Pith. sign in

REVIEW 5 major objections 4 minor 40 references

Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space

T0 review · 5 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that the lineage of open-weight LLMs — independent, same-series, or shared-base — can be read off from weight geometry alone, using spectral energy for coarse separation and subspace alignment for fine discrimination.

desk verdict Promising fine-grained subspace-alignment results, but the headline coarse-grained AUC rests on 37 distinct pairs padded to '110+' by near-duplicate noise, and an unvalidated layer-interpolation assumption. read the letter →

arxiv 2608.07786 v1 pith:JAKRLAFO submitted 2026-08-07 cs.AI cs.LG

Yiwei Chen , Bingqi Shang , Sijia Liu This is my paper · ORCID
classification cs.AIcs.LG
keywords LLMlineageweight-spacefingerprintsspectralenergysingularvaluedecompositionsubspacealignmentprincipalanglesmodelprovenanceshared-basemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that open-weight LLMs carry intrinsic fingerprints in their weight matrices, so that two models' relationship can be inferred without any training data or generated outputs. It defines a three-level hierarchy: independent-origin, same-series, and shared-base. It shows that spectral energy, computed cheaply as per-layer trace-based singular-value sums, separates the coarse levels, while subspace alignment, measured through principal angles between top singular-vector subspaces, separates fine variations inside the shared-base level, including dataset scale and post-training algorithm. If correct, model provenance, governance, and supply-chain auditing can be done passively from released weights.

What carries the argument

The machinery is a two-signal geometric fingerprint built from singular value decomposition of each weight matrix $W_c^{(l)}$. The first signal, spectral energy, is the Frobenius-norm-like trace $\sqrt{\mathrm{tr}((W_c^{(l)})^\top W_c^{(l)})}$, summarized as a per-layer vector and compared by correlation after interpolation; it encodes global magnitude patterns that separate families and scales. The second signal, subspace alignment, compares top-$k$ left singular vector subspaces through principal angles, aggregated over the least-aligned layers and components; it captures the directional rotations that post-training induces without changing the spectrum. Together they form the unified fingerprint that the paper uses for hierarchical lineage tracing.

What would settle it

Train two models with identical architecture, tokenizer, and training data but independent random initializations, then compute their spectral trace correlation. If the score lands above the same-series threshold, spectral energy is encoding generic architecture rather than shared lineage; if it stays near independent-origin levels, the coarse lineage reading is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that weight-space geometry provides a reliable and interpretable signal for LLM lineage, enabling both coarse-grained regime separation and fine-grained discrimination within shared-base models. The paper proposes two complementary fingerprints: the spectral trace fingerprint, $\tau_c(\theta) = (t(W_c^{(1)}), \ldots, t(W_c^{(L)}))$ with $t(W_c^{(l)}) = \sqrt{\mathrm{tr}((W_c^{(l)})^\top W_c^{(l)})} = \sqrt{\sum_i \sigma_i^2}$, which is aligned across different depths by linear interpolation over a normalized layer index and compared via Pearson correlation; and subspace alignment, which takes the singular values of the cross-subspace matrix $C_c^{(l)} = (U_{c,k}^{(l)}(\theta_a))^\top U_{c,k}^{(l)}(\theta_b)$ as cosines of principal angles and averages the $J$ smallest ones. On over 110 open-weight LLM pairs, spectral trace separates independent-origin from same-series pairs with AUC 0.850, versus 0.256 for AWM and 0.655 for PDF, and perfectly separates same-series from shared-base pairs. Within the shared-base regime, subspace alignment orders Alpaca-SFT variants by data scale (0.976 at 10% down to 0.889 at 100%) and gives distinct per-component signatures to DPO, PPO, and RAFT post-training.

Load-bearing premise

The load-bearing assumption is that linear interpolation over a normalized layer index makes layers at the same relative depth functionally comparable across different architectures and depths; if that mapping fails, the Trace separation of independent-origin from same-series pairs loses its grounding.

