Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Independent per-instance neural-field weights can serve as semantically structured data representations when they are learned as multiplicative low-rank adaptations of a shared frozen base network.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 19:07 UTC pith:EEEOM4AU

load-bearing objection A useful empirical paper on multiplicative LoRA weight spaces for neural fields, with a shaky theoretical justification for why masking works. the 3 major comments →

arxiv 2512.01759 v3 pith:EEEOM4AU submitted 2025-12-01 cs.LG cs.AI

Weight Space Representation Learning via Neural Field Adaptation

Zhuoqian Yang , Mathieu Salzmann , Sabine S\"usstrunk This is my paper
classification cs.LG cs.AI
keywords weight space representationimplicit neural representationsneural fieldsmultiplicative low-rank adaptationLoRAdiffusion modelspermutation symmetrylinear mode connectivity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that a neural network's weights can be treated as data—reconstructable, generatable, and classifiable—provided the weight space is constrained by adapting a shared, pre-trained base neural field with multiplicative low-rank adaptation (mLoRA). Whereas standard LoRA adds a low-rank update (W + BA), mLoRA scales the base weights elementwise (W ⊙ BA), which the paper claims preserves the base network's channel structure and avoids feature entanglement. Across 2D images (FFHQ) and 3D shapes (ShapeNet), mLoRA weights reconstruct instances with lower error than standalone MLP weights, form semantically separated clusters, and, when a diffusion model is trained over the weight space, generate new faces and shapes of higher quality than previous weight-space methods like HyperDiffusion. The paper also shows that asymmetric masking (zeroing fixed random entries of A) makes the mLoRA weight space smoother, with independent optimization runs converging to a linear mode, and this geometry correlates strongly with generation performance. The larger claim is that the opacity of network weights is not intrinsic—it is a consequence of how the optimization space is chosen.

Core claim

The core claim is that weight-space geometry is a design choice. By freezing a shared pre-trained neural field and learning per-instance low-rank factors A and B that act multiplicatively—W' = W ⊙ BA—the paper obtains weights that reconstruct signals, cluster semantically, and support diffusion-based generation. The supplement rewrites mLoRA as a sum of channel-scaling terms diag(b_i) W diag(a_i), and the paper argues that asymmetric masking (zeroing fixed random entries of A) removes permutation ambiguity, aligning the weights with base-network channels and yielding convergence to a linear mode. Empirically, mLoRA-Asym gives the best reconstruction (PSNR 36.9 on FFHQ), the best generation s

What carries the argument

The central object is multiplicative low-rank adaptation (mLoRA), which updates a frozen base weight matrix W elementwise as W' = W ⊙ BA, with A ∈ R^{r×d_in}, B ∈ R^{d_out×r}. The supplement's Theorem 2 rewrites this as W' = Σ_i diag(b_i) W diag(a_i), showing that each rank component independently scales input and output channels. The companion asymmetric masking—randomly freezing (zeroing) a fixed subset of entries in each A, shared across all instances—is the mechanism meant to break the rank-permutation symmetry, so that each (a_i, b_i) is pinned to a specific channel axis. This pair of choices (multiplicative interaction + symmetry-breaking mask) is what the paper credits for the smoothn

Load-bearing premise

The claim that asymmetric masking eliminates the ambiguity of mLoRA weights rests on removing only discrete permutation symmetries, while leaving intact the continuous per-rank rescaling a_i → c_i a_i, b_i → c_i^{-1} b_i; if that continuous gauge freedom remains, the reported 'alignment with base network channels' and the linear mode behaviour lack their stated mechanism.

