One integer sequence is enough to name the model: NumPrint identifies among 51 models at 92.72% from a single answer
Synopsis
Users of model APIs usually judge which model answered from the declared name, although the actual model may differ. The NumPrint benchmark collects 5,036 answers from 51 models, each answer giving 294 integers between 1 and 355 under Chinese, Japanese and English prompts; SMCT pairs a statistical-feature encoder with a causal attention encoder and reaches 92.72% accuracy and 91.10% macro-F1 in 51-class closed-set identification from one sequence, above every method compared.
Interpretation
A single fixed-length decimal integer sequence can itself serve as a language-model fingerprint, identifying the source model within a candidate set without relying on the model name a platform declares. The earlier integer-sequence fingerprinting method lm-detector mixed prompts and requested lengths in its reference bank, leaving too few comparable samples at a fixed observation length, and its representation relied mainly on frequency and coarse positional summaries with limited use of adjacent changes and dependencies. On 756 test answers over 51 classes, SMCT reaches 92.72% accuracy and 91.10% macro-F1 from one 294-integer answer, 1.45 and 1.53 points above the strongest baseline MLP at 91.27% and 89.57%, and 3.83 and 4.45 points above lm-detector at 88.89% and 86.65%.
Identification draws on two kinds of information at once: statistics of value preferences over the whole sequence, and the generation-order relation of each integer to those before it. The statistical-feature encoder maps 944 predefined statistics through a single-hidden-layer feedforward network to class scores, while the causal attention encoder embeds each integer as a 64-dimensional vector, applies single-layer single-head causal self-attention, and pools positions with a learned weighted readout. An adaptive classification module weights the two encoders' outputs by confidence and applies a per-class linear correction; in ablation, removing the statistical encoder lowers accuracy and macro-F1 by 5.29 and 5.83 points, and removing the causal encoder by 1.45 and 1.53 points.
Perspective
The demonstrated scope is closed-set identification among the 51 known models in NumPrint, using answers in which each model generates 294 integers from 1 to 355 under Chinese, Japanese and English prompts; the sequence-length experiment truncates prefixes of collected answers rather than comparing prompts that request different lengths.
Extending identification beyond the candidate set requires unknown-model detection and updates to the reference set; system prompts, generation settings, reasoning modes and conversation context were not independently controlled, so how these factors change numerical fingerprints and whether identification transfers between service configurations remain unclear.