Editorial extensions

If this is right

  • Lineage categories can be inferred from weights alone, without accessing training data, prompts, or model outputs.
  • Spectral trace gives a cheap prefilter: independent-origin versus same-series pairs are separable at AUC 0.850, while standard baselines fail.
  • Subspace alignment makes data-scale effects visible inside the shared-base regime, so one can rank post-training variants by how much weight geometry moved.
  • Different post-training algorithms leave distinctive per-component signatures, with the Q projection showing the largest method-dependent variation.
  • The coarse shared-base signal survives quantization, pruning, merging, and distillation, while subspace alignment tracks transformation severity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The three regimes are likely anchors of a continuous relatedness spectrum: the paper's cross-series Qwen2.5/Qwen3 mean of 0.479 sits between independent-origin and same-series means, suggesting any pair of open-weight models could be placed on the same graded scale.
  • Because subspace alignment is computed per component, it could serve as a diagnostic for which parts of a network a post-training method rewrites, potentially linking alignment drops to functional changes.
  • A natural testable extension is to separate data-sharing from code-sharing lineage: same architecture and tokenizer but different training data should produce different spectral-trace behavior than same data with different architecture.
  • The cheap trace fingerprint could be used as a screening stage, reserving the more expensive subspace-alignment computation for pairs that pass a coarse similarity threshold.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes a white-box, weight-only method for tracing LLM lineage. It defines three regimes of increasing shared information—independent-origin (S1), same-series (S2), and shared-base (S3)—and introduces two complementary geometric fingerprints: a spectral trace fingerprint (Eqs. 2–5) derived from layer-wise sums of squared singular values, and a subspace-alignment fingerprint (Eqs. 6–8) based on the smallest principal angles between top-k singular-vector subspaces. The spectral trace is claimed to separate S1 from S2 (AUC 0.850, Fig. 6), while subspace alignment is claimed to distinguish shared-base variants differing in data scale and post-training algorithm (Table 2, Fig. 9). The paper reports experiments on 'over 110' model pairs, including Gaussian-noise-perturbed duplicates, and concludes that weight-space geometry provides a robust and interpretable lineage signal.

Significance. The conceptual framing—treating spectral magnitude and subspace direction as complementary lineage signals—is a useful step beyond existing static fingerprints such as AWM, HuReF, and PDF, and the proposed metrics are simple, computationally cheap, and parameter-light (k=256, J=3, K_layer=3). The controlled Alpaca data-scale ablation and the RLHFlow/Tulu variant comparisons are well designed and, if validated, would make the fine-grained results genuinely useful for model governance and provenance auditing. The main limitation is statistical: the headline coarse-grained result rests on 37 distinct pairs, the noise-augmented '110+ pairs' are not independent replicates, and the layer-alignment assumption in Eq. (4) is unvalidated. These issues are fixable with additional analysis and reporting, so the underlying approach remains promising.

