pith. sign in

arxiv: 2606.26902 · v1 · pith:5RA3KR23new · submitted 2026-06-25 · 💻 cs.AI

Learning to Recover Task Experts from a Multi-Task Merged Model

Jinwook Jung , Taegyu Kim , Kumju Jo , Sungyong Baik This is my paper

Pith reviewed 2026-06-26 04:58 UTC · model grok-4.3

classification 💻 cs.AI
keywords multi-task model mergingparameter interferencetask expert recoveryoffset predictionSVD subspace signaturesgeneralization to unseen tasksadaptive interpolation
0
0 comments X

The pith

By predicting additive offsets from a merged checkpoint, ReTeX recovers over 95 percent of each task expert's original performance from a single model.

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

Multi-task merging combines separate expert models into one but creates interference that degrades individual task performance. The paper shows this interference can be treated as predictable additive offsets on the original parameters. ReTeX trains a predictor to estimate those offsets and a router-free identifier that matches inputs to tasks via SVD subspace signatures computed in advance. The result is recovery of expert-level accuracy without storing or loading separate models, plus improved handling of tasks never seen during merging. The offset prediction also produces emergent interpolation that helps on out-of-distribution inputs.

Core claim

Parameter interference during merging is modeled as an affine transformation that reduces to an additive offset; a network trained to predict the offset from the merged weights alone can restore each original expert, while an offline SVD-based identifier selects the correct expert at inference without a learned router.

What carries the argument

ReTeX offset predictor that maps merged weights to task-specific additive corrections, paired with SVD subspace signatures for task identification.

If this is right

  • A single merged checkpoint supports near-expert accuracy on every training task without redundant storage.
  • The same checkpoint generalizes better to tasks absent from the original merge set.
  • Offset prediction produces automatic interpolation among seen experts when the input lies outside the training distribution.
  • Task identity at inference is obtained from projection residuals onto precomputed SVD subspaces rather than a separate router network.

Where Pith is reading between the lines

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

  • Deployment cost for multi-task systems could drop because only one set of weights needs to be kept in memory.
  • The SVD identifier might transfer to other merging algorithms that do not explicitly train routers.
  • If offset prediction scales to very large models, it could serve as a lightweight post-merge adaptation step.
  • The emergent interpolation behavior suggests the method might support zero-shot composition of task knowledge.

Load-bearing premise

The changes each expert undergoes when merged can be reversed by learning a simple additive correction that depends only on the final merged weights.

What would settle it

Measure whether the offset predictor still recovers 95 percent performance when the merging procedure is changed to a method whose perturbations deviate from the assumed affine form.

Figures

Figures reproduced from arXiv: 2606.26902 by Jinwook Jung, Kumju Jo, Sungyong Baik, Taegyu Kim.

Figure 1
Figure 1. Figure 1: Component requirements for multi-task deployment. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall ReTeX framework. (a) Task identification constructs a task mem￾ory bank by fitting a low-dimensional SVD subspace to merged-model representations for each task. Given a test representation z, ReTeX computes projection residuals {Rt(z)} T t=1 and outputs a task ID estimate tˆ by an argmin rule. (b) Task expert re￾covery conditions a task embedding on tˆ and predicts low-rank offsets across layers. A… view at source ↗
Figure 3
Figure 3. Figure 3: Performance and efficiency trade-offs of ReTeX. (a) Normalized accuracy across ranks r for three base merged models (WA, TA, TIES). (b) Normalized accuracy as a function of task embedding dimension demb for 8, 14, and 20 tasks with ViT-B/32 and fixed r=256. (c) Normalized accuracy vs. normalized memory for ReTeX and ReTeX-E at fixed outer rank r while varying rg. Normalized memory is computed as the storag… view at source ↗
read the original abstract

Multi-task model merging aims to consolidate several task-specific experts into a unified model, yet static merging consistently suffers from parameter interference. While dynamic merging models aim to bridge this gap, many works rely on the costly storage and loading of redundant expert components at inference. In this work, from the perspective of task expert, we view parameter interference as parameter perturbation introduced to each expert during merging process. We show that such parameter perturbations can be modeled as affine transformation, which can be approximated as additive offsets. Motivated by these, we propose Recover Task eXpert (ReTeX), a framework that predicts those offsets, in order to undo parameter interference and recover task-expert performance from a single merged checkpoint. To recover the appropriate expert when task identity is unknown, we introduce a router-free task identifier based on SVD subspace signatures computed offline before inference. At inference, the identifier selects the task whose subspace yields the smallest projection residual for a given input. As a result, ReTeX recovers over 95% of individual-expert performance in both vision and NLP domains, while significantly improving generalization to unseen tasks. Crucially, we also show that the parameter offset prediction leads to emergent adaptive interpolation of expert knowledge for out-of-distribution (OOD) tasks. ReTeX adaptively interpolates seen expert knowledge to handle unseen tasks. Our code is available at https://github.com/BAIKLAB/ReTeX

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces ReTeX, a framework that models merging-induced parameter perturbations on task experts as affine transformations approximable by additive offsets, predicts these offsets from the merged checkpoint alone to recover individual expert performance, and employs an SVD-based subspace signature method for router-free task identification at inference. It reports recovering over 95% of per-expert performance on vision and NLP tasks while achieving improved generalization to unseen tasks via emergent adaptive interpolation of expert knowledge.