What would settle it

Fit the same instance twice with mLoRA-Asym under identical masks but different initializations, then check whether the two solutions differ only by per-rank rescalings (a_i → c_i a_i, b_i → c_i^{-1} b_i). If they do, the masking has not removed the continuous symmetry, and any observed linear mode connectivity cannot be attributed to the claimed channel-alignment mechanism; alternatively, normalize A rows before interpolation and see whether the barrier reappears.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Weight-space diffusion becomes viable at higher resolution: mLoRA-Asym produces recognizable 128×128 faces, a scale previous weight-space methods did not reach.
  • Because mLoRA weight space is semantically organized, discriminative tasks (1-NN, logistic regression, k-means) can be performed directly on the raw weight parameters without a learned encoder.
  • Weight-space smoothness—linear mode connectivity and high cross-initialization similarity—strongly predicts generation quality, motivating future work to engineer flatter weight manifolds.
  • The method transfers across modalities through a single base architecture, supporting reconstruction, generation, and classification for both 2D images and 3D shapes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same masking fails to give additive LoRA the linear-mode benefit, so the paper's hypothesis that feature entanglement (not just permutation symmetry) is the bottleneck would predict that a continuous symmetry-removing regularizer on mLoRA—one that fixes the per-rank rescaling gauge—should reproduce the linear-mode behaviour without freezing parameters.
  • If the claimed alignment of mLoRA with base-network channels is correct, individual rank components could be probed semantically by examining which channels they modulate, giving a new interpretability tool for neural-field generators.
  • The approach shares its modulation view with style-based generators (weight modulation via per-channel scaling), suggesting that weight-space learning could connect to existing latent-space editing and animation techniques for such models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes to treat the weights of neural fields as data representations by fitting each instance with a multiplicative low-rank adaptation (mLoRA) of a shared, pre-trained base network. The mLoRA weight update is W' = W ⊙ BA, optionally combined with asymmetric masking to reduce rank-permutation symmetry. The authors compare six representations (standalone MLP, additive LoRA, multiplicative LoRA, each with and without asymmetric masking) on FFHQ and ShapeNet across reconstruction, weight-space stability, diffusion-based generation, and classification/clustering. Their main empirical findings are that mLoRA and mLoRA-Asym give better reconstruction, smoother weight-space behavior, better generation metrics, and better semantic separability than additive LoRA and standalone MLP weights, and that mLoRA-Asym in particular exhibits low linear-interpolation barriers.

Significance. If the empirical results hold, this is a useful step for weight-space representation learning: it shows that independent optimization, when constrained by a shared base model and multiplicative adaptation, can produce weight representations that are simultaneously high-fidelity, learnable by diffusion, and semantically structured. The paper's strengths are the controlled comparisons across parameterization and symmetry-breaking, the use of multiple generation metrics (FD, MMD, mMD/COV/1-NNA), and the supplementary ablation of the hierarchical diffusion encoder. The main reservation is that the theoretical mechanism offered for the headline linear-mode result is not valid as written, and one implementation detail (the role of the base model's latent codes) is unspecified. These are fixable in revision.

major comments (3)
  1. [Supplementary §6.2, Corollary 2.2; main §3.2, §4.2] Corollary 2.2 is invalid as stated. Equation (18) gives W' = Σ_i diag(b_i) W diag(a_i). For any nonzero c_i, the transformation a_i -> c_i a_i and b_i -> c_i^{-1} b_i leaves each b_i a_i^T (and hence W') unchanged. Asymmetric masking only zeroes fixed entries of A; a diagonal scaling preserves the zero pattern, so this continuous gauge freedom survives the mask. The corollary's conclusion that 'each rank component i is uniquely identified' therefore does not follow; the proof assumes the uniqueness it is proving. Because §4.2 explicitly attributes the exceptional linear-mode connectivity of mLoRA-Asym to Corollary 2.2, the mechanistic explanation in the paper is unsupported. To retain this claim, the authors need to fix the gauge explicitly (e.g., canonicalize each a_i) or downgrade the claim to an empirical observation.
  2. [§3.3 and Algorithm 1] The paper never says what happens to the per-instance latent codes {z_i} learned during base-model training (Eq. 4) when the LoRA adaptation in Eq. (2) is performed. The base architecture (Fig. 7b) takes a style vector derived from z_i as input. Algorithm 1 omits z_i in Stage 2, and Eq. (2) writes f(p|LoRA(W, φ)) without a latent argument. If z_i is retained as a conditioning input, then the representation is not the LoRA weights alone, undermining the claim that each instance is represented by φ_i. If z_i is discarded or set to a fixed value, this must be stated, because the base network's behavior is otherwise undefined. Please clarify this step; it is needed to interpret the classification and generation results.
  3. [§4.2, Figure 3] The stability analysis reports cosine similarity and linear-interpolation barrier of raw LoRA factors. Both quantities are gauge-dependent: two optimizations can represent identical functions but differ by the residual scaling symmetry noted above, making the cosine similarity arbitrarily low while the functional barrier remains zero. The high similarity of mLoRA-Asym in Figure 3 may therefore reflect the shared initialization and the shared zero mask rather than a canonical alignment of optimization paths. To support the 'converge to a linear mode' claim, the authors should either gauge-align the factors before computing similarity or report a function-space or invariance-aware distance.
minor comments (6)
  1. [§4.3] Typo: 'recg-nizable' should be 'recognizable'.
  2. [§3.4] The phrase 'zeroing out frozen entries ... removes the corresponding rank component's contribution' is imprecise: zeroing one entry in A removes one input channel for that rank component, not the whole rank component.
  3. [§9.1] The Fréchet Distance formula divides by N_feature, which is nonstandard and changes the scale of the metric. Please state why this normalization is used, since it makes FD values difficult to compare with prior work.
  4. [Figure 3] Please define the range and meaning of λ on the x-axis and label the y-axis units explicitly; currently the reader must infer them from the text.
  5. [§4.3] The statement that HyperDiffusion 'fails to produce recognizable face images' is qualitative; Table 2 gives FD 0.241 for HyperDiffusion, so a brief quantitative characterization of the failure mode would strengthen the comparison.
  6. [References] Several references have inconsistent page-number formatting (e.g., [2], [19]); please harmonize.