major comments (5)
  1. [§4, Fig. 6, Appendix B/D] The S1-versus-S2 AUC of 0.850 is computed on only 27 independent-origin and 10 same-series distinct model pairs; the claimed 'over 110 pairs' is reached by adding Gaussian noise (sigma=1e-4) to every model, which nearly doubles the dataset with perturbed duplicates. These augmented pairs are not independent replicates, so the AUC and the robustness claim do not follow from the reported sample. Moreover, the perturbation is not harmless in all rows: Table A2 shows Trace for Gemma-2B vs Gemma-7B changing from 0.6169 to 0.0298 and for Mistral-7B-v0.3 vs Mixtral-8x7B from 0.8679 to 0.4265 under sigma=1e-4. This contradicts the statement in Appendix D that Trace 'remains stable under perturbation.' Please report the AUC with confidence intervals on the 37 distinct pairs, and present the Gaussian perturbation as a separate stability analysis rather than as additional samples.
  2. [Eq. (4), §4] The layer-alignment step in Eq. (4) linearly interpolates trace fingerprints over a normalized layer index, implicitly assuming that layer position as a fraction of total depth corresponds to functional similarity across architectures of different depths. This assumption is load-bearing for the coarse-grained separation (Finding 1), since many S2 pairs (e.g., Pythia-1B vs 1.4B, LLaMA-2-7B vs 13B) have different depths and the S1 set includes models with 28, 32, and 40 layers. No validation of this alignment is provided, and no alternative alignment (e.g., comparing only common-depth layers, or aligning by block type and distance from input/output) is tested. Please add an ablation or justification for why normalized-layer interpolation is appropriate, or show that the S1-vs-S2 result is insensitive to the alignment choice.
  3. [Appendix B.2, §3 definition of S2] The same-series regime is defined in §3 as pairs that 'share architectures and training pipelines, but differ in scale,' yet the S2 set in Appendix B.2 includes Mistral-7B-v0.3 vs Mixtral-8x7B-v0.1, which is a dense-versus-MoE pair, and Falcon-7B vs Falcon-11B, whose architectural differences are not documented. This inconsistency directly affects the measured S1-vs-S2 separation, because the MoE pair has the highest same-series Trace score (0.8679 in Table A2) and may inflate or distort the distribution. Please either exclude these pairs from the main S2 evaluation or analyze them as a separate intermediate category, as is already done for Qwen2.5-vs-Qwen3 in Appendix F.
  4. [Reproducibility statement, header] The paper's reproducibility claim is not fully met: the GitHub link in the header is malformed ('/githubGitHub/gl⌢beProject') and no working code repository is provided. Given that the paper's contribution is an empirical fingerprinting method with specific aggregation parameters (k=256, J=3, K_layer=3) and a particular interpolation rule, a public implementation is important for the community to verify the reported AUC and the subspace-alignment scores and to apply the method to new models.
  5. [§5, Table A7, Appendix K] The fine-grained shared-base results, while suggestive, are reported without any uncertainty quantification. For example, Table A7 lists a single subspace-alignment score per pair, and the observed differences (e.g., Qwen3-4B vs Qwen3-4B-Thinking at 0.184 versus LLaMA-3.1-8B vs Alpaca-10% at 0.976) are presented as exact. Since these scores are computed from a single model checkpoint per pair, there is no estimate of variability from weight initialization or sampling. Please provide repeated-measure or bootstrap-based error bars, or at least report the per-component ranges, so the reader can judge whether the claimed data-scale and algorithm-specific effects are larger than the noise in the measurement procedure.
minor comments (4)
  1. [Abstract and §4] The abstract and contribution list state 'over 110 diverse open-weight LLM pairs,' but the actual number of distinct pairs in the three regimes is 57, with the remainder being Gaussian-perturbed copies. Please state the distinct-pair count explicitly in the main text to avoid overstating the diversity of the evaluation.
  2. [Fig. 5] The caption says the shaded regions have 'boundaries defined by the mean trace scores,' but it is not clear whether these boundaries are the mean of the scenario distribution or of something else. Please define the shading rule precisely in the caption or in Appendix D.
  3. [§5, Algorithm A2] The main text describes aggregating 'the three lowest-similarity layers' for each component, and Algorithm A2 uses K_layer for the same purpose; please define K_layer=3 explicitly in the text near Eq. (8) rather than only in the algorithm pseudocode.
  4. [Appendix I, Table A6] The sensitivity analysis for J shows that scores increase monotonically with J, and the paper states that relative differences remain consistent, but no threshold or decision rule is derived from these scores. Since the paper does not provide a calibrated classifier for the shared-base regime, please clarify whether the fine-grained claims are intended as descriptive findings or as an operational discrimination method.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; one minor self-referential visualization in Fig. 5 does not affect the main AUC results.

  1. self definitional [Fig. 5 caption, Section 4 (Coarse-Grained Lineage Discrimination via Spectral Energy)]
    "shaded regions on the y-axis indicate desired scenario-specific score ranges, with boundaries defined by the mean trace scores."

