acceptodds
Under review as a conference paper at ICLR 2027

Provable Model Provenance Set for Large Language Models

Abstract

The rapid proliferation of customized large language models has intensified concerns over unauthorized model reuse, highlighting the need for reliable model provenance techniques for model auditing and tracking. However, existing work largely infers provenance relationships through heuristic fingerprint‑matching rules and lack of provably valid error control, leaving the credibility of provenance claims unverified. In this work, we formalize model provenance as a set‑valued inference problem and propose the Model Provenance Set (MPS), a novel principled method to adaptively construct a provenance set with rigorous coverage guarantee, i.e., covering the source models at a specified confidence level. Specifically, we introduce a novel equivalence test that yields a valid -value under the null hypothesis that all candidates are equally close to the target model, and develop a test-and-exclusion procedure to sequentially select the most probable source candidate until the null is rejected. By establishing the sequential structure of the test, MPS constructs a subset of candidates with a provable coverage guarantee. Extensive experiments demonstrate that MPS can effectively identify all source models with the intended coverage while strictly avoiding the inclusion of unrelated models.

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

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oKeUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a method MPS for model provenance. Given a target model and a candidate models set, it tries to find a small source set that contains all source models of the target model. MPS computes a distance for each prompt between each candidate and the target model. Then it centers the distances per prompt, computes a studentized t-statistic for each candidate (Eq. 6), takes T_min, and gets a p-value by permuting model labels within each prompt (Eq. 9). It then iteratively moves the argmin candidate into the source set while the null "all candidates are equally close" is rejected. The overall process is kinda like Model Confidence Set but in reverse [1]. Overall, the paper have a good theoretical grounding. The statistics machinery might not be new but the formulation is novel and applies to the problem well.

[1] Hansen, Peter R., et al. “The Model Confidence Set.” Econometrica, vol. 79, .no 2, Econometric Society, 2011, pp. 453-497, https://doi.org/10.3982/ECTA5771

Critical strengths
  1. The problem is important and timely. Treating multi-source provenance as set-valued inference, including the "no source" case, is a useful framing that many existing tests do not provide in a measurable way.

  2. The method needs only black-box access, works with any distance, is simple, and is cheap once outputs are cached.

  3. There is a sizeable benchmark built from real Hugging Face lineage metadata (562 models).

  4. The sensitivity analyses are broad (α, R, T, |M|). The same-family study (G.3) honestly reports failure modes.

  5. Code and data are released. The paper is generally clearly written.

Critical weaknesses
  1. The null and the assumptions can be unrealistic. The authors' own G.3 shows the assumption breaks with same-family candidates, and this can limit the effectiveness of the proposed method in cases the models in the candidate sets are all derived from the same ancestor model or shared large amount of overlapping training data.

  2. The benchmark sidesteps the hard cases. Negatives in the experiments are drawn from different families, however real audits face siblings, same-family models and descendants. In addition, since similarity is symmetric, MPS cannot tell ancestors from descendants or siblings.

  3. The guarantee is framed incorrectly. A rejection adds a model to source set M^\hat{M}. So the type-I error α\alpha governs false inclusion, not missed sources. Fig. 3(a) shows exactly what you would expect if α\alpha controls inclusions: coverage stays flat as α\alpha grows, while set size grows.

  4. The paper is internally inconsistent in several places.

    a. Fig. 10 contradicts Fig. 2. In Fig. 10, Semantic-MPS with |M*| = 1 has mean first-stage p ≈ 0.22-0.32. Since p ≤ 1, P(p ≤ α) ≤ (1−E[p])/(1−α), which caps first-stage rejection at about 0.72-0.82. A source is included only if the first stage rejects, so coverage cannot exceed this. Yet Fig. 2 reports coverage of about 0.95 in the same settings.

    b. |M| = 0 "coverage" is always 1.0, yet set sizes are above 0.* Fig. 3(a) shows set size ≈ 2 at α = 0.3 and |M| = 100; the bars in Fig. 2 are about 1. Under the paper's definition this coverage is trivially 1. Under the code's definition (mcs_evaluation), the output must be empty. Table 4 uses yet another, non-trivial value (0.97).

    c. Fig. 10's null mean p is about 0.85–0.98, whereas valid p-values under an exact null average about 0.5. The test is strongly conservative or uses a mismatched null. Reading a large p as "evidence of non-infringement" (G.5) confuses absence of evidence with evidence of absence.

    d. The code in the anonymous repo is inconsistent with the paper.