Circularity Check

0 steps flagged

No significant circularity: the mLoRA weight-space claims are empirical and benchmarked externally; the questionable Corollary 2.2 is a proof-gap issue, not a circular derivation.

full rationale

The paper's central derivation chain is empirical rather than definitional. Representations are obtained by optimizing L_recon, and then evaluated through reconstruction quality, weight-space similarity/linear-mode connectivity, diffusion-based generation, and discriminative tasks. None of these downstream evaluations is forced by construction: generation is assessed on newly sampled weights via distributional metrics against reference sets, classification/clustering use external ShapeNet labels, and the linear-mode connectivity result is an observed geometric phenomenon rather than a quantity derived from the fitting loss. The only mildly self-referential element is that reconstruction quality is measured on the same fitting objective used to create the representation, but this is an intrinsic benchmark of representation capacity, not a renamed prediction of an independent quantity. The paper does not rely on load-bearing self-citations; prior-work dependencies such as asymmetric masking [25], GL(r) equivalence [24], and LoRA [18] are external. Corollary 2.2 is invoked as a mechanistic explanation for the mLoRA-Asym linear-mode behavior, but even if its proof is mathematically incomplete—the masked multiplicative LoRA factors retain a continuous per-rank scaling gauge freedom, so 'uniquely identified' does not strictly follow from Eq. 18—that is a correctness/rigor concern, not circularity. No step reduces the paper's conclusions to its own inputs by definition or by fitted-parameter renaming.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central empirical recipe rests on two free design choices (rank, mask pattern) and on a symmetry-breaking assumption that is not fully justified. No new physical entities are introduced.

free parameters (6)
  • LoRA rank r = not reported
    Controls dimensionality and capacity of all mLoRA/LoRA representations; absent from implementation details, so replication requires guessing.
  • Fourier feature frequency omega_0 = 32 (FFHQ), 1 (ShapeNet)
    Section 8.1, chosen empirically and shared with LoRA representations; changes which frequencies the network can represent.
  • Asymmetric mask size = sqrt(d_out) frozen entries per row of A
    Section 3.4; inherited from [25]; the exact random mask pattern is not released.
  • Frozen-entry variance kappa = 6
    Section 3.4/4.1, empirically determined for additive LoRA; the paper notes large kappa hurts neural fields.
  • Latent regularization lambda_r = 1e-4
    Section 8.3; controls base-model training and shapes the adaptation subspace used by all LoRA representations.
  • Diffusion noise schedule beta_1, beta_T, T = 1e-4, 2e-2, 500
    Section 3.5/8.5; standard DDPM schedule, but chosen values affect generation quality and are not treated as sensitive.
axioms (5)
  • domain assumption A base neural field trained by variational autodecoding on a dataset captures transferable features that LoRA can adapt per instance.
    Section 3.3/8.3; if this fails, LoRA weights are not meaningful representations. Not proven beyond the reconstruction results.
  • domain assumption LoRA weight updates lie in a meaningful low-dimensional subspace so that constraining to LoRA induces structure.
    Motivation in Section 3.1; borrowed from Hu et al., not re-derived here.
  • domain assumption Permutation/GL(r) symmetry is the main cause of multi-modality in weight space, and breaking it yields unimodal, linearly connected weight spaces.
    Section 3.4 and Corollary 2.2; this is the underlying claim of the structure analysis.
  • ad hoc to paper Asymmetric masking removes internal symmetries for multiplicative LoRA as it does for additive LoRA.
    Section 3.4/Cor 2.2. For mLoRA, zeroing entries of A preserves the per-rank scaling symmetry of Eq. 18; the paper does not address this. Load-bearing for the linear-mode claim.
  • ad hoc to paper Feature entanglement in additive composition explains why additive LoRA is worse than multiplicative LoRA.
    Section 3.2/4.2; presented as a hypothesis, not independently tested against alternative explanations.

