acceptodds
Under review as a conference paper at ICLR 2027

Inferring Large Language Model Provenance via Token Ranking

Abstract

Model provenance detection determines whether a model was derived from a claimed parent through fine-tuning or other adaptations. Many methods infer provenance relationships by measuring the similarity between two models from their generated responses. However, these approaches rely solely on sampled tokens, leaving the rich information associated with alternative candidate tokens unexplored. In this paper, we propose **RankMPT**, a novel method that leverages the rank information of candidate tokens to detect model provenance. Specifically, RankMPT measures the average rank correlation between the tail candidate tokens of two models across the prompt tokens, which can be approximated via repeated high-temperature sampling in black-box settings. We then perform a hypothesis test to decide whether the score is significantly higher than the correlation level expected in the absence of a provenance relationship. Our method captures discriminative signals from rarely selected tail tokens, making it less sensitive to confounding similarities among models arising from shared domains or tasks. Extensive experiments demonstrate the superiority of RankMPT, improving the average accuracy by up to 23% in the black-box setting of LeafBench.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model provenance, model lineage, large language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NmxkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores model provenance tracking, where the authors hope to label related models (same family, finetuning, etc) as provenance and unrelated models as non-provenance. To do this, they use a method based on the rank of tail tokens in a model distribution. The idea is that tail output distributions provide a discriminative signal since tuning will change primarily the head distribution. The authors motivate this with some example analysis early in the paper. They then propose the whitebox version of the algorithm, which uses access to the full output distributions. They also show a relaxation to the blackbox case, where sampling under a very high temperature can be used to approximate the ranks. They then test their method over a variety of open weight models and settings.

Critical strengths
  • Interesting problem and increasingly topical as model releases are heavily incentivized, though the effort can range from creating a new model from scratch to simply tuning an existing model.
  • Method has a motivating example and seems reasonable, though I have concerns for the execution. Using sampling to access tail tokens is interesting.
  • Uses a wide variety of relationships between models, including families, architecture, and training (finetuning) relationships.
Critical weaknesses

While the results and analysis are interesting, one key piece of the system seems to be missing: how does this method account for different vocabularies/tokenizers across models? In the long tail of tokens, for a given vocabulary, we would expect very few matches. A low match score in the blackbox method could simply indicate that there are no matching tokens. If you use string matching rather than exact token match across different vocabularies, these details should be provided. What about text normalization? Since results include models from various families, it seems apparent that there were cross-vocabulary comparisons. Concretely, how do you implement line 4 in algorithm 1, as L^p and L^c are defined over different vocabularies?

Questions for authors

How do you account for different vocabularies using this method? If you normalize strings, can you account for the fact that this seems closer to tokenization matching rather than something more insightful? Can you show some statistics around the intersecting tokens?

Reviewer_bvSPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies LLM provenance detection, which asks whether a suspected model was derived from a claimed parent model. It proposes RankMPT, a method that compares the rankings of low-probability candidate tokens shared by two models. The motivation is that these rankings preserve information inherited from the parent while being less affected by shared tasks or similar output behavior. With logit access, RankMPT computes rank correlations directly. In the black-box setting, it estimates rankings through repeated high-temperature sampling. It then combines the prompt-level correlations through a statistical test to decide provenance. The evaluation uses BenchA and LeafBench, covering multiple model families and six adaptation types. Additional experiments examine prompt shifts, sampling budgets, and difficult negative pairs such as sibling models. The reported results show gains over several existing provenance detectors.

Critical strengths
  1. The use of tail-token rankings is well motivated.
  2. The method achieves strong results on BenchA.
  3. The evaluation covers several model families and six adaptation types, with useful ablations on token selection, sampling budget, and decision rules.
Critical weaknesses
  1. The statistical proof does not fully match the stated method. The null hypothesis is defined using the mean raw correlation, but Appendix C treats it as equivalent to a bound on the mean Fisher-transformed correlation. These hypotheses are not equivalent in general because the transformation does not preserve means. The proof also assumes separate prompts for token selection and testing, while Algorithm 1 uses the same prompt set. These gaps leave the claimed false-positive guarantee unsupported for the procedure as described.
  2. LeafBench contains only positive pairs, so the reported accuracy is actually recall. A detector that labels every pair as related would score 100% on this evaluation. The headline 23-point improvement therefore shows better detection of known provenance pairs, but does not establish better rejection of unrelated models. Negative pairs are needed to evaluate the trade-off between detecting true relationships and making false provenance claims.
  3. Table 5 reports rejection accuracies of 68% for the white-box version and 80% for the black-box version. This means that 32% and 20% of sibling pairs are incorrectly classified as direct provenance pairs. Since the paper explicitly defines provenance as a direct parent–child relationship, these errors concern the central task. The evaluation also includes only 20 sibling pairs, limiting the evidence for robustness to shared ancestry.
  4. The practical cost of black-box detection is insufficiently evaluated. The default setup uses 768 samples per prompt at temperature 10. With 1,500 prompts, this requires roughly 1.15 million single-token samples per model. The sampling-budget ablation shows how accuracy changes with additional samples, but the paper does not report runtime, inference cost, or comparisons under matched query budgets. Performance under lower temperature limits is also unclear. These results are needed to assess whether the gains justify the required access and computation.
  5. The paper does not compare with RAFP [1], Stemma [2], LLMPrint [3] or lineage-attestation method [4]. Comparisons under compatible access settings are needed to establish RankMPT’s advantage over these alternatives.

[1] RAFP: Identifying LLM Lineages via Rare-Region Fingerprints

[2] Stemma: Induced Decision Regions Reveal LLM Provenance

[3] Fingerprinting LLMs via Prompt Injection

[4] Attesting Model Lineage by Consisted Knowledge Evolution with Fine-Tuning Trajectory

Questions for authors

See weaknesses.

