acceptodds
Under review as a conference paper at ICLR 2027

Functional Network Fingerprint: A Unified Method for Cross-Architecture and Cross-Structure LLM Lineage Verification

Abstract

Training large language models (LLMs) is extremely expensive, yet released weights can be copied and re-released by third parties. We study a subtle disguise, which we call weight repackaging: an attacker copies a victim model's weights and then alters its structure—adding, removing, or interpolating layers, or expanding the hidden or MLP width—before continuing training. The resulting model differs from the victim in depth, width, and tokenizer while inheriting most of its parameters, making fingerprints that rely on direct weight or representation correspondence difficult to apply. We propose the Functional Network Fingerprint (FNF), a white-box method that requires no fingerprint-specific training and verifies whether a suspect model is derived from a victim model. Inspired by brain-fingerprinting research, FNF applies canonical independent component analysis (CanICA) to internal activations of Transformer blocks to identify co-activating neuron groups, and compares their activity trajectories between two models under identical input stimuli. Because FNF compares functional activity rather than weights or static representations, it does not require direct correspondence in depth, width, or tokenizer, and can operate across architectures and under common weight transformations. On a labeled set of model pairs spanning three provenance categories, FNF separates derived from independent models with an AUC of and identifies every derived model at zero false positives. On the pairs to which two non-invasive baselines also apply, the representation-based method REEF reaches an AUC of and the recent attention-difference method AttnDiff , and at zero false positives they recover only and of the related pairs. FNF thus provides a simple and interpretable tool for LLM provenance verification.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fingerprint, Large Language Models

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_byZkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work is about Functional Network Fingerprint (FNF). It is a training-free, white-box test of if a suspect LLM was derived from a victim LLM. Nothing is matched between the two models, so the method is supposed to work when depth, width or tokenizer differ. The authors propose CanICA, which extracts spatial maps and with a threshold, each map is filtered to form core neurons. The idea seems inspired from the causality perspective as well, since they try to find neurons that are causally responsible for certain task in the architecture. The authors frame the work as a “weight repackaging” attack, for example SOLAR-10.7B built from Mistral-7B by depth up-scaling.

Critical strengths
  • The method is simple, needs no correspondence between models.

  • Paper is clearly written and easy to follow

  • I think the idea is novel, and the inspiration from neuroscience makes sense and also the proposal of CanICA and its corresponding formulation is correct.

  • The experimental setup looks correct with no concerns around the evaluation scheme.

Critical weaknesses
  • The main claim of the paper is not sufficiently tested. The paper presents "weight repackaging" as changing depth, width and tokenizer, but all three repackaged pairs (Mistral-7B, SOLAR-10.7B, Mistral-11B) change depth only and use the same tokenizer. No derived model with expanded width, interpolated layers or a new tokenizer is evaluated.

  • I feel the whole work is a bit empirical and done under specific setting. I worry if the idea is generalizable, and whether we could have similar analysis on bigger, more complex models.

  • To justify that the FNF is not tuned to their own evaluation scenario and benchmark, is there any other setting where FNF can be evaluate for example LeaFBench?

  • I believe that the robustness is evaluated against the transformations that the FNF is invariant to by design. FNF is not affected by permutation and scaling by construction. What if there is a function-preserving orthogonal rotation that keeps outputs identical and likely breaks the neuron-aligned structure that spatial ICA relies on, while CKA-based REEF is unaffected by it. No such scenario is evaluated and I wonder if it would generalize

Questions for authors

I have asked my questions and raised my concerns as part of the weakness section.

Suggestion: Although the score is not affect by this, but I would recommend authors to make the plots better, if it is accepted eventually. The plot labels and legends are a bit hard to read. Also, even though authors acknowledged the use of AI for paper drafting it would be better if they could remove some of the many "--" double dashes from the paper, if it is avoidable.

1 comment

First of all, sorry for the late reply and thanks for your patience.

I appreciate the efforts of the authors and the justification provided during the rebuttal. Honestly, I am still not totally convinced about certain aspects of the work, especially regarding the scalability and practicality. Most of the work is done on small model setup.

That being said, I am taking nothing away from the methodology proposed and efforts taken by the authors to justify their research. I also understand the resource and other practical constraints.

The only reason I would think about increasing the scores is because of the efforts, which I feel is not fair, since I would have to take the decision based on the soundness of the work. Hence, I would like to maintain my rating.

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

This paper proposes Functional Network Fingerprint (FNF), a white-box method for verifying LLM lineage across architectural changes. The method uses CanICA to extract functional neuron groups from Transformer activations and compares their activity trajectories using Spearman correlation, without requiring direct structural correspondence. Experiments on fine-tuned, pruned, merged, and weight-repackaged models demonstrate promising detection performance, achieving an AUC of 1.00 on a benchmark of 40 model pairs.

