English | 简体中文

LAPD

This repository contains the code and data for the paper "Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy".

We provide LAPD and RAI, together with a set of zero-shot baselines (fast-detectgpt, lastde++, binoculars, dna-detectllm, ...) in a unified framework, and benchmark data covering M4, DetectRL, RAID, RealDet, XSum, WritingPrompts, and CReD.

Paper: arXiv:2604.16923 | Code: creator-xi/LAPD

• 4/18/2026: The LAPD preprint is available on arXiv.
• 6/27/2026: This repository provides the LAPD implementation, baseline runners, benchmark data, robustness evaluation scripts, and ablation utilities.

Brief Introduction

Detecting AI-generated text is difficult because modern LLMs produce fluent, coherent, and increasingly human-like passages. Traditional likelihood-based detectors are often confounded by intrinsic content complexity: a simple human sentence can receive higher likelihood than a complex AI-generated paragraph.

LAPD starts from a different observation: modern user-facing LLMs are aligned through supervised fine-tuning and preference tuning. This alignment process leaves a measurable distributional imprint. By comparing the log-likelihood assigned by an aligned model and its corresponding base model, we can capture how much extra probability a text receives from alignment-induced preferences rather than from general language modeling alone.

The paper derives this signal as Alignment Imprint and proposes Log-likelihood Alignment Preference Discrepancy (LAPD), a standardized information-weighted statistic that improves robustness in high-entropy regions.

Method	Main Benchmarks AUROC	Cross-Domain AUROC	TPR @ 0.5% FPR	Time / Text
Fast-DetectGPT	82.26	95.36	51.79	0.5578s
DNA-DetectLLM	86.12	98.36	59.56	0.7696s
RAI	85.60	95.69	69.92	0.3506s
LAPD	92.37	99.49	92.27	0.5792s

The main benchmark results are averaged over M4, DetectRL Multi-LLM, DetectRL Multi-Domain, RAID, and RealDet. Cross-domain results are averaged over XSum, WritingPrompts, and Arxiv with GPT-4 Turbo, Gemini-2.0 Flash, and Claude-3.7 Sonnet generations.

Project Overview

LAPD is a zero-shot detector. It does not train a classifier on labeled AI/human examples. Given an input text, it:

1. Computes token-level likelihoods under a base model.
2. Computes token-level likelihoods under the corresponding aligned or instruction-tuned model.
3. Forms the Raw Alignment Imprint (RAI) from the base-aligned likelihood discrepancy.
4. Applies information weighting and perturbation-based standardization to obtain the LAPD score.

The implementation supports:

• LAPD and RAI evaluation.
• Standard zero-shot baselines including likelihood, entropy, LogRank, DetectGPT, Fast-DetectGPT, Lastde++, Binoculars, DNA-DetectLLM, RoBERTa baselines, and IRM.
• Main benchmark evaluation on M4, DetectRL, RAID, RealDet, XSum, WritingPrompts, Arxiv, text-attack data, and CReD.
• Ablations for LAPD components, text length, low-FPR behavior, time efficiency, and model-pair sensitivity.

Environment

Create a Python 3.10 environment first and then install the requirements.

conda create -n lapd python=3.10
conda activate lapd
pip install -r requirements.txt

After installation, verify the environment:

python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
python -c "import method.lapd; print('lapd import ok')"

Models are loaded through method/model.py. The default cache directory is ./cache. Local model aliases are also defined in method/model.py; for example, llama2-7b maps to a local Llama-2-7B checkpoint under the cache directory.

Recommended default models:

• Sampling model: llama2-7b
• Base model: llama2-7b
• Aligned model: llama2-7b-instruct

Running Guide

Run a small LAPD evaluation from the repository root:

python -m method.lapd \
  --data_source main \
  --dataset xsum \
  --source_model gpt4o \
  --sampling_model llama2-7b \
  --base_model llama2-7b \
  --instruct_model llama2-7b-instruct \
  --max_samples 10 \
  --device cuda \
  --cache_dir ./cache

The results will be saved under:

./formal_exp/lapd/

Run the full LAPD batch script:

./method/scripts/run_lapd.sh

Run baseline methods:

./method/scripts/run_baselines

Available baseline methods include:

likelihood, logrank, entropy, fast, fast_origin, detectgpt,
lastde, dna_detectllm, binoculars, binoculars_zscore,
roberta_base, roberta_large, rai, rai_double_gpu

Useful experiment scripts:

# LAPD main evaluation
bash method/scripts/run_lapd.sh

# Baseline evaluation
bash method/scripts/run_baselines.sh

# Component ablation
bash method/scripts/component_abla.sh

# Text-length ablation
bash method/scripts/ablation_text_length.sh

# Time-efficiency evaluation
bash method/scripts/time_efficiency.sh

Recommended Directory Structure

LAPD/
├── cache/                     # model checkpoints
│   ├── llama-2-7b/
│   ├── llama-2-7b-instruct/
│   ├── falcon-7b/
│   └── falcon-7b-instruct/
├── data/
│   ├── Collected data/        # XSum, WritingPrompts, Arxiv data
│   ├── M4/                    # M4 benchmark
│   ├── DetectRL/              # DetectRL benchmark
│   ├── RAID/                  # RAID benchmark
│   ├── RealDet/               # RealDet benchmark
│   ├── Text_attack/           # Random and realistic attack data
│   └── CReD/                  # Chinese benchmark data
├── method/
│   ├── lapd.py                # Main LAPD entry point
│   ├── lapd_multi_gpu.py      # Multi-GPU LAPD variant
│   ├── component_abla.py      # Component ablation entry point
│   ├── args.py                # Centralized argument parser
│   ├── model.py               # Model and tokenizer loading
│   ├── core/                  # Core likelihood and discrepancy computation
│   ├── dataloader/            # Unified benchmark loaders
│   ├── baselines/             # Baseline detector implementations
│   ├── scripts/               # Batch experiment scripts
│   └── utils/                 # Utilities for saving, seeding, and processing
├── formal_exp/                # Saved experiment outputs
├── LAPD.pdf                   # Paper draft
├── README.md
└── README_zh.md

Please download detector models into ./cache. For local model aliases in method/model.py, the expected directory names are:

llama2-7b          -> ./cache/llama-2-7b
llama2-7b-instruct -> ./cache/llama-2-7b-instruct
llama3.1-8b        -> ./cache/Meta-Llama-3.1-8B
llama3.1-8b-instruct -> ./cache/Meta-Llama-3.1-8B-Instruct
gpt-j-6b           -> ./cache/gpt-j-6b
gpt-j-6b-instruct  -> ./cache/gpt-j-6b-instruct

You can override the data and output roots with:

export DATA_ROOT=/path/to/data
export FORMAL_EXP_ROOT=/path/to/outputs

Citation

If you find this work useful, please cite:

@article{wu2026alignment,
  title={Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy},
  author={Wu, Junxi and Huang, Kailin and Hu, Dongjian and Chen, Bin and Wu, Hao and Xia, Shu-Tao and Zou, Changliang},
  journal={arXiv preprint arXiv:2604.16923},
  year={2026}
}

Reference Repositories

This project borrows ideas, baselines, or evaluation conventions from the following works:

• Fast-DetectGPT
• Binoculars
• DNA-DetectLLM