AWM: Accurate Weight-Matrix Fingerprint for Large Language Models
ICLR 2026

Training-free model fingerprint with LAP-aligned unbiased CKA

     

This repository contains the official implementation for the paper "[AWM: Accurate Weight-Matrix Fingerprint for Large Language Models](https://arxiv.org/abs/2510.06738)"

AWM is a fingerprinting method to determine whether one large language model (LLM) is derived from another base model. The method combines Linear Assignment Problem (LAP) based dimension alignment with unbiased CKA on attention weights, and is robust to common post-training changes.

News

• [2026.1] AWM is accepted at ICLR 2026!
• [2025.10] Code and paper released.

Table of Contents

• Quick Feature Summary
• Results
  • ROC Curves
  • Absolute Z-score Table for correlated models
• Installation
• Quick Start
  • A. Compare Two Local Models
  • B. Run a Predefined Experiment
• Model Download
• Experiment Suites
• Repository Structure
• Citation

Quick Feature Summary

Feature Category	AWM Capability
Goal	Independency test of open-source LLMs based on their weights
Method	1️⃣ Linear-Assignment-Problem-based (LAP-based) alignment 2️⃣ Unbiased Central Kernel Alignment (UCKA)
Detectable Weights	✅ Attention weights ✅ (informal) FFN weights
Computation Cost	✅ Training-free ✅ ~30s/pair on one RTX3090
Accuracy	✅~0 similarity scores for independent models ✅High similarity scores for correlated models
Robustness to training recipes	✅ Supervised fine-tuning (SFT) ✅ Continual pre-training (CPT) ✅ Reinforcement learning post-training (RL) ✅ Multimodal tuning ✅ Pruning ✅ MoE upcycling
Robustness to weight manipulations	✅ Constant scaling ✅ Permutaion matrices ✅ Signature matrices ✅ Orthogonal matrices

Results

For more detailed results, please refer to the paper.

ROC Curves

We show that our method consistently separates correlated model pairs from independent pairs across thresholds.
In the low-false-positive region, our ROC curves stay higher and achieve larger partial AUC (pAUC), indicating stronger practical fingerprint verification performance.

Absolute Z-score Table for correlated models

Absolute Z-score measures how many standard deviations a positive-pair similarity score is away from the mean of the independent-pair (negative) distribution. Larger values mean stronger statistical separation and lower ambiguity in lineage identification.

We further show that our method produces substantially larger absolute Z-scores than prior baselines. Across SFT, CPT, MoE upcycling, multimodal tuning, RL post-training, and pruning settings, we consistently achieve the strongest separation margins between correlated and independent pairs.