Critical strengths

1.The paper addresses the practical challenge of verifying LLM lineage after structural changes, where direct weight and representation comparisons become difficult.

2.The proposed FNF method uses CanICA to identify functional neuron groups and compares their activity trajectories without requiring explicit alignment of model layers or dimensions.

3.Experiments covering fine-tuning, pruning, model merging, and weight repackaging demonstrate promising verification performance, achieving an AUC of 1.00 on a benchmark of 40 model pairs.

Critical weaknesses
  1. The definition of model lineage needs further clarification. The evaluation groups verified derivation and same-family relationships under the broader category of related models, although sharing a common ancestor is not equivalent to direct inheritance from a specific victim. This makes the exact scope of lineage verification somewhat unclear.
  2. Insufficient discussion of closely related work. The paper does not adequately discuss recent studies on LLM lineage verification. In particular, GhostSpec [1] has already investigated model provenance verification under architectural transformations, including pruning and model expansion, which substantially overlap with the weight-repackaging scenario emphasized in this paper. The lack of discussion and comparison with such closely related work makes it difficult to assess the novelty and advantages of the proposed method.
  3. The cross-tokenizer generalizability remains insufficiently evaluated. FNF relies on resampling activation trajectories to align models with different tokenizers, but the experiments provide limited analysis of varying tokenizer mismatches and challenging unrelated pairs. More systematic evidence is needed to establish robustness across substantially different tokenization schemes.
  4. The reliability of threshold calibration on unseen models remains uncertain. Although the reported AUC results are promising, the evaluation uses a limited set of related model pairs, and the paper acknowledges that FNF scores can vary with model scale. Testing a fixed, independently calibrated threshold on unseen model families would better support its practical applicability.

[1] Wang, S., Ma, Z., Xinyi, L., & Li, Z. (2026). Ghost in the Transformer: Detecting Model Reuse with Invariant Spectral Signatures. Proceedings of the AAAI Conference on Artificial Intelligence, 40(40), 33648–33656.

Questions for authors

See Weaknesses

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

This paper proposes FNF, a white-box LLM lineage verification method targeting weight repackaging. It uses CanICA to extract functional networks from hidden states and compares their activity trajectories using Spearman correlation. Experiments cover various model modifications and include some meaningful ablations and discussions. However, the evaluation of weight repackaging is limited to a few (eg 2-3) models, and the threshold relies on calibration.

Critical strengths
  1. Model lineage verification after structural modifications (e.g., changes in network depth, width, or even tokenizer) is a challenging and truly meaningful problem in this domain. This setup also has enough practical relevance.

  2. Intuitive and straightforward method. The method uses CanICA to extract functional networks and then compares their activity time courses. The overall pipeline is relatively clear and does not require modifying the original model, which is also a practical advantage.

  3. The experiments cover various model modifications and include some meaningful ablations and discussions. The authors test depth expansion, pruning, model merging, parameter permutation, and scaling transformations, and discuss the effects of input distribution, component count, etc.

Critical weaknesses

My main concern is about the effectiveness and reliability of the proposed method.

  1. Some cases do not seem to perform well. eg, Table 3 claims successful detection, but the FNF score is only ~0.57. In Table 23, the FNF scores for Qwen2.5-7B-IT vs. Qwen1.5-7B-IT and Qwen2.5-1.5B-IT are only 0.40and 0.47, respectively. The margin between the lowest positive score and the highest negative score (0.3757) is only 0.0286.

  2. I would therefore suggest more experiments, especially for weight repackaging, which is the main motivation of this paper. Table 3 only includes three pairs from the same Mistral lineage, two of which are direct derivation pairs. Although the authors claim that FNF can handle changes in depth, width, and even tokenizer, there is a lack of formal evaluation on width expansion and tokenizer modifications (!!) across different model families. Most experiments (e.g., pruning and merging) are case studies on only 2–3 models, and mainly report similarity scores rather than actual detection performance.

  3. Another relevant question is how the threshold τ=0.42\tau=0.42 is selected. The authors use this threshold in Appendix B, but do not clearly explain how it is determined. Is it selected by optimizing the TPR-FPR trade-off on the 40 model pairs? If so, this could be tricky and may affect the practical effectiveness of the method. The method seems quite sensitive to the threshold (e.g., using 0.42 would already miss one related pair in Table 23), and seems not able to determine whether any two models share a lineage without calibrating a threshold.

Questions for authors

Is FNF less effective for checkpoints from the early stages of pretraining? In Sec 6.3 and Appendix C, the derivation pairs involving the step-1,000 checkpoint have FNF scores of only 0.38–0.54, which substantially overlap with independently trained same-family pairs (0.29–0.51).

My intuition is that FNF may rely on stable functional organization formed during training. So it would be worth testing this on models with much longer training (eg, OLMo 3 series). Whether FNF works or not would be meaningful.

Comments on the paper

2 comments

Thanks to the authors for the additional experiments and clarifications. Several points are now clear to me:

  1. Threshold. The threshold is tuned directly on the evaluated pairs. This is somewhat tricky, but I did not penalize it, since many similarity/correlation-based fingerprinting methods rely on such a calibrated threshold that does not generalize. It is still worth discussing in the limitations (at least this is a limitation to me, because it means that, given an arbitrary pair of models, the method cannot tell whether they share a lineage).

  2. Long post-training. I am satisfied with the OLMo-3, CodeLlama and Llemma experiments.

  3. Width expansion. I'm also satisfied with the width-expanded variant experiments, please add them into the revision, they're important.

  4. Early pretraining inconsistency. The reliance on converged behavior, and hence the limited applicability to the pretraining stage, is also worth discussing in the limitations; this is a known limitation of the baseline compared here, REEF, as well.

However, I am now more confused on the first point.

I missed this at first: the paper seems to mix up "same family" and "same lineage". All Qwen models are the same family, sure, but across versions and scales we normally do not consider them to share a lineage. So what is the basis for labeling Qwen2.5-7B-IT / Qwen1.5-7B-IT and Qwen2.5-7B-IT / Qwen2.5-1.5B-IT as related? Is there any evidence of weight inheritance?

My original concern (1) therefore remains: derived and same-family models do not seem to be distinguishable. Same-family pairs often get a higher FNF than derived ones, e.g. the Pythia step-1,000 derived pairs are only 0.38–0.54, which overlaps with the same-family range (0.29–0.51). And the negatives used in existing evaluations (independent pairs) are all easy ones. If only strictly derived pairs are counted as positives, and same-family and independent are both negatives, what are the AUC and the TPR at zero false positives? Could the aligned-network count you added help here (i.e., differentiate same-lineage from same-family)?

The only pairs that really matter for the claim of this paper seem to be Mistral-7B / Solar-10.7B and Mistral-7B / Mistral-11B (correct me if I'm wrong). Again, weight repackaging is really a strong selling point in the fingerprinting literature, and I can neglect all the weaknesses mentioned above for that. Might be worth trying to find more real examples, or showing cases similar to the width-expanded variants you constructed yourselves (with some continued training after the weight repackaging).

A few minor things:

  1. Is LLaMA2 here the chat or the base model? Table 23 writes "LLaMA2-chat" and marks the pairs as Deriv., but CodeLlama and Vicuna-v1.5 are trained from the LLaMA2 base.

  2. One wrong label: CodeLlama / Llemma is marked as Fam., but Llemma is initialized from Code Llama 7B.

  3. Tokenizer: I can buy the argument that cross-tokenizer comparison is supported by design. But the added experiment is not very convincing: isn't CodeLlama still using LLaMA2's tokenizer, with just 4 extra infilling special tokens? More different tokenizers might be worth testing. This one is less critical to me now though.

Thanks to the authors for the very detailed experiments and responses. This is honestly a hard decision for me. I can see the effort put into the rebuttal, and the width expansion + continued training results are what I asked for, so I'm willing to raise my score, even though quite a few of my concerns turned out to be real weaknesses. I'm not asking for any more experiments, just listing where things stand so they can be stated explicitly in the revision.

Resolved:

  • Long post-training (OLMo-3, CodeLlama, Llemma) and width expansion, including the continued-training variants.
  • Label corrections (LLaMA2 base vs. chat, CodeLlama / Llemma).

Remaining issues or real weaknesses:

  • Width expansion + continued training is only shown on small models (Pythia-410M / 1B) with 2k / 4k steps, and the signal is decaying (aligned networks 43$\rightarrow$9, 30$\rightarrow$3). Longer training is unknown, and real-world cases are still only the Mistral ones.
  • Same-family (but not necessarily related) models cannot be distinguished from truly same-lineage ones, and the aligned-network count does not help.
  • Under the strict evaluation, TPR@0FPR is only 3/8, which is low for a fingerprinting method.
  • Tokenizer: the evidence only covers near-identical tokenizers, so no general claim can be made.

Acceptable limitations:

  • The threshold is tuned on the evaluated pairs and does not generalize, so the method cannot decide on an arbitrary pair of models.
  • Reliance on converged behavior, hence limited applicability to early pretraining (same for REEF).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.