Research map / Large language models / Agents, code and math
SK2Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin
Hanzhuo Tan, Weihao Li, Xiaolong Tian, Siyi Wang, Jiaming Liu, Jing Li, Yuqun Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| split decompilation into structure recovery and identifier naming via IR (ablation) | SK2Decompile SFT · HumanEval | re-executability rate (%): 54.86 -> 63.75 | Improved |
| split decompilation into structure recovery and identifier naming via IR (ablation) | SK2Decompile SFT · MBPP | re-executability rate (%): 47.51 -> 52.83 | Improved |
| add RL with compiler and semantic rewards to the structure recovery model (ablation) | pseudo-ir · HumanEval | re-executability rate (%): 62.56 -> 68.84 | Improved |
| add RL with compiler and semantic rewards to the structure recovery model (ablation) | pseudo-ir · MBPP | re-executability rate (%): 47.25 -> 57.06 | Improved |
| add RL to both phases of the two-phase model (ablation) | pseudo-ir-src · HumanEval | re-executability rate (%): 63.75 -> 69.00 | Improved |
| add RL to both phases of the two-phase model (ablation) | pseudo-ir-src · MBPP | re-executability rate (%): 52.83 -> 59.63 | Improved |
| add naming phase after RL structure recovery (ablation) | pseudo-ir-rl · HumanEval | re-executability rate (%): 68.84 -> 69.00 | No clear effect |
| add naming phase after RL structure recovery (ablation) | pseudo-ir-rl · MBPP | re-executability rate (%): 57.06 -> 59.63 | Improved |
| add independent structure recovery phase only (IR), SFT (ablation) | pseudo-src · MBPP | re-executability rate (%): 47.51 -> 47.25 | No clear effect |
| add IR-based two-phase decompilation | SK2Decompile vs GPT-5-mini · HumanEval | re-executability rate (%): 56.75 -> 69.00 | Improved |
| add IR-based two-phase decompilation | SK2Decompile vs Idioms · GitHub2025 | R2I: 57.97 -> 74.99 | Improved |
| add IR-based two-phase decompilation | SK2Decompile vs LLM4Decompile · MBPP | re-executability rate (%): 43.05 -> 59.63 | Improved |
Closest papers by meaning
- Decompile-Bench: Million-Scale Binary-Source Function Pairs for Real-World Binary Decompilation (NeurIPS 2025 · similarity 0.87)
- From I/O to Code with Discovery Agent (NeurIPS 2026 · similarity 0.87)
- ASAP: Assembly-Source Aligned Pseudocode Refinement For Binary Decompilation (NeurIPS 2026 · similarity 0.87)
- CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules (ICLR 2024 · similarity 0.87)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.86)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