    The shaded regime boundaries are set to the mean Trace scores of the same model pairs whose separation the figure is used to demonstrate. Two groups with different means will always produce separated mean-centered intervals, so the visual 'hierarchy' in Fig. 5 is partly guaranteed by construction. This is a presentation artifact rather than a fitted prediction: the actual S1-vs-S2 discrimination is quantified by the ROC/AUC in Fig. 6 and Table A4, which does not use these mean-defined boundaries, and the Trace metric itself has no fitted parameters. The step is therefore minor and not load-bearing for the central claim.

full rationale

The derivation chain for the Trace and subspace-alignment similarities is self-contained: Eq. (2) defines trace-based spectral energy from SVD singular values, Eqs. (3)-(5) define layer-wise fingerprints and Pearson-correlation pooling with no learned parameters, and Eqs. (6)-(8) define subspace alignment from principal angles with fixed k=256, J=3, and K_layer=3. Lineage labels come from public model documentation, not from the metric, so the S1/S2/S3 ordering is an empirical finding rather than an input. The cited self-references (Chen et al. 2026a,b; Shang et al. 2025) are background citations and carry no load-bearing argument. The honest caveats in the paper are statistical, not circular: the S1-vs-S2 AUC rests on 27 independent-origin and 10 same-series distinct pairs, the Gaussian augmentation mostly reuses near-duplicate scores, and Eq. (4)'s layer interpolation is unvalidated; these are robustness concerns, not equivalences-to-inputs. Overall, the central claim that weight-space geometry separates lineage regimes is not forced by construction, so circularity is minimal.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on modeling assumptions about layer correspondence, subspace choice, and aggregation hyperparameters, but no parameters are fitted to lineage labels. The empirical evaluation uses known lineage labels only to measure separation.

free parameters (4)
  • k (top singular vectors) = 256
    Fixed number of leading left singular vectors used to define subspaces in Eq. (6). No sensitivity analysis is reported for k.
  • J (smallest singular values averaged) = 3
    Number of least-aligned directions averaged in Eq. (8). Sensitivity analysis is reported in Table A6.
  • K_layer (lowest-similarity layers averaged) = 3
    Per-component aggregation averages the three layers with lowest similarity. No sensitivity analysis is reported for this choice.
  • Gaussian noise sigma = 0.0001
    Noise standard deviation used to create synthetic duplicate pairs, doubling the pair count. The value is chosen to preserve functionality but is not varied.
assumptions (4)
  • domain assumption Linear interpolation aligns layers across different depths and architectures for Pearson correlation of trace profiles.
    Eq. (4) resamples layer-wise fingerprints to a common length. This is valid only if normalized layer index preserves functional correspondence across models.
  • domain assumption Top-k left singular subspaces of weight matrices carry lineage-relevant directional information.
    Eqs. (6)-(8) define subspace alignment on top-k left singular vectors, supported mainly by visualizations in Fig. 8 and the appendix.
  • domain assumption The three regime labels (independent-origin, same-series, shared-base) are ground truth for the model zoo.
    Appendix B assigns labels from public documentation. Labeling errors would bias all AUC and separation results.
  • standard math Standard SVD and principal-angle properties used in Eqs. (7)-(8).
    Singular values of a cross-subspace matrix are cosines of principal angles between subspaces, a standard linear algebra fact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space." pith.science (2026). https://pith.science/paper/JAKRLAFO

@misc{pith2026260807786,
  author       = {Pith},
  title        = {Pith review of: Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAKRLAFO}},
  note         = {Machine review of arXiv:2608.07786}
}
read the original abstract

