Reinforcement Learning for LLM-Based Decompilation

This is a very wild idea: Use reinforcement learning to enhance the current LLM-based decompilation.

Is there any existing work?

Yes — there is now some directly relevant work, but it is quite recent and still sparse, rather than a large established line of research.

Baselines and Precursors

• LLM4Decompile (EMNLP 2024) is a major baseline for LLM-based decompilation. Its training recipe centers on large-scale supervised fine-tuning, data augmentation, data cleaning, and two-stage training — not reinforcement learning. Highly relevant as the baseline that RL methods build upon.

RL-Based Approaches

• D-LiFT (2025) explicitly improves an LLM-based decompiler backend using code-quality-aware reinforcement learning. It aims to improve decompiled-code quality while preserving accuracy via an integrated scoring system. This is RL for post-processing / backend improvement rather than pure end-to-end decompilation from binary.

• SK2Decompile (2025) describes a two-phase binary decompilation pipeline that uses reinforcement learning in both phases — one reward for structure recovery following compiler-like syntactic/semantic rules, and another reward for identifier naming based on semantic similarity. This is very close to “using RL to improve LLM-based decompilers” end-to-end.

• RlDecompiler (ICPC 2026) — “Enhancing LLM-based Decompilation via Reinforcement Learning with a Multi-Faceted Reward Function.” An explicitly decompilation-focused RL paper at a top program comprehension venue. Very recent; I have not yet found an openly accessible full PDF.

Two Emerging Directions

These papers split into two directions:

1. RL for improving decompiler outputs after or around an existing decompiler pipeline (e.g., D-LiFT).
2. RL inside the decompilation model itself, training the LLM’s generation with RL objectives (e.g., SK2Decompile, RlDecompiler).

RlDecompiler

The anonymous repo has expired, but the model is still available:

• Original: ri-char/rldecompile-1.3b
• Static quants: mradermacher/rldecompile-1.3b-GGUF

This 1.3B model is basically useless. It cannot correctly decompile even the easiest assemblies.

D-LiFT

Possibly still under peer review, so they have not released their model weights.

They use GRPO to perform RL on the base models Qwen2.5-Coder and Llama3.2-3B.

They propose a reward function called D-SCORE, which performs both syntax and semantic checks to ensure the correctness of decompiled code. They also assess readability if both accuracy checks pass. For readability metrics, see b&w + R2I.

Their method is quite interesting, but I am not sure if symbolic-execution-based semantic checks would guarantee semantic correctness. I think it will lead to reward hacking if the design is poor.

SK2Decompile

They did not use RL to enhance semantic correctness, but only used it to enhance structural recovery.

Takeaway

Before 2025, the best-known LLM decompilation work (LLM4Decompile) was mostly supervised. By 2025–2026, papers are explicitly introducing RL-based training objectives for decompilation quality, structure recovery, readability, and backend refinement. The idea is valid and already emerging, but still early-stage rather than saturated.