pith-pipeline@v1.3.0-alltime-deepseek · 18623 in / 19068 out tokens · 195131 ms · 2026-08-03T19:07:12.909236+00:00 · methodology

0 comments
read the original abstract

We investigate the potential of weights to serve as effective representations, focusing on neural fields. Our key insight is that constraining the optimization space through a pre-trained base model and low-rank adaptation (LoRA) can induce structure in weight space. Across reconstruction, generation, and analysis tasks on 2D and 3D data, we find that multiplicative LoRA weights achieve high representation quality while exhibiting distinctiveness and semantic structure. When used with latent diffusion models, multiplicative LoRA weights enable higher-quality generation than existing weight-space methods.

Figures

Figures reproduced from arXiv: 2512.01759 by Mathieu Salzmann, Sabine S\"usstrunk, Zhuoqian Yang.

Figure 1
Figure 1. Figure 1: LoRA based weight space representation with neural [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diffusion Transformer with hierarchical LoRA layer encoder architecture. For each layer l, we treat vector pairs (a (i) l , b (i) l ) as tokens. Vector-level positional encodings capture rank dimension indices, followed by multi-head attention that models interactions among the r rank components within the layer. This hierarchical design enables the model to learn both local (within-layer) dependencies amo… view at source ↗
Figure 3
Figure 3. Figure 3: Weight space structure analysis. We measure weight similarity (cosine similarity) and the linear mode connectivity bar￾rier (Chamfer distance) as a function of initialization perturbation strength λ. Each data point is averaged from 30 instances, the un￾derlying shades are indicative of standard deviation. Results. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative generation results. Generated samples from diffusion models trained on different weight space representations. The top 2 rows show results generated by the Airplane model, followed by 2 rows from the Multi-class model. The bottom rows show 2D FFHQ generations. matical form rather than using modality-specific implemen￾tations like FID [16] or KID [2], as this allows for con￾sistent evaluation ac… view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE visualization of weight spaces. Each point rep￾resents one instance from the ShapeNet ten-category dataset, col￾ored by object category. Multiplicative LoRA weight spaces ex￾hibit semantic structure. mLoRA delivers the best results overall, achieving 90% ac￾curacy with a linear classifier. Given that mLoRA-Asym exhibits better weight space structure in terms of linear mode connectivity, this result i… view at source ↗
Figure 6
Figure 6. Figure 6: Illustrating permutation symmetries within LoRA. (a) Permutation symmetry in additive LoRA. Low rank matri￾ces A and B could be seen as an encoder layer and a decoder layer. The order of the intermediate neurons could swapped with￾out changing the output. (b) multiplicative LoRA could be seen as parallel pathways with different scaling factors for the input and output. The order of the pathways could be sw… view at source ↗
Figure 7
Figure 7. Figure 7: Network architectures for weight space representations. (a) Standalone MLP architecture with Fourier Feature layer followed by two linear layers. (b) Base model architecture with modulated fully connected layers. The network takes spatial coordinates p and style vector s as inputs, applying multiplicative weight modulation at each layer. (c) LoRA adaptation applied to the base model, where low rank matrice… view at source ↗
Figure 9
Figure 9. Figure 9: Novelty check. For each generated FFHQ image (left), we show its nearest neighbor from the training set retrieved by CLIP feature similarity (right). Generated samples are visually distinct from training data, confirming generalization rather than memorization. 13. Limitation and Future Work While our work establishes that neural network weights can serve as effective data representations, several limitati… view at source ↗
Figure 8
Figure 8. Figure 8: visualizes linear interpolations between pairs of mLoRA weight representations on FFHQ. We linearly inter￾polate between two instances’ LoRA weight pairs (A1, B1) and (A2, B2), evaluating the resulting neural field at each interpolation step. While the interpolated weights do not always produce smooth, perceptually gradual transitions between instances (as would be expected from a contin￾uous learned laten… view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative generation results on ShapeNet - Airplanes. HyperDiffusion MLP-Asym LoRA LoRA-Asym mLoRA mLoRA-Asym [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional qualitative generation results on ShapeNet - Multi. 7 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional qualitative generation results on FFHQ. 8 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 9 linked inside Pith

