SeedPrints: Fingerprints Can Even Tell Which Seed Your Large Language Model Was Trained From
In a groundbreaking study published on arXiv (reference: 2509.26404v2), researchers have unveiled a novel approach to fingerprinting Large Language Models (LLMs) known as SeedPrints. This technique provides a robust method for provenance verification and model attribution, addressing significant gaps in existing fingerprinting methodologies.
Background on Fingerprinting LLMs
Fingerprinting LLMs has become increasingly critical as the demand for model accountability and traceability grows in the artificial intelligence community. Traditional fingerprinting techniques focus primarily on models after they have undergone fine-tuning, a stage where they develop stable signatures influenced by their training data and optimization processes. However, it is during the pretraining phase that a model acquires most of its capabilities, making the need for effective lineage verification during this phase essential.
Challenges with Existing Techniques
Existing fingerprinting methods have been found to be unreliable in the pretraining context. They typically depend on post-hoc signatures that emerge only after a significant amount of training has occurred. This reliance on later training stages contradicts the classical understanding of a fingerprint as an intrinsic and consistent identifier. Thus, the challenge lies in identifying a method that can ascertain model lineage from the very beginning of the training process.
Introducing SeedPrints
The research team proposes SeedPrints, an innovative approach that capitalizes on random initialization biases as enduring, seed-dependent identifiers. This method asserts that even before formal training commences, untrained models display reproducible prediction biases that can be traced back to their initialization seed. These biases are not merely ephemeral; they persist throughout the training process, enabling high-confidence lineage verification.
Key Features of SeedPrints
SeedPrints boasts several advantages over previous fingerprinting techniques:
- Persistence: The seed-dependent identifiers are intrinsic to the model and detectable from the outset of training.
- Robustness: Unlike prior methods that falter during early pretraining or under shifting distributions, SeedPrints maintains effectiveness throughout all training phases.
- Comprehensive Evaluation: Experiments conducted on LLaMA-style and Qwen-style models demonstrate the method’s ability to distinguish models at the seed level and facilitate identity verification from the moment of initialization through to full pretraining and adaptation.
Empirical Validation
The research findings include extensive evaluations conducted on large-scale pretraining trajectories alongside real-world fingerprinting benchmarks. These evaluations confirm SeedPrints’ robustness, showing that it remains reliable under prolonged training, domain shifts, and modifications to model parameters.
Conclusion
SeedPrints represents a significant advancement in the field of model fingerprinting, offering a method that not only addresses the limitations of existing techniques but also enhances the ability to trace LLMs back to their origins. As the AI landscape continues to evolve, such innovations will play a crucial role in ensuring accountability and transparency in model development.