Open-weight large language models (LLMs) are increasingly developed through complex, multi-stage pipelines, leading to intricate lineage relationships that reflect model origin, ownership, and evolution. Understanding these relationships is important for model provenance, governance, and supply-chain integrity. In this work, we investigate the notion of LLM "biometrics" (analogous to human biometrics) to ask whether LLMs exhibit intrinsic fingerprints in weight space alone, without access to input data, that reveal their origin and lineage. We formulate this as a lineage discrimination problem, distinguishing among independent-origin, same-series, and shared-base models. To characterize these relationships, we propose a unified geometric fingerprinting framework that analyzes weight matrices from two complementary perspectives: (i) spectral energy, captured by singular value distributions to encode global magnitude patterns, and (ii) subspace alignment, quantified via subspace deviations to capture directional geometry. Our analysis uncovers a clear hierarchy of structural similarity in weight space: spectral energy reliably distinguishes independently trained models and different model families, while subspace alignment enables fine-grained discrimination among closely related models, including variations in dataset scale and post-training procedures. Extensive experiments on over 110 diverse open-weight LLM pairs demonstrate that weight-space geometry provides a robust and interpretable signal for model lineage, enabling coarse-grained regime separation and fine-grained discrimination within shared-base models.

Figures

Figures reproduced from arXiv: 2608.07786 by the authors.

Figure 1
Figure 1. Overview of the proposed geometric fingerprinting framework for LLM lineage anal￾ysis. Given model weight matrices, we apply SVD to decompose them into two complementary representations: (i) spectral energy (Σ), which captures global magnitude patterns and enables coarse￾grained discrimination across model families and scales, and (ii) subspace alignment (U), which models directional relationships via subspace align… view at source ↗
Figure 2
Figure 2. Singular value distributions of middle-layer weight matrices for three component types c in (a)–(c) across models (top-256 singular values, accommodating varying matrix sizes). The x-axis denotes the singular value index k (descending), and the y-axis shows normalized singular values (scaled by the largest value). The middle layer is the ⌊L/2⌋-th layer of a model with L layers. Qwen2.5- 3B and Qwen2.5-7B are from th… view at source ↗
Figure 3
Figure 3. Normalized similarity scores in [0, 1] for model pairs under two relationship scenarios: independent-origin and same-series. Each point repre￾sents a model pair, positioned by AWM(y-axis) and PDF(x-axis) scores. Dashed lines denote the decision thresholds of AWM (horizontal) and PDF (vertical), separating regimes. Despite the clear spectral patterns in weight space, existing methods fail to capture these distinction… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Model relationships form three regimes with increasing degrees of shared information (left to right). Independent-origin: models are trained independently without shared initialization, data, or training pipeline. Same-series: models from the same series share architec…
Figure 5
Figure 5. Figure 5: Distribution of Trace-based similarity scores (5) (normalized to [0,1]) for distinguishing model relationships across three scenarios: (S1) independent-origin, (S2) same-series, and (S3) shared-base. For comparison, distributions of baseline methods (AWM, HuReF, PDF) a…
Figure 6
Figure 6. Figure 6: ROC curves for distinguishing independent￾origin and same-series model pairs, experiment setting is consistent with [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison between Trace and PDF on distinguishing independent-origin and same-series pairs. Experiment setup stay consistent with [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison between a base model (Qwen2.5-7B) and its post-trained variant (Qwen2.5- 7B-Instruct) using the attention projection Q at the middle transformer layer (14th), with top-k = 256 singular components. (a) Spectral magnitude: normalized singular values vs. index.…
Figure 9
Figure 9. Figure 9: Per-component subspace alignment across post-training strategies with respect to the base model Qwen2.5-Math-7B. Per-component subspace alignment is defined as the average of the three smallest layer-wise similarity scores (8) across all layers, capturing worst-case di…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 12 canonical work pages

  1. [1]

    Constitu- tional ai: Harmlessness from ai feedback.arXiv preprint arXiv:2212.08073,

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitu- tional ai: Harmlessness from ai feedback.arXiv preprint arXiv:2212.08073,

    arXiv

  2. [2]

    3: independent-origin (Scenario 1), same- series (Scenario 2), and shared-base (Scenario 3)