  1. [1]

    Image generators with conditionally-independent pixel synthesis

    Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 14278–14287, 2021. 1, 2, 3, 5

    2021

  2. [2]

    Demystifying mmd gans.arXiv preprint arXiv:1801.01401, 2018

    Mikołaj Bi ´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans.arXiv preprint arXiv:1801.01401, 2018. 6, 7, 5

    Pith/arXiv arXiv 2018

  3. [3]

    pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis

    Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. arxiv e- prints, page.arXiv preprint arXiv:2012.00926, 2020. 1, 2, 3, 5

    Pith/arXiv arXiv 2012

  4. [4]

    Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015. 5

    Pith/arXiv arXiv 2015

  5. [5]

    Transformers as meta- learners for implicit neural representations

    Yinbo Chen and Xiaolong Wang. Transformers as meta- learners for implicit neural representations. InEuropean Conference on Computer Vision, pages 170–187. Springer,

  6. [6]

    Interpreting the weight space of customized dif- fusion models.Advances in Neural Information Processing Systems, 37:137334–137371, 2024

    Amil Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei Efros, and Kfir Aberman. Interpreting the weight space of customized dif- fusion models.Advances in Neural Information Processing Systems, 37:137334–137371, 2024. 2

    2024

  7. [7]

    Coin: Compression with implicit neural representations.arXiv preprint arXiv:2103.03123,

    Emilien Dupont, Adam Goli ´nski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations.arXiv preprint arXiv:2103.03123,

    Pith/arXiv arXiv

  8. [8]

    Emilien Dupont, Hyunjik Kim, S. M. Ali Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. InIn- ternational Conference on Machine Learning, 2022. 2

    2022

  9. [9]

    Hyperdiffusion: Generating implicit neural fields with weight-space diffusion

    Ziya Erkoc ¸, Fangchang Ma, Cengiz ¨Oztireli, and Pascal Fua. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. InInternational Conference on Computer Vision, 2023. 1, 2, 3, 4, 6, 7, 5

    2023

  10. [10]

    Where do we stand with implicit neural representations? a technical and performance survey.arXiv preprint arXiv:2411.03688,

    Amer Essakine, Yanqi Cheng, Chun-Wun Cheng, Lipei Zhang, Zhongying Deng, Lei Zhu, Carola-Bibiane Sch¨onlieb, and Angelica I Aviles-Rivero. Where do we stand with implicit neural representations? a technical and performance survey.arXiv preprint arXiv:2411.03688,

    Pith/arXiv arXiv

  11. [11]

    Linear mode connectivity and the lot- tery ticket hypothesis

    Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lot- tery ticket hypothesis. InInternational Conference on Ma- chine Learning, pages 3259–3269. PMLR, 2020. 6

    2020

  12. [12]

    Sur la distance de deux lois de probabilit´e

    Maurice Fr ´echet. Sur la distance de deux lois de probabilit´e. InAnnales de l’ISUP, pages 183–198, 1957. 6, 5

    1957

  13. [13]

    Revisiting model merging: A statistical perspective.arXiv preprint, 2024

    Y Gao and Others. Revisiting model merging: A statistical perspective.arXiv preprint, 2024. 2, 3

