Binary Decompilation LLM with Feedback-Driven Multi-Turn Refinement

2025-01-14 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

AutoDecompiler is a reinforcement-learning-trained large language model designed for feedback-driven multi-turn binary decompilation. Unlike traditional single-turn LLM methods that generate code once, AutoDecompiler iteratively refines decompiled code using compilation, execution, and input/output testing feedback. Its training incorporates decompilation-specific rewards for code validity, recompilability, execution consistency, and semantic fidelity. It also utilizes stage-aware diagnostic feedback, progress-aware trajectory rewarding, and turn-aware advantage reweighting to guide beneficial revisions. Evaluated across HumanEval and ExeBench benchmarks, AutoDecompiler consistently outperforms single-turn LLM counterparts of the same size and input setting, demonstrating significant improvements in behavioral re-executability. For instance, AutoDecompiler-E2E 1.3B improved the best prior Re-exe result by 5.19%, and the 30B variant improved Re-exe over DeepSeek-V3 by 16.16%. The model family, including 1.3B, 6.7B, and 30B parameter scales, was trained on a comparatively smaller dataset of 0.31 million samples.

Key takeaway

For AI Engineers or ML Scientists developing code generation models for security tasks, you should move beyond single-turn generation and integrate iterative, feedback-driven refinement. This approach significantly enhances functional correctness and re-executability, even with smaller datasets. Consider implementing multi-dimensional rewards and stage-aware diagnostic feedback to improve model reliability and reduce hallucinations in critical applications like vulnerability discovery and malware inspection.

Key insights

Binary decompilation improves significantly by treating it as an iterative, feedback-driven refinement process using reinforcement learning.

Principles

• Decompilation quality requires multi-dimensional rewards.
• Iterative refinement needs explicit progress signals.
• Heterogeneous feedback must be actionable guidance.

Method

AutoDecompiler fine-tunes a base LLM with SFT, then optimizes it via RL for multi-turn refinement. It validates generated code, computes scalar rewards, and constructs natural-language diagnostic feedback for subsequent turns.

In practice

• Implement feedback loops for LLM-generated code.
• Design multi-dimensional rewards for complex tasks.
• Utilize RL to learn iterative code refinement.

Topics

• Binary Decompilation
• Reinforcement Learning
• Large Language Models
• Code Refinement
• Software Security
• Functional Correctness

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Vibe-Coding: Feedback-Based Automated Verification with no Human Code Inspection, a Feasibility Study — Precise, fine-grained feedback is critical for reliable LLM-generated code verification in runtime-adaptive systems.
• How Generation Architecture Shapes Code Complexity in Multi-Agent LLM Systems: A Paired Study on HumanEval — Multi-agent LLM code generation architectures significantly impact code complexity without necessarily improving functional correctness.
• Imperfect Visual Verification for Code Edition : A Case Study on TikZ — Imperfect visual verifiers can significantly improve LLM-based code customization through iterative feedback, especially for weaker models.
• FBOS-RL: Feedback-Driven Bi-Objective Synergistic Reinforcement Learning — Feedback-driven bi-objective reinforcement learning significantly boosts LLM training efficiency and performance by synergizing exploitation and exploration.
• From Risk Classification to Action Plan Remediation: A Guardrail Feedback Driven Framework for LLM Agents — TRIAD uses iterative, verbal guardrail feedback to guide LLM agents, enabling plan revision and preserving benign task components.
• Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery — Aggressive 2-bit quantization in LRMs causes specific generation pathologies, not just accuracy drops, which can be mitigated.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.