One Sequence Is Enough: Fingerprinting Language Models from Fixed-Length Integer Sequences
Organizations: Faculty of Computing, Harbin Institute of Technology Harbin, China
Abstract
Users of model APIs and third-party platforms often rely on the declared model name to identify the model responding to their requests. However, the actual model may differ from the declared one, motivating identity verification based on observable outputs. Previous studies have shown that preferences in simple numerical-choice tasks can serve as model fingerprints. We investigate whether a single fixed-length integer sequence can combine multiple numerical observations to identify its source model. For this task, we introduce NumPrint, a benchmark of 5,036 answers from 51 models, collected using Chinese, Japanese, and English prompts that request 294 integers between 1 and 355 in one response. We propose the Shallow MLP and Causal Transformer (SMCT) to capture aggregate numerical preferences and behavioral patterns in generated sequences. In 51-class closed-set evaluation, SMCT achieves 92.72% accuracy from a single sequence, outperforming all methods compared in this study. Using two and four answers increases accuracy to 96.06% and 97.12%, respectively. These results show that a single integer sequence contains useful model fingerprints and that additional answers further improve identification.
Figures & tables
| Prompt language | Train | Val | Test | Total |
|---|---|---|---|---|
| Chinese | 1148 | 229 | 245 | 1622 |
| Japanese | 1188 | 260 | 263 | 1711 |
| English | 1189 | 266 | 248 | 1703 |
| Total | 3525 | 755 | 756 | 5036 |
| Method | Acc. | Macro-F1 |
|---|---|---|
| Random forest | 68.39 | 61.88 |
| Transformer | 79.10 | 75.52 |
| CatBoost | 82.54 | 79.83 |
| XGBoost | 82.80 | 80.01 |
| MultiROCKET + statistics | 86.90 | 83.86 |
| Ridge | 87.96 | 85.13 |
| Variant | Acc. | Macro-F1 |
|---|---|---|
| SMCT (full) | 92.72 | 91.10 |
| w/o statistical encoder | 87.43 | 85.27 |
| w/o causal encoder | 91.27 | 89.57 |
| Model errors | Cross-family errors | Family Acc. | Family Macro-F1 | ||
|---|---|---|---|---|---|
| 1 | 756 | 55 | 18 | 97.62 | 96.49 |
| 2 | 5867 | 231 | 27 | 99.54 | 98.82 |
| 3 | 5867 | 194 | 1 | 99.98 | 99.94 |
| 4 | 5867 | 169 | 0 | 100.00 | 100.00 |
| 5 | 5867 | 153 | 0 | 100.00 | 100.00 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Collected label | Train | Val | Test | Collected label | Train | Val | Test |
|---|---|---|---|---|---|---|---|
| gpt-6-luna | 469 | 100 | 101 | qwen3.8-27b | 52 | 11 | 12 |
| grok-4.6 | 176 | 38 | 38 | qwen3.8-max | 52 | 11 | 11 |
| gemini-3.8-flash | 132 | 28 | 28 | gpt-6-astra | 51 | 11 | 11 |
| claude-opus-5.5 | 116 | 25 | 25 | qwen3.7-max | 50 | 11 | 10 |
| glm-5.3-flash | 113 | 24 | 24 | qwen3.5-397b-a17b | 50 | 11 | 10 |
| muse-spark-1.3 | 113 | 24 | 24 | glm-5.1 | 48 | 10 | 11 |
| Feature family | Dimensions |
|---|---|
| Integer histogram | 355 |
| Four positional histograms | 64 |
| Final-digit histogram | 10 |
| Modulo histograms | 65 |
| Tens/hundreds histograms | 14 |
| Signed/absolute increments | 48 |
| Method | Acc. | Macro-F1 |
|---|---|---|
| LDA | 91.27 | 89.45 |
| Logistic regression | 89.68 | 88.04 |
| Ridge | 87.96 | 85.13 |
| RBF-SVM | 88.49 | 86.67 |
| Random forest | 68.39 | 61.88 |
| XGBoost | 82.80 | 80.01 |
| Evaluated | Acc. | Macro-F1 | |
|---|---|---|---|
| 1 | 756 | 92.72 | 91.10 |
| 2 | 5,867 | 96.06 | 94.15 |
| 3 | 5,867 | 96.69 | 95.22 |
| 4 | 5,867 | 97.12 | 95.64 |
| 5 | 5,867 | 97.39 | 96.03 |
| Acc. | Macro-F1 | |
|---|---|---|
| 32 | 50.26 | 42.58 |
| 64 | 70.63 | 64.98 |
| 128 | 83.60 | 80.88 |
| 192 | 88.36 | 86.50 |
| 256 | 91.40 | 90.38 |
| 294 | 92.72 | 91.10 |
| Hidden width | Parameters | Acc. | Macro-F1 |
|---|---|---|---|
| 2,039,859 | 91.14 | 89.52 | |
| 1,019,955 | 91.27 | 89.57 | |
| 628,275 | 90.61 | 88.97 | |
| 320,819 | 90.48 | 88.69 | |
| 281,395 | 90.74 | 89.21 | |
| 160,563 | 86.51 | 83.39 |
Language Model Fingerprinting
Language model fingerprinting identifies and verifies models or their lineage from distinctive outputs, internal representations, or embedded signals.
See all 87 papers