- 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?
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:
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.
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.
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.