Training Method	Base	Offspring	HuRef	REEF	Ours
SFT	Llama2-7B	vicuna-7b-v1.5	44.22	1.95	355.09
	Llama2-7B	llama-2-7b-finance	44.20	1.94	355.02
	Llama2-7B	selfrag_llama2_7b	44.47	1.95	355.20
	Llama2-7B	LLaMA-2-7B-32K	44.22	1.88	354.84
	Llama2-7B	WizardMath-7B-V1.0	44.51	1.95	355.23
	Llama2-13B	llama-2-7b-guanaco	44.12	1.94	354.55
	Llama2-13B	vicuna-13b-v1.5	44.25	1.94	355.02
	Llama2-13B	Nous-Hermes-Llama2-13b	44.48	1.95	355.20
	Llama2-13B	LLaMA2-13B-Estopia	43.09	1.91	343.14
	Llama2-13B	firefly-llama2-13b	44.51	1.95	354.59
CPT	Llama2-7B	llemma_7b	8.15	1.66	250.32
	Llama2-7B	CodeLlama-7b-hf	9.41	1.31	250.78
	Llama2-7B	CodeLlama-7b-Python-hf	9.19	1.77	253.28
	Gemma-2B	codegemma-2b	28.97	0.32	198.18
	Gemma-7B	codegemma-7b	39.51	1.52	268.72
	Qwen2.5-7B	Qwen2.5-Math-7B	0.28	1.04	177.57
	Qwen2.5-7B	Qwen2.5-Coder-7B	0.57	1.29	135.17
	Qwen2-7B	Qwen2-Math-7B	1.01	1.09	183.49
	Llama2-70B	CodeLlama-70b-hf	3.43	1.93	241.12
	Llama2-70B	CodeLlama-70b-Python-hf	2.78	1.92	232.92
MoE Upcycling	Mistral-7B	Nous-Hermes-2-Mixtral-8x7B-DPO	7.97	1.47	239.01
	Llama3-8B	LLaMA-MoE-v2-3_8B-2_8-sft	42.50	0.47	332.12
	Llama2-7B	LLaMA-MoE-v1-3_5B-4_16	24.58	0.64	332.23
	Llama2-7B	LLaMA-MoE-v1-3_0B-2_16	21.59	0.64	326.49
	Llama2-7B	LLaMA-MoE-v1-3_5B-2_8	24.91	0.62	332.12
	Llama2-7B	LLaMA-MoE-v1-3_0B-2_16-sft	21.59	0.63	326.49
	Llama2-7B	LLaMA-MoE-v1-3_5B-2_8-sft	24.91	0.63	332.12
	Llama2-7B	LLaMA-MoE-v1-3_5B-4_16-sft	24.58	0.62	332.23
	Qwen-1.8B	Qwen1.5-MoE-A2.7B	4.03	0.57	173.36
	Minicpm-2B	MiniCPM-MoE-8x2B	11.39	1.49	150.15
Multimodal tuning	Llama2-7B	llava-v1.5-7b	44.06	0.40	354.98
	Llama2-7B	Video-LLaVA-7B-hf	44.03	0.36	354.98
	Qwen2-7B	Qwen2-VL-7B-Instruct	39.94	0.84	342.72
	Qwen-7B	Qwen-Audio	38.79	0.19	339.86
	Qwen-7B	Qwen2-Audio-7B	21.59	0.45	317.97
	Qwen-7B	Qwen-VL	37.41	0.77	336.55
	Qwen2.5-7B	Qwen2.5-VL-7B-Instruct	24.28	0.48	290.08
	Qwen2.5-3B	Qwen2.5-VL-3B-Instruct	23.95	0.61	298.46
	Llama3-8B	llama3-llava-next-8b-hf	44.30	0.26	355.05
	Llama2-13B	llava-v1.5-13b	44.09	0.07	354.91
RL post-training	Open-llama3B	hh_rlhf_rm_open_llama_3b	44.52	1.94	355.23
	Qwen2.5-7B	Nemotron-Research-Reasoning-Qwen-1.5B	44.58	1.93	355.27
	Qwen2.5-1.5B	Open-Reasoner-Zero-1.5B	44.58	1.94	355.27
	Mixtral	Nous-Hermes-2-Mixtral-8x7B-DPO	44.57	1.75	355.23
	Mistral-7B	Nous-Hermes-2-Mistral-7B-DPO	44.53	1.48	355.23
	Mistral-7B	dolphin-2.6-mistral-7b-dpo	44.54	1.21	355.23
	Minicpm-2B	MiniCPM-2B-dpo-bf16	44.58	1.92	355.27
	Qwen3-4B	Qwen3_Medical_GRPO	44.58	1.78	355.27
	Chatglm-6B	chatglm-fitness-RLHF	44.58	1.96	355.27
	Llama3-8B	LLaMA3-iterative-DPO-final	44.52	1.96	355.23
Pruning	Llama-3-8B	Llama-3.1-Minitron-4B-Depth-Base	28.29	0.52	344.07
	Llama-3-8B	Llama-3.1-Minitron-4B-Width-Base	22.23	0.53	343.00
	Llama-3-8B	Llama-3.2-1B	0.33	1.15	12.14
	Llama-3-8B	Llama-3.2-3B	0.73	0.92	106.29
	Llama2-7B	Sheared-LLaMA-2.7B-Pruned	22.88	1.75	328.81
	Llama2-7B	Sheared-LLaMA-2.7B-ShareGPT	16.00	1.78	312.44
	Llama2-7B	Sheared-LLaMA-2.7B	15.86	1.77	312.80
	Llama2-7B	Sheared-LLaMA-1.3B-Pruned	10.06	1.80	317.79
	Llama2-7B	Sheared-LLaMA-1.3B	7.64	1.79	297.50
	Llama2-7B	Sheared-LLaMA-1.3B-ShareGPT	7.64	1.79	297.50

Installation

git clone https://github.com/LUMIA-Group/AWM
cd AWM
pip install -r requirements.txt

Quick Start

A. Compare Two Local Models

Download or place two model checkpoints (e.g. Llama-2-7b and CodeLlama-7b-hf) locally, then run:

python main.py \
  --model_paths ./checkpoints/Llama-2-7b ./checkpoints/CodeLlama-7b-hf \
  --device auto

--device supports auto, cpu, and cuda.

B. Run a Predefined Experiment

python main.py --config sft_pairs --device auto

This mode uses model groups and analysis pairs defined in configs.py.

Model Download

AWM provides two ways to prepare checkpoints.

1. Direct download with huggingface-cli:

huggingface-cli download meta-llama/Llama-2-7b --local-dir ./checkpoints/Llama-2-7b
huggingface-cli download codellama/CodeLlama-7b-hf --local-dir ./checkpoints/CodeLlama-7b-hf

2. Batch download with helper script:

chmod +x download_models.sh
./download_models.sh             # list supported config keys
./download_models.sh sft_pairs   # download models for one config
./download_models.sh all         # download all mapped models

Note: keep CHECKPOINT_BASE_DIR in configs.py and download_models.sh consistent with your local storage path.

Experiment Suites

Predefined config keys:

• sft_pairs
• cpt_pairs
• rl_pairs
• multimodal_pairs
• pruning_pairs
• all_moe_pairs
• sft_13b_7b_pairs
• independent_pairs
• independent_pairs_13b

Suggested reproducibility flow:

1. Download models for one suite, e.g. ./download_models.sh cpt_pairs.
2. Run analysis, e.g. python main.py --config cpt_pairs --device auto.
3. Compare summary scores in the console output (Wq_weights, Wk_weights, Wq_Wk_weights).

Repository Structure

AWM/
├── checkpoints/            # model checkpoints
├── asset/                  # static assets (logo/icons)
├── main.py                 # experiment entry point
├── similarity_metrics.py   # CKA, weight loading, LAP alignment
├── configs.py              # experiment definitions + model mapping
├── download_models.sh      # batch downloader for Hugging Face models
└── requirements.txt        # dependencies

Citation

If you find AWM useful in your research or applications, we would appreciate it if you could cite our work:

@article{zeng2025awm,
  title={{AWM: Accurate Weight-Matrix Fingerprint for Large Language Models}},
  author={Boyi Zeng and Lin Chen and Ziwei He and Xinbing Wang and Zhouhan Lin},
  year={2025},
  journal={arXiv preprint arXiv:2510.06738},
}