    2024

  14. [14]

    D’oh: Decoder-only ran- dom hypernetworks for implicit neural representations

    Cameron Gordon, Lachlan E MacDonald, Hemanth Saratchandran, and Simon Lucey. D’oh: Decoder-only ran- dom hypernetworks for implicit neural representations. In Proceedings of the Asian Conference on Computer Vision, pages 2507–2526, 2024. 2

    2024

  15. [15]

    Hypernetworks

    David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. InInternational Conference on Learning Representations,

  16. [16]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017. 6, 7, 5

    2017

  17. [17]

    Meta- learning in neural networks: a survey

    T Hospedales, A Antoniou, P Micaelli, and A Storkey. Meta- learning in neural networks: a survey. arxiv preprint arxiv: 200405439. 2020. 1

    2020

  18. [18]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 1, 3

    2022

  19. [19]

    Re- thinking fid: Towards a better evaluation metric for image generation

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Re- thinking fid: Towards a better evaluation metric for image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307– 9315, 2024. 5

    2024

  20. [20]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 2, 5

    2019

  21. [21]

    Alias-free generative adversarial networks.Advances in neural infor- mation processing systems, 34:852–863, 2021

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks.Advances in neural infor- mation processing systems, 34:852–863, 2021. 1, 2, 3, 5

    2021

  22. [22]

    Hypernetwork functional image representation

    Sylwester Klocek, Łukasz Maziarka, Maciej Wołczyk, Jacek Tabor, Jakub Nowak, and Marek ´Smieja. Hypernetwork functional image representation. InInternational Conference on Artificial Neural Networks, pages 496–510. Springer,

  23. [23]

    Graph neural net- works for learning equivariant representations of neural net- works

    M Kofinas, B Knyazev, Y Zhang, Y Chen, G J Burghouts, E Gavves, C G M Snoek, and D W Zhang. Graph neural net- works for learning equivariant representations of neural net- works. InThe Twelfth International Conference on Learning Representations, 2024. 2

    2024

  24. [24]

    Learning on lo- ras: Gl-equivariant processing of low-rank weight spaces for large finetuned models.arXiv preprint arXiv:2410.04207,

    D Lim, Y Gelberg, S Jegelka, H Maron, et al. Learning on lo- ras: Gl-equivariant processing of low-rank weight spaces for large finetuned models.arXiv preprint arXiv:2410.04207,

    Pith/arXiv arXiv

  25. [25]

    The empirical impact of neural param- eter symmetries, or lack thereof.Advances in Neural Infor- mation Processing Systems, 37:28322–28358, 2024

    Derek Lim, Theo Putterman, Robin Walters, Haggai Maron, and Stefanie Jegelka. The empirical impact of neural param- eter symmetries, or lack thereof.Advances in Neural Infor- mation Processing Systems, 37:28322–28358, 2024. 1, 4

    2024

  26. [26]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837–2845, 2021. 7, 5

    2021

  27. [27]

    Equivariant architectures for learning in deep weight spaces

    A Navon, A Shamsian, I Achituve, E Fetaya, G Chechik, and H Maron. Equivariant architectures for learning in deep weight spaces. InInternational Conference on Machine Learning, pages 25790–25816, 2023. 2, 3 9

    2023

  28. [28]

    Fusion of graph convolutional net- works via optimal transport.arXiv preprint, 2025

    O Ormaniec and Others. Fusion of graph convolutional net- works via optimal transport.arXiv preprint, 2025. 2

    2025

  29. [29]

    Deepsdf: Learning con- tinuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con- tinuous signed distance functions for shape representation. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 165–174, 2019. 2, 3

    2019

  30. [30]

    Learning to learn with genera- tive models of neural network checkpoints.arXiv preprint arXiv:2209.12892, 2023

    William Peebles, Ilija Radosavovic, Tim Brooks, Alexei A Efros, and Jitendra Malik. Learning to learn with genera- tive models of neural network checkpoints.arXiv preprint arXiv:2209.12892, 2023. 1, 2, 4

    Pith/arXiv arXiv 2023

  31. [31]

    Scalable diffusion mod- els with transformers

    William S Peebles and Saining Xie. Scalable diffusion mod- els with transformers. 2023 ieee. InCVF International Con- ference on Computer Vision (ICCV), 2022. 4

