AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

SeedPrints: Fingerprints Can Even Tell Which Seed Your Large Language Model Was Trained From

Yao Tong, Haonan Wang, Siquan Li, Kenji Kawaguchi, Tianyang Hu

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use initialization-bias SeedPrints test instead of post-hoc fingerprints for early pretraininglineage verification · seed vs trained model, 12-layer LLaMA-stylelineage detection: baselines 0.580 PCS / 0.375 REEF below 0.8 threshold vs p=3.37e{-}26Improved
test SeedPrints across different init seeds with identical data and order (ablation)SeedPrints · cross-seed pairs on OpenWebTextp-value: p-values 0.109 to 0.883 (>0.01, distinguishes seeds)Improved
continual pretraining on code data (The Stack) distribution shift (ablation)baselines (Intrinsic, REEF, PCS, ICS) · continual training from seed-1000 baselineage detection: Intrinsic 0.489, REEF 0.557, PCS 0.585, ICS 0.123 on true descendantHurt
continual pretraining on The Stack for SeedPrints (ablation)SeedPrints · continual training from seed-1000 baselineage detection: p approximately 0 for true descendantNo clear effect
evaluate on early OLMo-2-7B checkpoints (5B tokens onward)baselines vs SeedPrints · OLMo-2-7B Stage 1 vs final checkpointlineage detection: baselines ICS~0, PCS<0.3, REEF<0.6 in first trillion tokens; SeedPrints p<<0.01Improved
large-scale finetuning up to 500B-700B tokensPCS and ICS baselines · Llama-2-7B to CodeLlama-7B / Llemma-7Bsimilarity score: ICS 0.3369 (CodeLlama), 0.2905 (Llemma)Hurt
weight-space merging with unrelated modelsSeedPrints · LeaFBench MergeAUC: 1.000 (Instruct) -> 0.959 (Merge)Hurt
increase number of random input sequences n (ablation)SeedPrints · seed identificationaccuracy: 0.7368 -> 0.9375Improved
increase random sequence length l (ablation)SeedPrints · seed identificationaccuracy: 0.9048 -> 1.0000Improved
increase fingerprint size m to 800 (ablation)SeedPrints · seed identificationempirical FPR: 0 -> 0.1463Hurt
use final-layer hidden states instead of logits as model output (ablation)SeedPrints · random-input bias estimationnumber of random inputs neededImproved
use mean response vector instead of per-sample argmin frequency (ablation)SeedPrints · identity dimension extractionstabilityImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy