Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Update Your Transformer to the Latest Release: Re-Basin of Task Vectors

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A structured two-level permutation moves a fine-tuned Transformer's task vector to a new checkpoint, preserving functional equivalence without data or retraining.

desk verdict Useful empirical recipe for transferring task vectors across Transformer checkpoints, but the functional-equivalence claim is weakened by modifying residual connections rather than using pure weight permutations. read the letter →

arxiv 2505.22697 v1 pith:S5DO3HAK submitted 2025-05-28 cs.LG

Filippo Rinaldi , Giacomo Capitani , Lorenzo Bonicelli , Donato Crisostomi , Federico Bolelli , Elisa Ficarra , Emanuele Rodolà , Simone Calderara
show 1 more author
Angelo Porrello
This is my paper
classification cs.LG
keywords taskvectortransportmodelre-basinweightpermutationmulti-headattentionalignmentspectraldistancedata-freetransferlinearmodeconnectivityTransformerfine-tuning
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

Foundation-model releases normally leave every downstream fine-tune obsolete, forcing expensive retraining and access to data that may have been deleted. This paper claims the fine-tune can instead be ported: take the task vector $\tau = \theta_A^{\mathrm{ft}} - \theta_A$, apply a permutation $\pi$, and add it to the new checkpoint as $\theta_B + \pi(\tau)$, with no data and no training steps. The recipe is built for Transformers: a two-level permutation first reorders attention heads and then permutes units inside matched heads, and residual connections are handled by composing permutations so both branches of every skip connection transform together. If the claim holds, model updates become cheap, private-data constraints stop blocking reuse, and multiple fine-tunes can still be merged after transport.

What carries the argument

The central machinery is a structured block permutation $P_{\mathrm{attn}} = \sum_{i} E_{i,\pi(i)} \otimes P^{(i)}_{\mathrm{intra}}$ acting on the query, key, and value projections of every multi-head attention layer. Inter-head permutation $\pi$ is selected with the Hungarian algorithm on a cost built from Euclidean distances between singular-value spectra of head matrices, a rotation-invariant comparison; intra-head permutations $P^{(i)}_{\mathrm{intra}}$ are then selected by linear assignment inside each matched pair. The structure prevents head contamination and yields the equivariance identity $O' = O P_{\mathrm{attn}}$, meaning the permuted attention block computes the same function up to a permutation of its output. Residual connections are repaired by replacing the identity skip with compositions such as $I_i = P_{W_0} P_{\mathrm{in}}^{\top}$, so the two addends of every sum always carry the same permutation. The whole matching pass runs in $O(L\,d_m^3)$, polynomial in depth and width rather than dependent on training data.

What would settle it

Take a fixed input batch, run $\theta_A$ and the permuted model $\pi(\theta_A)$ with layer normalization enabled, and check whether the output relation $O' = O P_{\mathrm{attn}}$ holds to numerical precision. If it does not, or if recomputing the transport with an exact LayerNorm-aware handling fails to improve over plain $\theta_B + \tau$, the paper's central claims are refuted.

Watch

Extended reading notes

Core claim

TransFusion claims that the task vector $\tau = \theta_A^{\mathrm{ft}} - \theta_A$ can be transported to a new checkpoint $\theta_B$ as $\theta_B + \alpha\,\pi(\tau)$, with $\pi$ a structured permutation produced by a data-free weight-matching pass. The pass has two levels: heads of the attention projection tensors are matched by Euclidean distance between singular-value spectra (inter-head), and then rows within each matched head pair are assigned by linear assignment (intra-head). The composed block permutation $P_{\mathrm{attn}} = \sum_i E_{i,\pi(i)} \otimes P^{(i)}_{\mathrm{intra}}$ is proven to make attention equivariant, $O' = O P_{\mathrm{attn}}$, and residual connections are handled by replacing the skip identity with compositions such as $I_i = P_{W_0}\,P_{\mathrm{in}}^{\top}$ so both addends transform alike. On CLIP ViT and GLUE experiments, $\theta_B + \pi(\tau)$ improves task accuracy relative to $\theta_B$ while roughly preserving support-set accuracy.

Load-bearing premise

The load-bearing premise is that layer normalization can be treated as a standard linear projection when the permutations are composed across residual connections; if that simplification fails on real Transformers, the derived permutations do not guarantee functional equivalence and the transported task vector is distorted.

Editorial extensions

If this is right

  • A fine-tuned model becomes portable: the new checkpoint $\theta_B + \pi(\tau)$ can be used immediately, so a released upgrade of a backbone no longer forces downstream retraining.
  • Because attention output is only permuted, interpolation $L((1-\alpha)\pi(\theta_A)+\alpha\theta_B)$ stays on a low-loss path, and the permuted model is lossless with respect to the original.
  • One alignment $\pi$ between two checkpoints serves all task vectors from the same source, so the cost of re-basin amortizes over any number of downstream tasks.
  • Task-vector merging remains valid at the destination: transported vectors can still be combined, so re-basin composes with model merging rather than replacing it.
  • Transport keeps the new model's zero-shot capabilities on unseen data, so updating the backbone improves task performance without sacrificing generality.

Reading between the lines

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

  • One extension the paper leaves implicit is batch-wise amortization: a single alignment $\pi$ between two checkpoints can port any number of task vectors from the same source model, reducing per-task cost to a single addition after the one-time matching pass.
  • Because the inter-head cost compares only singular-value spectra, heads with identical spectra but different internal bases are treated as interchangeable; an activation-aware or basis-aligned matching cost would be a direct test of whether the SVD shortcut is sufficient.
  • The same two-level scheme should in principle apply to decoder-only and encoder-decoder Transformers, where causal masking and tied input/output embeddings add compatibility constraints not analyzed in the paper.
  • The paper notes that a weak source expert degrades transport; that suggests an operational rule: check source task accuracy before transporting, and shrink $\alpha$ when the source accuracy is low.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes TransFusion, a data-free method for transferring a task vector from a fine-tuned Transformer checkpoint to a newly released checkpoint without retraining. The method aligns the old and new backbones by a two-level permutation strategy: first it permutes attention heads using a spectral, permutation-invariant distance based on singular values, then it permutes units within matched heads via linear assignment. Residual connections are handled by replacing identity skip connections with compositions of permutation matrices. The authors prove an equivariance theorem for multi-head attention under their structured permutations, provide complexity analysis, and report gains over naive task-vector transport and existing re-basin baselines on several visual and NLP tasks.

Significance. If the transport guarantee were fully established, this would be a practically valuable contribution: it would allow updating a fine-tuned model to a new pretrained checkpoint without data or training, and it offers a concrete way to handle the multi-head structure that prior re-basin methods ignore. The inter-head spectral matching idea is novel, the equivariance proof for the attention module is clean and self-contained, the main tables fix the scaling at alpha=1 rather than tuning it on the target benchmark, and code is provided. The weakness is that the global functional-equivalence claim currently rests on an incomplete treatment of residual connections and normalization layers; this is the main gap that the revision must close.

major comments (4)
  1. [Sec. 3.1, Step 3; Eq. (12); Appendix A.3] The residual handling replaces the identity skip connection with non-identity permutation matrices I_i = P_W0 P_in^T and I_out = P_W2 P_W0^T. This changes the computation graph of the permuted model, so pi(theta_A) is not a weight-permutation reparameterization of theta_A: a fixed identity branch has been replaced by explicit linear maps. The proof of Theorem 3.1 covers only the multi-head attention module, and the residual derivation in Appendix A.3 explicitly ignores the ReLU nonlinearity, so the global functional-equivalence claim is not established. The paper should either derive a permutation assignment that leaves the identity skips intact, or explicitly present the modified architecture as part of the method and discuss the consequences for Eq. (14).
  2. [Sec. 3.1, Step 3] The statement that LayerNorm 'can be regarded as a standard linear projection' is not correct: LayerNorm involves a nonlinear normalization step with input-dependent statistics. Although LayerNorm is permutation-equivariant when its affine parameters are permuted consistently, the paper does not specify how LayerNorm weights are included in the permutation pi, and the proofs omit LayerNorm entirely. Since real ViT models contain LayerNorm, the functional-equivalence claim for the full network needs a precise treatment of these layers, or the claim must be weakened.
  3. [Sec. 4.2; Fig. 5] The paper infers functional equivalence from low loss at the interpolation endpoint alpha=0, but it does not directly measure whether pi(theta_A) and theta_A produce identical outputs. Given that the residual handling changes the computation graph, a direct check of output equivalence (for example, the maximum activation difference over a validation set) is necessary to support the central claim. Without such a check, the empirical interpolation plots do not distinguish exact functional equivalence from approximate alignment.
  4. [Sec. 3.2; Eq. (14)] The transport formula theta_B + alpha*pi(tau) is justified through the functional equivalence of pi(theta_A), but the computation graph of theta_B + pi(tau) is not specified. If the non-identity residual maps I_i are part of pi, then the transported model is an architecture different from the released backbone theta_B; if they are not, then pi was derived under one graph and applied to another. The paper should state explicitly which computation graph is used for theta_B + pi(tau) and how the residual permutation insertions are (or are not) carried over.
minor comments (5)
  1. [Table 2] The column header 'C OLA' should read 'CoLA'.
  2. [Sec. 3.1, Step 1; Eq. (7)] The notation h_B^i = [\tilde{W}]^B_{i,:,:} is ambiguous because it does not indicate whether the head is taken from the query, key, or value tensor; the subsequent definition of D_ij as a sum over q, k, v should be made explicit in the main text.
  3. [Algorithm 2, line 7] The composition notation P_inter ∘ {P^(h)_intra}_{h=1}^H is terse; presenting the block structure as in Eq. (24) would make the construction of P_attn unambiguous.
  4. [Fig. 5] The figure reports single-curve loss landscapes without error bars or repeated-seed variability; since the paper makes comparative claims about interpolation quality, some measure of variance would strengthen the presentation.
  5. [Table 1] The baseline name 'GiT Re-Basin' is a typo and should be 'Git Re-Basin' throughout the tables.

Circularity Check

1 steps flagged · score 4.0 of 10

Residual-block 'functional equivalence' is enforced by redefining the skip connection; the task-vector transport itself is not fitted to the evaluation data.

  1. self definitional [Section 3.1, Step 3 (Managing of Residual Connections), Eq. (12)]
    "To enforce this consistency, we redefine the identity mapping made by the residual connection. We replace it with a composition, Ii = PW0 P ⊤ in , consisting of two permutations — one to address issue I and another for issue II — as follows: zi = PW0 zattn + IiPinx = PW0 zattn + PW0 x, (12)"

    The derivation needs the transformed residual to satisfy z_i = P_W0 z_attn + P_W0 x. A genuine permutation reparameterization of z_i = z_attn + x leaves the identity skip untouched and applies the same permutation to both addends. Instead, the paper inserts a non-identity linear map I_i = P_W0 P_in^T into the skip path, changing the architecture to make the equation hold. The claimed functional equivalence of π(θ_A) is thus true by definition of the new skip connection, not derived from a symmetry of the original residual block. Since this equivalence underpins the use of π as a valid permutation transport for τ, the residual-handling step reduces to its own construction; Theorem 3.1 covers only the multi-head attention module, not this residual redefinition.

full rationale

The core task-vector experiment is not circular: π is computed solely from the two pretrained checkpoints θA and θB via spectral inter-head matching and per-head LAPs; no downstream labels, task-vector entries, or support-set accuracies are used to select π. The scaling α is fixed at 1 in the main tables and swept only as sensitivity analysis, so α is not fitted to the benchmark. The comparisons against Git Re-Basin, Optimal Transport, and vanilla transport are genuine external evaluations. The permutation-invariance of singular values is a standard mathematical fact cited to external literature. Self-citations (Crisostomi et al. 2024; Porrello et al. 2025) appear only in related work and are not load-bearing. The Discussion's admission that the source expert must perform strongly on the target task is a scope caveat, not a circularity. However, the residual-handling step in Sec. 3.1 is self-definitional: the identity skip is replaced by I_i = P_W0 P_in^T to force z_i = P_W0 z_attn + P_W0 x, making the 'functional equivalence' of the transformed model a matter of definition rather than a consequence of the original architecture's symmetries. The LayerNorm simplification ('can be regarded as a standard linear projection') is an unsupported assumption affecting validity, but it is not circularity. Overall the central empirical transport result retains independent content, so the score is moderate.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The core method introduces no new physical entities. It relies on one questionable simplification (LayerNorm as linear), one unproven domain assumption about task-vector alignment, and the standard permutation-symmetry machinery. The single free parameter is the task-vector scaling, fixed at 1 in the main results.

free parameters (1)
  • task vector scaling alpha = 1 (fixed in main results; swept in sensitivity analysis)
    Eq. (14) uses alpha to scale the transported task vector; it is set to 1 in the main tables, not fitted to the target data. Sensitivity analysis in Fig. 4 shows performance varies with alpha.
assumptions (4)
  • ad hoc to paper LayerNorm can be treated as a standard linear projection for permutation bookkeeping
    Sec. 3.1, Step 3: 'we omit layer normalization as it can be regarded as a standard linear projection'. Real ViT-B/16 uses LayerNorm; the true function is nonlinear and data-dependent, so functional equivalence may be broken.
  • standard math ReLU commutes with permutation symmetries
    Used implicitly in the residual derivations in Appendix A.3; because ReLU is elementwise, P ReLU(x) equals ReLU(P x) for any permutation P.
  • standard math Singular values are invariant to row and column permutations of a head matrix
    Proved in Proposition A.1; used in Eq. (7) for inter-head matching so that the ordering of units inside a head does not affect the distance.
  • domain assumption The permutation computed on base weights also aligns the fine-tuned model's task vector to theta_B
    Sec. 3.2 applies pi to the task vector tau assuming the task vector inherits the same permutation-symmetry structure as the base weights; this is not proven and is supported only empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Update Your Transformer to the Latest Release: Re-Basin of Task Vectors." pith.science (2026). https://pith.science/paper/S5DO3HAK

@misc{pith2026250522697,
  author       = {Pith},
  title        = {Pith review of: Update Your Transformer to the Latest Release: Re-Basin of Task Vectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S5DO3HAK}},
  note         = {Machine review of arXiv:2505.22697}
}
read the original abstract

Foundation models serve as the backbone for numerous specialized models developed through fine-tuning. However, when the underlying pretrained model is updated or retrained (e.g., on larger and more curated datasets), the fine-tuned model becomes obsolete, losing its utility and requiring retraining. This raises the question: is it possible to transfer fine-tuning to a new release of the model? In this work, we investigate how to transfer fine-tuning to a new checkpoint without having to re-train, in a data-free manner. To do so, we draw principles from model re-basin and provide a recipe based on weight permutations to re-base the modifications made to the original base model, often called task vector. In particular, our approach tailors model re-basin for Transformer models, taking into account the challenges of residual connections and multi-head attention layers. Specifically, we propose a two-level method rooted in spectral theory, initially permuting the attention heads and subsequently adjusting parameters within select pairs of heads. Through extensive experiments on visual and textual tasks, we achieve the seamless transfer of fine-tuned knowledge to new pre-trained backbones without relying on a single training step or datapoint. Code is available at https://github.com/aimagelab/TransFusion.

