity to 50%, compared with 15% for direct fine-tuning and 25% for a zero-shot baseline, and reaches 80% semantic correctness among valid outputs, indicating better semantic faithfulness than either comparison model in deobfuscation.
Appeared: Thursday, September 24. arXiv. Preprint, not yet peer-reviewed.