  5. There's label noise: unsloth/* models (34 of the 462 depth-1 chains) are mostly re-uploads of the identical model, about half of them have byte-identical weight files comparing to the upstream model and the other half are mostly bf16 converted versions. They're not derived child and detecting them is trivial. Shouldn't weaken the main conclusion of the experiments, though.

Questions for authors
  1. Please restate exactly what α controls. Report coverage separately from the false-inclusion rate, preferably including for |M*| = 0.

  2. How do you reconcile Fig. 10 with Fig. 2?

  3. How are Bench-A pools of size 50 built? Does the FWER extension correct for multiplicity across candidates? How was REEF's threshold tuned?

Reviewer_kNq9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies LLM provenance as the inference of a set of possible source models, including multiple ancestors in a derivation chain. It proposes Model Provenance Set (MPS), which repeatedly applies a permutation test to relative model distances and selects the most statistically distinguishable candidate until testing stops. The framework accommodates different behavioral similarity scores and presents asymptotic coverage and exact-recovery guarantees under separability and exchangeability assumptions. Evaluation on a collected benchmark of 562 models and an existing provenance benchmark examines source recovery, set size, sensitivity to testing parameters, and robustness to related distractor models.

Critical strengths
  1. The set-valued formulation addresses a meaningful limitation of single-source attribution. Separating the retention of true ancestors from the inclusion of unnecessary candidates provides a useful framework for provenance screening and uncertainty assessment.

  2. The evaluation covers a substantial collection of model lineages, two behavioral similarity measures, and varying candidate pools. The analysis of same-family distractors is particularly informative because it reveals how strongly provenance inference can depend on the composition of the comparison set.

Critical weaknesses
  1. The central statistical guarantee rests on an unresolved distinction between ancestry and similarity. The exact-recovery proof treats a set of unrelated models as satisfying the equal-distance null, although unrelated models can have systematically different similarities to the target. The stated assumptions do not establish this implication, so the procedure can continue selecting unrelated candidates after all true sources have been found. The separate coverage theorem is asymptotic and does not establish the finite-query false-negative control suggested by parts of the empirical discussion.

  2. The reported calibration evidence is not consistently aligned with the defined statistical quantities. Table 4 reports coverage below one when there are no true sources, although coverage as defined is automatically satisfied in that case. Appendix D also describes a nearly zero-centered, symmetric permutation distribution that is difficult to reconcile with the defined minimum statistic. Together with coverage values below the nominal target in Figure 2, these inconsistencies undermine the claimed empirical confirmation of the guarantees.

Questions for authors
  1. What assumption ensures that unrelated candidates satisfy the equal-distance null required by the exact-recovery proof? How does the guarantee apply when unrelated candidates differ systematically in similarity to the target?

  2. What quantities were actually computed for the no-source coverage results and the permutation diagnostics? How do these correspond to the definitions used in the theoretical analysis?

Reviewer_Ks5oOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies a set of possible source models for a target LLM. Its method, MPS, repeatedly tests model similarities and selects likely sources. The paper provides asymptotic guarantees under stated assumptions and evaluates the method on collected model lineages.

Critical strengths
  • Predicting a source set supports models with multiple ancestors rather than only one source.

  • Explicit statistical guarantees make the uncertainty of provenance claims easier to assess.

Critical weaknesses
  • The guarantees rely on strong separability and exchangeability assumptions that may fail for closely related models.

  • The main benchmark uses unrelated models from different families, making it less representative of difficult same-family attribution.

Questions for authors
  • How reliable is MPS when the candidate pool contains many same-family models that are not direct ancestors?

  • How should users interpret the coverage guarantee with a limited number of prompts?

  • Can MPS distinguish actual ancestry from similarity caused by shared training data?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.