    15 Published as a conference paper at COLM 2026 B Model Zoo of Each Regime This section catalogs the model pairs used in our experiments, organized by the three lineage discrimination regimes defined in Sec. 3: independent-origin (Scenario 1), same- series (Scenario 2), and shared-base (Scenario 3). We reserve the same-series regime for scale variants wit...

    work page 2026

  3. [3]

    We distinguish two subcategories based on fine-grained variation sources

    Model pairs in this regime originate from a common pretrained base model and diverge through post-training procedures. We distinguish two subcategories based on fine-grained variation sources. Data-Scale Variants.These pairs share the same base model and post-training algorithm but differ in the size of the fine-tuning dataset: Standard instruction tuning...

    work page 2026

  4. [6]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    arXiv

  5. [7]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  6. [8]

    A2 summarizes the proposed Subspace Alignment algorithm

    22 Published as a conference paper at COLM 2026 H Spectral Directional-based Fingerprint Algorithm Alg. A2 summarizes the proposed Subspace Alignment algorithm. It computes layer-wise similarity from the J smallest singular values of the cross-subspace matrix, then aggregates the least-aligned layers and model components into an overall score. Algorithm A...

    work page arXiv 2026

  7. [9]

    Hide and seek: Fingerprinting large language models with evolutionary learning.arXiv preprint arXiv:2408.02871,

    Dmitri Iourovitski, Sanat Sharma, and Rakshak Talwar. Hide and seek: Fingerprinting large language models with evolutionary learning.arXiv preprint arXiv:2408.02871,

    arXiv

  8. [10]

    The universal weight subspace hypothesis.arXiv preprint arXiv:2512.05117,

    Prakhar Kaushik, Shravan Chaudhari, Ankit Vaidya, Rama Chellappa, and Alan Yuille. The universal weight subspace hypothesis.arXiv preprint arXiv:2512.05117,

    arXiv

Show all 40 references
  1. [11]

    Principal angles between subspaces in an a-based scalar product: algorithms and perturbation estimates.SIAM Journal on Scientific Computing, 23(6):2008–2040,

    Andrew V Knyazev and Merico E Argentati. Principal angles between subspaces in an a-based scalar product: algorithms and perturbation estimates.SIAM Journal on Scientific Computing, 23(6):2008–2040,

    2008

  2. [13]

    Tulu 3: Pushing frontiers in open language model post-training.arXiv preprint arXiv:2411.15124,

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. Tulu 3: Pushing frontiers in open language model post-training.arXiv preprint arXiv:2411.15124,

    arXiv

  3. [15]

    Training language models to follow instructions with human feedback.NeurIPS,

    12 Published as a conference paper at COLM 2026 Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.NeurIPS,

    2026

  4. [16]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    arXiv

  5. [18]

    Sok: Large language model copyright auditing via fingerprinting.arXiv preprint arXiv:2508.19843,

    Shuo Shao, Yiming Li, Yu He, Hongwei Yao, Wenyuan Yang, Dacheng Tao, and Zhan Qin. Sok: Large language model copyright auditing via fingerprinting.arXiv preprint arXiv:2508.19843,

  6. [19]

    Lora vs full fine-tuning: An illusion of equivalence.arXiv preprint arXiv:2410.21228,

    Reece Shuttleworth, Jacob Andreas, Antonio Torralba, and Pratyusha Sharma. Lora vs full fine-tuning: An illusion of equivalence.arXiv preprint arXiv:2410.21228,

  7. [20]

    Idiosyncrasies in large language models.arXiv preprint arXiv:2502.12150,

    Mingjie Sun, Yida Yin, Zhiqiu Xu, J Zico Kolter, and Zhuang Liu. Idiosyncrasies in large language models.arXiv preprint arXiv:2502.12150,

    arXiv

  8. [21]

    Natural fingerprints of large language models

    Teppei Suzuki, Ryokan Ri, and Sho Takase. Natural fingerprints of large language models. arXiv preprint arXiv:2504.14871,