Figures

Figures reproduced from arXiv: 2505.22697 by the authors.

Figure 1
Figure 1. Transporting task vector τ from a fine-tuned base model θ f t A = θA + τ to a new release θB. ably decreasing the amount of data and training time needed to tailor models for specific downstream tasks. For this reason, pre-trained backbones such as OpenAI’s CLIP (Rad￾ford et al., 2021) are being extensively utilized as base foun￾dation models. As a result, the corresponding fine-tuned versions play a crucial role in… view at source ↗
Figure 2
Figure 2. Inter- (Step 1) and intra-head alignment (Step 2). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Residu￾als block and per￾mutations. To clarify the interaction between permutations in residual blocks, con￾sider [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Zero-shot gain/drop relative to θB of naive θB + ατ (blue) and our strategy θB + απ(τ ) (red) varying α. NLP Classification Tasks. Herein, we investigate a different setting that involves closed-vocabulary text clas￾sification — specifically, a set of tasks from the GL…
Figure 5
Figure 5. Figure 5: Loss values on CIFAR-10 test set during model [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Zero-shot gain/drop relative to θB of naive θB + ατ (blue) and our strategy θB + απ(τ ) (red) varying α. A.4. Extended Comparison on the Application of the Task Vector In [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. NNiT: Width-Agnostic Neural Network Generation with Structurally Aligned Weight Spaces

    cs.LG 2026-02 conditional novelty 6.0 of 10

    NNiT, a diffusion transformer that tokenizes weight matrices into patches, generates functional MLP policies for ManiSkill3 tasks at architecture topologies unseen during training, outperforming vector-based baselines.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Git Re-Basin: Merging Models modulo Permutation Symmetries

    Ainsworth, S., Hayase, J., and Srinivasa, S. Git Re-Basin: Merging Models modulo Permutation Symmetries . In International Conference on Learning Representations, 2023

    work page 2023

  3. [3]

    Reproducible scaling laws for contrastive language-image learning

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning . In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2023

    work page 2023

  4. [4]

    Fusing finetuned models for better pretraining

    Choshen, L., Venezian, E., Slonim, N., and Katz, Y. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022

    arXiv 2022

  5. [5]

    Describing Textures in the Wild

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and Vedaldi, A. Describing Textures in the Wild . In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014

    work page 2014

  6. [6]

    C 2 M 3 : Cycle-Consistent Multi-Model Merging

    Crisostomi, D., Fumero, M., Baieri, D., Bernard, F., and Rodola, E. C 2 M 3 : Cycle-Consistent Multi-Model Merging . Advances in Neural Information Processing Systems, 2024

    work page 2024

  7. [7]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale . In International Conference on Learning Representations, 2021

    work page 2021

  8. [8]

    Essentially No Barriers in Neural Network Energy Landscape

    Draxler, F., Veschgini, K., Salmhofer, M., and Hamprecht, F. Essentially No Barriers in Neural Network Energy Landscape . In International Conference on Machine Learning, 2018

    work page 2018

Show all 45 references
  1. [9]

    The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks

    Entezari, R., Sedghi, H., Saukh, O., and Neyshabur, B. The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks . In International Conference on Learning Representations, 2022

    2022

  2. [10]

    K., Roy, D., and Carbin, M

    Frankle, J., Dziugaite, G. K., Roy, D., and Carbin, M. Linear Mode Connectivity and the Lottery Ticket Hypothesis . In International Conference on Machine Learning, 2020

    2020

  3. [11]

    Freeman, C. D. and Bruna, J. Topology and Geometry of Half-Rectified Network Optimization . In International Conference on Learning Representations, 2017

    2017

  4. [12]

    French, R. M. Catastrophic forgetting in connectionist networks . Trends in Cognitive Sciences, 3 0 (4), 1999

    1999

  5. [13]

    Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al

    Gadre, S. Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al. Datacomp: In search of the next generation of multimodal datasets . Advances in Neural Information Processing Systems, 2024

    2024

  6. [14]

    P., and Wilson, A

    Garipov, T., Izmailov, P., Podoprikhin, D., Vetrov, D. P., and Wilson, A. G. Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs . Advances in Neural Information Processing Systems, 2018

    2018

  7. [15]

    EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification

    Helber, P., Bischke, B., Dengel, A., and Borth, D. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification . IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12 0 (7), 2019

    2019

  8. [16]

    The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization

    Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., and Gilmer, J. The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization . IEEE International Conference o...

    2021

  9. [17]

    OpenCLIP , 2021

    Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., and Schmidt, L. OpenCLIP , 2021

    2021

  10. [18]

    Y., Song, S., Hajishirzi, H., Kornblith, S., Farhadi, A., and Schmidt, L

    Ilharco, G., Wortsman, M., Gadre, S. Y., Song, S., Hajishirzi, H., Kornblith, S., Farhadi, A., and Schmidt, L. Patching open-vocabulary models by interpolating weights . Advances in Neural Information Processing Systems, 2022

    2022

  11. [19]

    T., Wortsman, M., Schmidt, L., Hajishirzi, H., and Farhadi, A

    Ilharco, G., Ribeiro, M. T., Wortsman, M., Schmidt, L., Hajishirzi, H., and Farhadi, A. Editing Models with Task Arithmetic . In International Conference on Learning Representations, 2023

    2023

  12. [20]

    Imfeld, M., Graldi, J., Giordano, M., Hofmann, T., Anagnostidis, S., and Singh, S. P. Transformer Fusion with Optimal Transport . In International Conference on Learning Representations, 2024

    2024

  13. [21]

    Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., and Wilson, A. G. Averaging Weights Leads to Wider Optima and Better Generalization . arXiv preprint arXiv:1803.05407, 2018

    2018 arXiv

  14. [22]

    and Volgenant, T

    Jonker, R. and Volgenant, T. A shortest augmenting path algorithm for dense and sparse linear assignment problems . In Papers of the 16th Annual Meeting of DGOR in Cooperation with NSOR/Vortr \"a ge der 16. Jahrestagung der DGOR zusammen mit der NSOR , 1988

    1988

  15. [23]

    REPAIR: REnormalizing Permuted Activations for Interpolation Repair

    Jordan, K., Sedghi, H., Saukh, O., Entezari, R., and Neyshabur, B. REPAIR: REnormalizing Permuted Activations for Interpolation Repair . In International Conference on Learning Representations, 2023

    2023

  16. [24]

    and Stani \'c , Z

    Jovanovi \'c , I. and Stani \'c , Z. Spectral distances of graphs . Linear Algebra and its Applications, 436 0 (5), 2012

    2012

  17. [25]

    Learning Multiple Layers of Features from Tiny Images

    Krizhevsky, A., Hinton, G., et al. Learning Multiple Layers of Features from Tiny Images . Technical Report, University of Toronto, 2009

    2009

  18. [26]

    Y., Chen, B., Williamson, D

    Lu, M. Y., Chen, B., Williamson, D. F., Chen, R. J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Le, L. P., Gerber, G., et al. A visual-language foundation model for computational pathology . Nature Medicine, 30, 2024

    2024

  19. [27]

    P., Liu, M

    Mall, U., Phoo, C. P., Liu, M. K., Vondrick, C., Hariharan, B., and Bala, K. Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment . In International Conference on Learning Representations, 2024

    2024

  20. [28]

    Matena, M. S. and Raffel, C. A. Merging Models with Fisher-Weighted Averaging . Advances in Neural Information Processing Systems, 35, 2022

    2022

  21. [29]

    and Cohen, N

    McCloskey, M. and Cohen, N. J. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem . In Psychology of learning and motivation, volume 24, pp.\ 109--165. Academic Press, 1989

