acceptodds
Under review as a conference paper at ICLR 2027

Same Artifact, Two Faces: Why Black-Box Lineage Fingerprints Look Both Robust and Fragile

Abstract

Knowing which training branch a deployed model is produced from matters for model auditing and policy. We ask to what extent the post-training branch of same-family models can be identified from black-box fingerprints alone, i.e. output texts. Qwen3 base/IT pairs are the most comparable case: under a pre-training continuation format, both branches generate readable text. We collect 100 to 200 MATH500 problems per model and train a 16-feature output classifier. Within a single scale the separation is near-perfect (0.941 / 0.947), but only when the reading is uncontrolled: it is not stable across the five scales we ran, and it moves further than the headline effect we report when we change the prefix length we keep, chosen arbitrarily. Once length and problem difficulty are controlled together, the leave-one-model-out transfer is no longer distinguishable from chance, and it does not collapse to a clean negative either. Disguise, prompt and answer rewriting, template and temperature perturbation, and cross-lingual attacks all leave the reading without a discernible decrease, because the signal carrying it is termination and length; the one arm that moved the reading moved it by shortening the output, which is that channel.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model provenance, measurement validity, behavioral fingerprinting, black-box attribution, post-training branches, degenerate separators

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vuVJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether output text alone can distinguish base models from instruction-tuned models within the same family and scale. Using 16 text statistics and logistic regression, the authors obtain strong within-scale classification performance on Qwen3. Through length controls, question pairing, prefix truncation, and feature-group analysis, they show that cross-scale performance depends substantially on the evaluation protocol. The main contribution is to demonstrate that high black-box lineage classification scores may largely reflect differences in output termination and length, and should therefore be interpreted in the context of the measurement conditions.

Critical strengths

The problem is practically relevant. The paper examines overlooked confounders in black-box model auditing and cautions against interpreting high classification accuracy as evidence of a stable lineage fingerprint.

The control experiments are systematic. Section 3.4 progressively controls output length, question identity, and visible text. Section 4.3 extends the analysis to five model scales and 20 directed transfer pairs, helping assess the stability of the initial findings.

The negative findings are presented cautiously. The authors distinguish performance degradation from evidence that no signal exists. They also disclose problems in earlier attack construction, feature extraction, and evaluation procedures, demonstrating useful transparency.

Cross-family checks provide additional value.

Critical weaknesses

The controls introduce sample-selection effects. Different control conditions retain different subsets of examples, making it difficult to separate the effects of removing confounders from changes in sample composition. Comparisons on a shared subset would strengthen the conclusions.

The transfer results remain statistically inconclusive. Performance deteriorates under stricter controls, but substantial uncertainty remains. The evidence supports sensitivity to evaluation conditions more clearly than the absence of transferable lineage information.

The mechanistic interpretation is not fully established. The feature-group analysis does not clearly demonstrate that termination-related features have a unique explanatory role. The paper should distinguish predictive associations from evidence about the mechanism underlying classification.

The attacks cover a limited threat model. Most attacks preserve the output length and termination behavior identified as important. Their failure therefore provides limited evidence of robustness against interventions that directly target these channels.

Key validation details need clarification. The main text should explain whether data splitting prevents overlap between outputs for the same question and whether confidence estimation accounts for their dependence. These choices affect the reliability of the reported results.

Questions for authors

Can the existing results in Table 1 be reported on the same final subset of questions across conditions, separating the effects of sample selection from those of the controls? If this analysis already exists, please identify its location and key results.

Is within-scale cross-validation grouped by question ID, ensuring that outputs or generation variants for the same question cannot appear in different folds? What is the resampling unit used for bootstrapping?

Given the wide confidence intervals under strict controls and the fact that none of the 20 transfer pairs satisfies H1, what level of useful transfer performance can the current evidence rule out?

Since the random-subset comparison does not establish the special role of the degenerate feature group, what independent evidence supports the claim that termination dominates? Which conclusions should be interpreted only as predictive associations?

Do any existing attacks change output length or termination behavior? If not, would the authors agree to limit the attack conclusion to robustness against surface perturbations that preserve these channels?

Reviewer_LKCsOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the robustness of black box LLM classification method. It proposes to study the impact of input feature design compared to attacks on the text used to produce these features and their impact on classification accuracy. The findings show that the feature choice impacts more the classifier accuracy than any form of attack on these features.