  9. [22]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arxiv 2023.arXiv preprint arXiv:2302.13971, 10,

    2023 arXiv

  10. [23]

    Rofl: Robust fingerprinting of language models.arXiv preprint arXiv:2505.12682,

    Yun-Yun Tsai, Chuan Guo, Junfeng Yang, and Laurens van der Maaten. Rofl: Robust fingerprinting of language models.arXiv preprint arXiv:2505.12682,

    arXiv

  11. [24]

    Gradient-based model fingerprinting for llm similarity detection and family classification.arXiv preprint arXiv:2506.01631, 2025a

    Zehao Wu, Yanjie Zhao, and Haoyu Wang. Gradient-based model fingerprinting for llm similarity detection and family classification.arXiv preprint arXiv:2506.01631, 2025a. Zhaomin Wu, Haodong Zhao, Ziyang Wang, Jizhou Guo, Qian Wang, and Bingsheng He. Llm dna: Tracing model evol...

    2026 arXiv

  12. [25]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement.arXiv preprint arXiv:2409.12122,

    arXiv

  13. [26]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    arXiv

  14. [27]

    A fingerprint for large language models.arXiv preprint arXiv:2407.01235,

    Zhiguang Yang and Hanzhou Wu. A fingerprint for large language models.arXiv preprint arXiv:2407.01235,

  15. [28]

    Intrinsic fingerprint of llms: Continue training is not all you need to steal a model!arXiv preprint arXiv:2507.03014,

    Do-hyeon Yoon, Minsoo Chun, Thomas Allen, Hans Müller, Min Wang, and Rajesh Sharma. Intrinsic fingerprint of llms: Continue training is not all you need to steal a model!arXiv preprint arXiv:2507.03014,

    arXiv

  16. [29]

    Awm: Accurate weight- matrix fingerprint for large language models.arXiv preprint arXiv:2510.06738,

    Boyi Zeng, Lin Chen, Ziwei He, Xinbing Wang, and Zhouhan Lin. Awm: Accurate weight- matrix fingerprint for large language models.arXiv preprint arXiv:2510.06738,

  17. [30]

    Self: A robust singular value and eigenvalue approach for llm fingerprinting.arXiv preprint arXiv:2512.03620,

    Hanxiu Zhang and Yue Zheng. Self: A robust singular value and eigenvalue approach for llm fingerprinting.arXiv preprint arXiv:2512.03620,

  18. [31]

    Easydetector: Using linear probe to detect the provenance of large language models

    Jie Zhang, Jiayuan Li, Haiqiang Fei, Lun Li, and Hongsong Zhu. Easydetector: Using linear probe to detect the provenance of large language models. In2024 IEEE 23rd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom), pp. 2410–2417...

    arXiv

  19. [32]

    Tamm: Triadapter multi-modal learning for 3d shape understanding

    Zhihao Zhang, Shengcao Cao, and Yu-Xiong Wang. Tamm: Triadapter multi-modal learning for 3d shape understanding. InCVPR, 2024c. Zhihao Zhang, Abhinav Kumar, Girish Chandar Ganesan, and Xiaoming Liu. Unleashing the power of chain-of-prediction for monocular 3d object detection....

  20. [33]

    The path not taken: Rlvr provably learns off the principals.arXiv preprint arXiv:2511.08567,

    Hanqing Zhu, Zhenyu Zhang, Hanxian Huang, DiJia Su, Zechun Liu, Jiawei Zhao, Igor Fedorov, Hamed Pirsiavash, Zhizhou Sha, Jinwon Lee, et al. The path not taken: Rlvr provably learns off the principals.arXiv preprint arXiv:2511.08567,

  21. [34]

    Angles between subspaces and their tangents.arXiv preprint arXiv:1209.0523,

    Peizhen Zhu and Andrew V Knyazev. Angles between subspaces and their tangents.arXiv preprint arXiv:1209.0523,

    arXiv

  22. [37]

    The complete set comprises 27 model pairs across diverse model families: Qwen cross-family pairs (7 pairs): • Qwen/Qwen2.5-7B vs

    Model pairs in this regime originate from different organizations or training pipelines, with no shared base model. The complete set comprises 27 model pairs across diverse model families: Qwen cross-family pairs (7 pairs): • Qwen/Qwen2.5-7B vs. meta-llama/Llama-2-7b-hf. • Qwe...

