CodeOrigin

Domain-Invariant AI-Generated Code Detection via Style-Content Disentanglement and Hierarchical Contrastive Learning

Detecting AI-generated code that generalizes across unseen languages, domains, and model families.

Problem

Current AI-generated code detectors achieve near-perfect accuracy on in-distribution data but collapse under distribution shift (new languages, new domains, new generators). On AICD-Bench, even the best trained model (DeBERTa, 34.13 F1) scores below random guessing (45.73 F1) on the out-of-distribution test set.

Root cause: Models overfit to surface-level, domain-specific artifacts (e.g., competitive programming templates) instead of learning generalizable authorship signals.

Method

CodeOrigin disentangles how code is written (style) from what code does (content), enabling domain-invariant detection.

Source Code
    |
    v
[Multi-Granularity Encoder]
    |-- Token-level:  ModernBERT on source tokens
    |-- AST-level:    BiLSTM on AST node type sequences
    |-- Graph-level:  MLP on structural features
    |
    v  (Cross-Attention Fusion)
    |
[Style-Content Disentanglement]
    |-- z_style:   authorship fingerprint (naming, control flow, comments)
    |-- z_content: semantic meaning (algorithm, domain, language)
    |
    |-- MI Minimization (CLUB): ensure z_style independent of z_content
    |-- KL Regularization + Reconstruction
    |
    v
[Task Heads on z_style only]
    |-- T1: Binary classifier + Domain Adversarial Alignment (GRL)
    |-- T2: Prototypical Contrastive Head (learnable family prototypes)
    |-- T3: Focal Loss classifier (handles class imbalance)

Key Contributions

1. Style-Content Disentanglement -- First to explicitly separate authorship fingerprints from code semantics for detection, directly addressing OOD collapse
2. Multi-Granularity Code Representation -- Token + AST + structural features fused via cross-attention, capturing complementary signals
3. Hierarchical Prototypical Contrastive Learning -- Learnable family prototypes for attributing unseen generators to known model families
4. Unified Multi-Task Framework -- Single model handles all 3 AICD-Bench tasks with shared representations

Benchmark: AICD-Bench

• 77 generators across 11 families (DeepSeek, Qwen, Llama, GPT-4o, Gemini, ...)
• 9 programming languages (Python, Java, C++, C, Go, PHP, C#, JavaScript, Rust)
• OOD test splits: unseen languages, unseen domains, unseen generators

Results

Quick Start

Run on Kaggle

# Cell 1: Install dependencies
!pip install datasets transformers accelerate tree-sitter-languages

# Cell 2: Run
from exp00_codeorigin import main

# Task 1: Robust Binary Classification
main(task="T1")

# Task 2: Model Family Attribution
main(task="T2")

# Task 3: Fine-Grained Classification
main(task="T3")

Train on AICD-Bench (separate benchmark)

python experiment/exp00_codeorigin.py --benchmark aicd --task T1

Train on DroidCollection (separate benchmark)

python experiment/exp00_codeorigin.py --benchmark droid --task T1

• Each run uses exactly one benchmark (aicd or droid).
• Train/validation/test are all from the selected benchmark.

Custom Configuration

from exp00_codeorigin import main, Config

config = Config(
    task="T1",
    epochs=5,
    batch_size=32,
    max_train_samples=100_000,
    max_val_samples=20_000,
    max_test_samples=50_000,
    encoder_name="answerdotai/ModernBERT-base",
    lr_encoder=2e-5,
    lr_heads=1e-4,
    fp16=True,
)
main(task="T1", config=config)

Project Structure

ai_code_detection/
├── README.md
├── experiment/                 # Baseline: exp00_codeorigin.py (Kaggle-ready single-file)
├── Exp_DM/                     # Novel methods on AICD + Droid (exp01..30)
│   └── dm_tracker.md           # Results tracker for AICD + Droid
├── Exp_CodeDet/                # Novel methods on CoDET-M4 (exp11..38)
│   └── tracker.md              # Leaderboard for CoDET-M4
├── docs/
│   ├── CLAUDE.md               # Full repo context — start here for any AI assistant
│   ├── performance_tracker.md  # Legacy tracker (superseded by Exp_DM/dm_tracker.md)
│   ├── references/             # Source papers, dataset schemas, research notes
│   └── archive/                # Completed session artifacts
├── Slide/                      # Proposal deck (tex/pdf) + presentation script
└── Formatting_Instructions_For_NeurIPS_2026/   # NeurIPS 2026 LaTeX template

New to the repo? Read docs/CLAUDE.md first.

Requirements

• Python >= 3.8
• PyTorch >= 2.0
• transformers >= 4.36
• datasets
• scikit-learn
• tree-sitter-languages (optional, falls back to regex-based AST extraction)

Citation

@article{codeorigin2026,
  title={CodeOrigin: Domain-Invariant AI-Generated Code Detection via Style-Content Disentanglement and Hierarchical Contrastive Learning},
  author={},
  year={2026}
}

AICD-Bench:

@article{orel2026aicdbench,
  title={AICDBench: A Challenging Benchmark for AI-Generated Code Detection},
  author={Orel, Daniil and Azizov, Dilshod and Paul, Indraneil and Wang, Yuxia and Gurevych, Iryna and Nakov, Preslav},
  journal={arXiv preprint arXiv:2602.02079},
  year={2026}
}

License

MIT