Critical strengths
  • The paper attempts at answering a very interesting and important question: where is the robustness in black box fingerprinting techniques?
  • It seems to be vary precise about what it claims and what they does not claim
Critical weaknesses
  • The paper is extremely hard to read. The phrasing is really odd and particularly hard to follow. Most paragraphs needs to be rewritten in a clearer English because I am afraid most researchers will not be capable of understanding it easily. I spent few hours trying to understand what the authors did and I am still unsure about what some sections mean.

Example from 5.2

Within a scale, at a fixed length, a residual lexical signal separates the branches at 0.719 and 0.696, so a reading that does not control this channel measures the channel.
Questions for authors
  • Controls are added additively (Figure 2 - first length control, then same problems, then length free features etc..). What about testing each of them independently to see if the impact of controls is linked to adding them together or from a single one of them?
  • Where do these 16 features come from and how the factual results shown in the paper relate to the state of the art of LLM fingerprinting? In other words, why adding these specific controls in this specific order is informative about the difficulties in state of the art LLM fingerprinting? Most of the citations are pushed to appendix A, and only 2 of them are included in the main text which makes it very hard for readers to situate the paper in its context and to understand implementation choices supporting the claims of the paper.
  • Nowadays some fingerprinting techniques purely based on model outputs use more complex or automated features for fingerprinting (ex. PhyloLM [1], LLM DNA [2]), how do the paper's findings generalise to these complex/automated features?

[1] Yax, N., Oudeyer, P.-Y., & Palminteri, S. (2025). PhyloLM: Inferring the phylogeny of large language models and predicting their performances in benchmarks. In International Conference on Learning Representations (ICLR 2025). https://doi.org/10.48550/arXiv.2404.04671

[2] Wu, Z., Zhao, H., Wang, Z., Guo, J., Wang, Q., & He, B. (2026). LLM DNA: Tracing model evolution via functional representations. In International Conference on Learning Representations (ICLR 2026). https://doi.org/10.48550/arXiv.2509.24496

Reviewer_51YKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper mainly studies a simple classifier that distinguishes Qwen3 base and instruction-tuned models based on their generated texts, ablates the settings under which it succeeds or fails, and tries to draw broader conclusions about black-box lineage fingerprinting. I have concerns with almost every step of its arguments.

Critical strengths

None

While the authors claim to have used AI only for "editing the manuscript for readability," and not for "the drafting of the main body of the paper," I find the paper hard to follow, AI-written (even autoresearch) in style, and lacking a clear logical flow. I also noticed sentences like "our earlier version reported...", "a first version lacked", and "Earlier scripts in this line used 17 features", which look like obvious traces of an AI-generated paper.

Critical weaknesses

The paper mainly studies a simple classifier that distinguishes Qwen3 base and instruction-tuned models based on their generated texts, ablates the settings under which it succeeds or fails, and tries to draw broader conclusions about black-box lineage fingerprinting. I have concerns with almost every step of these arguments.

W1. Unclear relevance to lineage fingerprinting. Being able to distinguish Qwen3 Base/IT on fixed prompts and tasks doesn't necessarily make this a useful lineage fingerprinting method. The authors should first validate it in more realistic settings (e.g., downstream deployed models, quantization, prolonged post-training models) to prove that this is a REAL "fingerprinting method".

W2. Rather brittle classifier. Just 16 output statistics + logistic regression. Not surprising that such a simple classifier is sensitive to different settings (e.g., model scale in Sec. 4.1–4.3). I don't see why its generalization limitations are particularly interesting in this specific setup (and back to W1)

W3. Overclaiming. The authors conduct many ablations on this simple classifier but draw broader conclusions about black-box fingerprinting. The claims should be validated first on existing lineage fingerprinting methods, with all these findings and results able to be reproduced.

W4. Unclear significance of the findings. Length and termination are real behavioral differences between models, so why treat them as confounders rather than useful fingerprinting signals? Also, Mistral and OLMo show very different patterns: if all findings, eg the dominance of length/termination features, seem Qwen3-specific, what's the value of these ablation/analysis?

W5. Some ablations are not well controlled. Table 1 changes both the feature set (6 to 4) and truncation between rows 3‘ and 4, but attributes the drop to truncation alone. The original baseline also uses different features and scaler fitting.

Questions for authors

n/a

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.