    1989

  22. [30]

    Equivariant Architectures for Learning in Deep Weight Spaces

    Navon, A., Shamsian, A., Achituve, I., Fetaya, E., Chechik, G., and Maron, H. Equivariant Architectures for Learning in Deep Weight Spaces . In International Conference on Machine Learning, 2023

    2023

  23. [31]

    Y., et al

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A. Y., et al. Reading Digits in Natural Images with Unsupervised Feature Learning . In Neural Information Processing Systems Workshops. Granada, 2011

    2011

  24. [32]

    Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models

    Ortiz-Jimenez, G., Favero, A., and Frossard, P. Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models . Advances in Neural Information Processing Systems, 2024

    2024

  25. [33]

    Pe \ n a, F. A. G., Medeiros, H. R., Dubail, T., Aminbeidokhti, M., Granger, E., and Pedersoli, M. Re-basin via implicit Sinkhorn differentiation . In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2023

    2023

  26. [34]

    A Second-Order Perspective on Model Compositionality and Incremental Learning

    Porrello, A., Bonicelli, L., Buzzega, P., Millunzi, M., Calderara, S., and Cucchiara, R. A Second-Order Perspective on Model Compositionality and Incremental Learning . In International Conference on Learning Representations, 2025

    2025

  27. [35]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning Transferable Visual Models From Natural Language Supervision . In International Conference on Machine Learning, 2021

    2021

  28. [36]

    Diverse Weight Averaging for Out-of-Distribution Generalization

    Rame, A., Kirchmeyer, M., Rahier, T., Rakotomamonjy, A., patrick gallinari, and Cord, M. Diverse Weight Averaging for Out-of-Distribution Generalization . In Advances in Neural Information Processing Systems, 2022

    2022

  29. [37]

    Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalization

    Ram \'e , A., Ahuja, K., Zhang, J., Cord, M., Bottou, L., and Lopez-Paz, D. Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalization . In International Conference on Machine Learning, 2023

    2023

  30. [38]

    Singh, S. P. and Jaggi, M. Model Fusion via Optimal Transport . Advances in Neural Information Processing Systems, 2020

    2020

  31. [39]

    The German Traffic Sign Recognition Benchmark: A multi-class classification competition

    Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C. The German Traffic Sign Recognition Benchmark: A multi-class classification competition . In The 2011 international joint conference on neural networks. IEEE, 2011

    2011

  32. [40]

    ZipIt! Merging Models from Different Tasks without Training

    Stoica, G., Bolya, D., Bjorner, J., Ramesh, P., Hearn, T., and Hoffman, J. ZipIt! Merging Models from Different Tasks without Training . In International Conference on Learning Representations, 2024

    2024

  33. [41]

    Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., and Bowman, S. R. GLUE A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding . In International Conference on Learning Representations, 2019

    2019

  34. [42]

    Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A

    Wortsman, M., Ilharco, G., Gadre, S. Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A. S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time . In Internatio...

    2022

  35. [43]

    W., Li, M., Kornblith, S., Roelofs, R., Lopes, R

    Wortsman, M., Ilharco, G., Kim, J. W., Li, M., Kornblith, S., Roelofs, R., Lopes, R. G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al. Robust fine-tuning of zero-shot models . In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2022 b

    2022

  36. [44]

    A., and Bansal, M

    Yadav, P., Tam, D., Choshen, L., Raffel, C. A., and Bansal, M. TIES-Merging: Resolving Interference When Merging Models . Advances in Neural Information Processing Systems, 36, 2024

    2024

  37. [45]

    Z., Albert, P., Rodriguez-Opazo, C., van den Hengel, A., and Abbasnejad, E

    Zhang, F. Z., Albert, P., Rodriguez-Opazo, C., van den Hengel, A., and Abbasnejad, E. Knowledge Composition using Task Vectors with Learned Anisotropic Scaling . Advances in Neural Information Processing Systems, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.