    2023

  32. [32]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017. 7, 5

    2017

  33. [33]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 6, 7, 5

    2021

  34. [34]

    Model fusion via optimal transport

    C Singh and M Jaggi. Model fusion via optimal transport. In Advances in Neural Information Processing Systems, 2020. 2

    2020

  35. [35]

    Hyper-align: Efficient modality alignment via hy- pernetworks

    Jaisidh Singh, Diganta Misra, and Boris Knyazev6 Antonio Orvieto. Hyper-align: Efficient modality alignment via hy- pernetworks. InWorkshop on Neural Network Weights as a New Data Modality. 1

  36. [36]

    Fourier features let networks learn high frequency functions in low dimen- sional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen- sional domains.Advances in neural information processing systems, 33:7537–7547, 2020. 3

    2020

  37. [37]

    Lion: Latent point dif- fusion models for 3d shape generation.Advances in Neural Information Processing Systems, 35:10021–10039, 2022

    Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point dif- fusion models for 3d shape generation.Advances in Neural Information Processing Systems, 35:10021–10039, 2022. 7, 5, 6

    2022

  38. [38]

    Learning transferable features for implicit neural representations.Advances in Neural Information Processing Systems, 37:42268–42291, 2024

    Kushal Vyas, Ahmed I Humayun, Aniket Dashpute, Richard G Baraniuk, Ashok Veeraraghavan, and Guha Bal- akrishnan. Learning transferable features for implicit neural representations.Advances in Neural Information Processing Systems, 37:42268–42291, 2024. 2

    2024

  39. [39]

    Scaling weight space generative mod- els.arXiv preprint, 2025

    K Wang and Others. Scaling weight space generative mod- els.arXiv preprint, 2025. 2

    2025

  40. [40]

    Neural network diffusion.arXiv preprint arXiv:2402.13144,

    K Wang, Z Xu, Y Zhou, Z Zang, T Darrell, Z Liu, and Y You. Neural network diffusion.arXiv preprint arXiv:2402.13144,

    Pith/arXiv arXiv

  41. [41]

    Neural fields in visual computing and beyond

    Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp- kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. InComputer graphics forum, pages 641–676. Wiley Online Library, 2022. 1, 2

    2022

  42. [42]

    Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666, 2024

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xi- aochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666, 2024. 1

    Pith/arXiv arXiv 2024

  43. [43]

    A structured dictionary perspective on implicit neural representations

    Gizem Y ¨uce, Guillermo Ortiz-Jim ´enez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228–19238, 2022. 3, 6

    2022

  44. [44]

    Symmetry in neural network parameter spaces.arXiv preprint arXiv:2506.13018, 2025

    Bo Zhao, Robin Walters, and Rose Yu. Symmetry in neural network parameter spaces.arXiv preprint arXiv:2506.13018, 2025. 1, 3

    arXiv 2025

  45. [45]

    Permutation equivariant neural functionals.Advances in neural information processing systems, 36:24966–24992,

    Allan Zhou, Kaien Yang, Kaylee Burns, Adriano Cardace, Yiding Jiang, Samuel Sokota, J Zico Kolter, and Chelsea Finn. Permutation equivariant neural functionals.Advances in neural information processing systems, 36:24966–24992,

  46. [46]

    Neural functional transformers.Advances in neural infor- mation processing systems, 36:77485–77502, 2023

    Allan Zhou, Kaien Yang, Yiding Jiang, Kaylee Burns, Win- nie Xu, Samuel Sokota, J Zico Kolter, and Chelsea Finn. Neural functional transformers.Advances in neural infor- mation processing systems, 36:77485–77502, 2023. 1, 2, 5

    2023

  47. [47]

    3d shape generation and completion through point-voxel diffusion

    Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 5826–5835, 2021. 7, 5 10 Weight Space Representation Learning via Neural Field Adaptation Supplementary Material This supplementary material provides additional theoreti- ...

    2021

  48. [48]

    Permutation Symmetry in LoRA We provide a formal proof that permutation symmetry ex- ists within both additive and multiplicative LoRA parame- terizations. 6.1. Permutation Symmetry in Additive LoRA Theorem 1.The adapted weight matrix from additive LoRA exhibits permutation symmetry with respect to the rank di- mensions. Proof.Consider the additive LoRA f...