2 comments

Thank you for the response. The corrected null hypothesis and proof resolve my first concern, and I accept the clarification that LeafBench measures recall. Three issues remain:

  1. On the expanded set, 42% of sibling pairs in the white-box setting and 23% in the black-box setting are still labeled as direct provenance, while the test runs at a 5% level. This means the null hypothesis does not cover siblings, so the claim of robustness to confounding similarity should be narrowed. The gap between RankMPT-BX and NT-FWER is also within one standard deviation.

  2. Your cost table shows that RankMPT needs T times more queries than NT-FWER, with T at least 500. The reported 0.21 seconds excludes sampling, which is the real cost. I would still like to see a comparison under the same query budget, and results at temperatures that public APIs allow.

  3. Stemma and LLMPrint verify the suspect model with black-box queries, but they are compared only with RankMPT-WB. RankMPT-BX reaches 84% on BenchA, which is below the 85% you report for RAFP. Please add the black-box comparison.

Thank you for the additional results. The new baseline table in Appendix E.3 resolves my third point, but I will keep my rating.

The black-box variant is evaluated only in a local simulation at temperature 10, which public APIs do not allow. Results at lower temperatures are still missing, and the revised paper mentions neither this limit nor the runtime comparison. On sibling pairs, the expanded results are also absent from the paper: the main text still reports the original 20 pairs, and the introduction still claims robustness on hard negatives.

The white-box results are strong, but the black-box contribution needs evidence under realistic access before the paper is ready.

Reviewer_6AcfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RankMPT, a new method to detect LLM provenance, and more specificaly which can detect parent-child relationship between two LLMs CC (a potential child) and PP (the potential parent), where one has been obtained by adapting the other (via general fine-tuning, instruction fine-tuning, distillation, or other transformations).

The intuition of the paper is that the ranks assigned by two models to tail tokens (the tokens with low probability and low logits) is a signal that can detect provenance. In particular, the authors observe that the rank of tail tokens is better preserved between a parent and a child model (when the provenance relationship is effective) than for pairs of models without provenance relationship.

The technique called RankMPT is simple, and it works as follows: for a set of prompts used as inputs, RankMPT looks at the average logits produced by the two LLMs for each token in the vocabulary, than it selects the two list of tail tokens and takes the intersection of both lists. For each prompt, it then compares the ranks (ie, the positions is the lists) of these tokens between the two models using a Spearman correlation. If C is obtained by adapting P, then the correlation is higher than if not. A statistical test is used to decide whether the correlation is sufficiently above a given threshold to conclude about the parent-child relationship.

Without access to the logits (black box mode), the ranks can be approximated by repeating the token generation with a high temperature (typically T=10) and by computing the frequency of each token. This gives a ranking which is less precise than using the logits, and all the same tail tokens are not recovered, but the results show that it also provides a useful signal to detect provenance.

The experiments are conducted on two datasets and many models (110 and 65 respectively), with 15 reference parent models. The experiments show that RankMPT performs better than the methods used for comparison. This shows that the ranking of low probability tokens is a good signal for detecting parent–child relationship both in white box (using logits) and black box, compared with existing methods.

Critical strengths
  • The problem addressed by the paper is relevant and the presentation of the method called RankMTP method is clear

  • The experimental evaluation is diverse, it tests many pairs of models, and it considers several transformations, thus showing the gain provided by the method compared with related work

  • The support of the black box mode (without logits) is a nice addition, although its practical cost remains to be clarified

Critical weaknesses
  • The paper defines provenance as the direct parent–child Relationship between two models, but it seems that RankMPT has difficulty to work well when sibling models are considered, which is an important difficulty (I appreciate it is well tested in the experiments).

  • The contribution seems incremental compared to existing work. For example, SeedPrints (which is discussed in the paper and compared in the experiments) already exploits low mean scores and rank correlations, and it potentially uses logits. RankMPT differs by comparing ranks across tokens for each prompt, which is weel explained in the paper.

  • The black box support is very interesting, but it is more costly (as saifd in the paper), because it requires many samples. A comparison of the results of RankMPT-balck-box with the other methods, for given budgets (in time or in tokens), would help better evaluate their respective peformances in practice.

Questions for authors
  • After the Fisher transformation which is non linear, I am wondering if the test provided in equation 6 still corresponds to hypothesis H_0. Is it the case, or should H_0 be formulated using the transformed correlations ?

  • The experiments compares RankMPT with SeedPrints, which seems a very close method. But the comparison with SeedPrints uses the internal representaiotn but not the logits. I guess the results would be the same using the logits (as SeedPrint reports comparable results with logits and internal representations), but a comparison with the same logits, the same prompts and the same lists of tail tokens might help better identify the source of the gains. Did you test this comparaison ?

  • Did you measure the gains of RankMPT in the black-box under fixed/given budgets (in number of token processed and generated or in computation time) with competitors ?

  • The results in Tables 5 and 6 seems to be obtained with alpha = 5%. But if so, could you clarify if the pairs without parent–child relationship in these experiments satisfy H_0 ? I am wondering how the false detection rates of around 20 to 32% should be interpreted.

  • In Figure 1(a) what is the precise definition of the "rank shift", with the normalization and the aggregation ? I an wondering if the benefit of looking at the tail persist after controlling for the effect of the initial rank and sensitivity to logit perturbations ?

  • Could you discuss a bit more the case of reasoning LLMs ? it seems to me that it would possibly make the technique more difficult to use, if reasonning steps preceedes the analysed outputs. Would the high temperature approach used for the black-box mode still provide a usefull signal with reasoning LLMs ?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.