Skip to main content
Search

Peipei Liu

1 related articles

Subscribe to updates
  1. arXiv

    AutoDecompiler turns binary decompilation into feedback-driven multi-turn refinement with reinforcement learning, consistently beating single-turn baselines at matched scale and input while improving behavioral re-executability

    The work presents AutoDecompiler, a decompilation-specialized LLM trained with reinforcement learning that reframes binary decompilation from single-turn generation into an iterative refinement process driven by compilation, execution, and input/output testing feedback, with decompilation-specific rewards covering code validity, recompilability, execution consistency, and semantic fidelity, plus stage-aware diagnostic feedback from compiler errors, execution failures, and failed test cases and progress-aware trajectory rewarding with turn-aware advantage reweighting; across input settings, model scales, and benchmarks it consistently outperforms its single-turn counterparts at the same model size and input setting, with clear gains in behavioral re-executability.

You’re all caught up