    2026

  23. [40]

    • Gradient clipping: max norm 1.0

    • Optimizer: AdamW withβ 1 =0.9,β 2 =0.999, weight decay 0.01. • Gradient clipping: max norm 1.0. • Precision: bfloat16 mixed precision training. Data preparation: For each scale, we perform stratified random sampling from the full Alpaca dataset to ensure balanced representat...

    2026

  24. [1980]

    Gemma: Open models based on gemini research and technology.arXiv preprint arXiv:2403.08295,

    11 Published as a conference paper at COLM 2026 Gemma Gemma, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupati- raju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology....

    2026 arXiv

  25. [2002]

    Blackbox model provenance via palimpsestic membership inference.arXiv preprint arXiv:2510.19796,

    Rohith Kuditipudi, Jing Huang, Sally Zhu, Diyi Yang, Christopher Potts, and Percy Liang. Blackbox model provenance via palimpsestic membership inference.arXiv preprint arXiv:2510.19796,

  26. [2012]

    Experiment setup is consistent with Fig

    14 Published as a conference paper at COLM 2026 Appendix A Additional Visualizations of Spectral Signals (a)Kprojection (b)Vprojection (c)Oprojection (d) FFN DOWN projection Figure A1:Singular value distributions of middle-layer weight matrices for component types c∈ {K,V,O, F...

    2026

  27. [2017]

    Forgetting to forget: Attention sink as a gateway for backdooring llm unlearning.arXiv preprint arXiv:2510.17021,

    Bingqi Shang, Yiwei Chen, Yihua Zhang, Bingquan Shen, and Sijia Liu. Forgetting to forget: Attention sink as a gateway for backdooring llm unlearning.arXiv preprint arXiv:2510.17021,

  28. [2021]

    Black-box access is insufficient for rigorous ai audits

    Stephen Casper, Carson Ezell, Charlotte Siegmann, Noam Kolt, Taylor Lynn Curtis, Ben- jamin Bucknall, Andreas Haupt, Kevin Wei, Jérémy Scheurer, Marius Hobbhahn, et al. Black-box access is insufficient for rigorous ai audits. InProceedings of the 2024 ACM Conference on Fairnes...

    2024

  29. [2022]

    Detecting stylistic fingerprints of large language models.arXiv preprint arXiv:2503.01659,

    Yehonatan Bitton, Elad Bitton, and Shai Nisan. Detecting stylistic fingerprints of large language models.arXiv preprint arXiv:2503.01659,

    arXiv

  30. [2023]

    Model provenance testing for large language models.arXiv preprint arXiv:2502.00706,

    Ivica Nikolic, Teodora Baluta, and Prateek Saxena. Model provenance testing for large language models.arXiv preprint arXiv:2502.00706,

  31. [2024]

    Raft: Reward ranked finetuning for generative foundation model alignment.arXiv preprint arXiv:2304.06767,

    Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment.arXiv preprint arXiv:2304.06767,

    arXiv

  32. [2025]

    Does math reasoning improve gen- eral llm capabilities? understanding transferability of llm reasoning.arXiv preprint arXiv:2507.00432,

    Maggie Huan, Yuetai Li, Tuney Zheng, Xiaoyu Xu, Seungone Kim, Minxin Du, Radha Poovendran, Graham Neubig, and Xiang Yue. Does math reasoning improve gen- eral llm capabilities? understanding transferability of llm reasoning.arXiv preprint arXiv:2507.00432,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.