  49. [49]

    INRs repre- sent signals as continuous functionsΦ :R n →R m, where a neural network mapsn-dimensional coordinates tom- dimensional quantities

    Related Works -Implicit Neural Representations Implicit Neural Representations (INRs), also known as neu- ral fields, are continuous functions parameterized by neural networks that map coordinates to signal values. INRs repre- sent signals as continuous functionsΦ :R n →R m, where a neural network mapsn-dimensional coordinates tom- dimensional quantities....

  50. [50]

    Standalone MLP As shown in Figure 7(a), The standalone MLP is a Fourier Feature [36] layerα 1 = sin(ω0 ·(W 1p+b 1))followed by 2linear layers

    Implementation Details 8.1. Standalone MLP As shown in Figure 7(a), The standalone MLP is a Fourier Feature [36] layerα 1 = sin(ω0 ·(W 1p+b 1))followed by 2linear layers. •ω 0 is the frequency scaling factor for the Fourier Feature layer. For 2D FFHQ, we setω 0 = 32; For 3D ShapeNet, we setω 0 = 1. This value is chosen empiracally and shared with its LoRA...

  51. [51]

    For 3D, we also calculate distance-based metrics

    Evaluation Metrics for Generation We calculate distributional metrics for both 2D and 3D. For 3D, we also calculate distance-based metrics. All metrics are calculated between2,048generated samples and2,048 reference samples. 9.1. Distributional Difference These metrics operate on features extracted by deep learned models, as deep learned feature extractor...

  52. [52]

    Following the practice of KID [2], we choose γp = 1/Nfeature

    Polynomial kernel (MMD-P) ψp(x,y) = (γ p ·x T y+c) d with degreed= 3and offset c= 1. Following the practice of KID [2], we choose γp = 1/Nfeature

  53. [53]

    Gaussian RBF kernel (MMD-G) ψg(x,y) = exp(−γ g∥x−y∥ 2 2)withγ g = 1/(2σ2 g); we chooseσ g =N feature. 9.2. Distance-based Metrics for 3D Shapes For 3D shapes, we calculate distance-based metrics follow- ing [9, 26, 37, 47]. We denote the distance function as D(x,y)for the Chamfer Distance between two shapesx andy. The metrics are defined as: mMD(P, Q) =Ey...

  54. [54]

    To isolate its contribution, we train a baseline diffusion model without the layer encoder on the ShapeNet multi-category dataset

    Ablation Study We examine the effectiveness of the hierarchical LoRA layer encoder introduced in Section 3.5. To isolate its contribution, we train a baseline diffusion model without the layer encoder on the ShapeNet multi-category dataset. This baseline treats each weight matrix as an independent token, directly feeding flattened LoRA matrices into the t...

  55. [55]

    We linearly inter- polate between two instances’ LoRA weight pairs(A1,B 1) and(A 2,B 2), evaluating the resulting neural field at each interpolation step

    Weight Space Interpolation Figure 8 visualizes linear interpolations between pairs of mLoRA weight representations on FFHQ. We linearly inter- polate between two instances’ LoRA weight pairs(A1,B 1) and(A 2,B 2), evaluating the resulting neural field at each interpolation step. While the interpolated weights do not always produce smooth, perceptually grad...

  56. [56]

    Please see Figure 10 for results on ShapeNet Air- planes, Figure 11 for results on ShapeNet Multi, and Fig- ure 12 for results on FFHQ

    Additional Qualitative Results We provide additional qualitative results on diffusion gen- eration. Please see Figure 10 for results on ShapeNet Air- planes, Figure 11 for results on ShapeNet Multi, and Fig- ure 12 for results on FFHQ. To verify that generated samples are novel rather than memorized reproductions of training data, we perform a CLIP-based ...

  57. [57]

    First, our approach requires all instances to share the same pre-trained base model and initialization

    Limitation and Future Work While our work establishes that neural network weights can serve as effective data representations, several limitations present opportunities for future research. First, our approach requires all instances to share the same pre-trained base model and initialization. This is a practical limitation: the base model may not suit all...