Significance. If the affine-to-additive modeling assumption is validated with quantitative bounds and the recovery results prove robust across baselines and tasks, the approach would offer a practical route to high-fidelity multi-task inference from a single checkpoint, reducing the storage overhead of dynamic merging methods while providing a mechanism for OOD task handling without retraining.

major comments (3)
  1. [Abstract] Abstract: the central modeling claim that merging perturbations can be represented as affine maps and then approximated as additive offsets is asserted without derivation, without a quantitative bound on residual error after the best additive fit, and without empirical validation of approximation quality; this assumption is load-bearing for the recovery performance claim.
  2. [Abstract] Abstract: the headline result of >95% recovery of individual-expert performance is stated without reference to the number of tasks, datasets, baselines, statistical error bars, or verification protocol, preventing assessment of whether the offset predictor genuinely undoes interference or merely learns a compensatory adapter.
  3. [Abstract] Abstract / method description: the SVD subspace signature task identifier is introduced at a high level with no detail on subspace construction, dimensionality, or how projection residuals are computed and thresholded; because this component is required for the router-free claim, its correctness directly affects the reported inference results.
minor comments (2)
  1. [Abstract] The abstract would benefit from a concise statement of the number of tasks, domains, and primary baselines used to support the 95% recovery figure.
  2. Notation for the merged weights, expert weights, and predicted offsets should be introduced consistently and early to aid readability of the recovery equations.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on the abstract and method presentation. We will revise the manuscript to incorporate the requested derivations, quantitative details, experimental specifics, and method expansions while preserving the core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central modeling claim that merging perturbations can be represented as affine maps and then approximated as additive offsets is asserted without derivation, without a quantitative bound on residual error after the best additive fit, and without empirical validation of approximation quality; this assumption is load-bearing for the recovery performance claim.

    Authors: We agree the abstract presents the claim at a high level. We will add an explicit derivation of the affine transformation model from the merging process, a quantitative bound on the residual error after the additive fit, and empirical validation of approximation quality to the revised manuscript (new subsection in Section 3), with a concise reference included in the abstract. revision: yes

  2. Referee: [Abstract] Abstract: the headline result of >95% recovery of individual-expert performance is stated without reference to the number of tasks, datasets, baselines, statistical error bars, or verification protocol, preventing assessment of whether the offset predictor genuinely undoes interference or merely learns a compensatory adapter.

    Authors: We will revise the abstract to specify the number of tasks and datasets, the baselines used, statistical error bars from repeated runs, and the verification protocol (performance measured against individually fine-tuned experts on held-out data). This will clarify the nature of the recovery. revision: yes

  3. Referee: [Abstract] Abstract / method description: the SVD subspace signature task identifier is introduced at a high level with no detail on subspace construction, dimensionality, or how projection residuals are computed and thresholded; because this component is required for the router-free claim, its correctness directly affects the reported inference results.

    Authors: We will expand the method description to detail subspace construction (via SVD on task-specific parameter matrices), chosen dimensionality, projection residual computation (norm of the component orthogonal to the subspace), and thresholding procedure. These specifics will be added to Section 3 and referenced in the abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper frames ReTeX as a predictive framework that learns to approximate merging-induced perturbations (modeled as affine maps and approximated by additive offsets) from the merged checkpoint alone, then uses an independent SVD-based task identifier at inference. No equations or steps reduce a claimed prediction to a fitted quantity by construction, nor does any load-bearing premise rest on self-citation chains or imported uniqueness results. The offset-prediction step is presented as a learned mapping whose success is evaluated against held-out expert performance rather than being definitionally equivalent to its inputs. The derivation therefore retains independent empirical content.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The approach rests on the domain assumption that merging interference takes the form of predictable additive offsets; no free parameters or invented entities are explicitly quantified in the abstract.

axioms (1)
  • domain assumption Parameter perturbations during merging can be modeled as affine transformations approximable as additive offsets
    Directly stated as the motivation for offset prediction in the abstract.
invented entities (1)
  • recoverable parameter offsets no independent evidence
    purpose: Undo interference to restore expert performance from merged checkpoint
    Introduced as the target of the prediction network; no independent evidence outside the method itself.

pith-pipeline@v0.9.1-grok · 5791 in / 1198 out tokens · 23090 ms · 2026-06-26T04:58:57.080153+00:00 · methodology

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

103 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    In: ICCCSP (2021)

    Ahmed, M.I., Mamun, S.M., Asif, A.U.Z.: Dcnn-based vegetable image classifica- tion using transfer learning: A comparative study. In: ICCCSP (2021)

    2021

  3. [3]

    Data in Brief (2023)

    Ahmed, S.I., Ibrahim, M., Nadim, M., Rahman, M.M., Shejunti, M.M., Jabid, T., Ali, M.S.: Mangoleafbd: A comprehensive image dataset to classify diseased and healthy mango leaves. Data in Brief (2023)

    2023

  4. [4]

    Alessio, C.: Animals-10.https : / / www . kaggle . com / datasets / alessiocorrado99/animals10

  5. [5]

    In: ICCV (2021)

    Baik, S., Choi, J., Kim, H., Cho, D., Min, J., Lee, K.M.: Meta-learning with task-adaptive loss function for few-shot learning. In: ICCV (2021)

    2021

  6. [6]

    In: NeurIPS (2020)

    Baik, S., Choi, M., Choi, J., Kim, H., Lee, K.M.: Meta-learning with adaptive hyperparameters. In: NeurIPS (2020)

    2020

  7. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Baik, S., Choi, M., Choi, J., Kim, H., Lee, K.M.: Learning to learn task-adaptive hyperparameters for few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  8. [8]

    In: CVPR (2020)

    Baik, S., Hong, S., Lee, K.M.: Learning to forget for meta-learning. In: CVPR (2020)

    2020

  9. [9]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    Baik, S., Hong, S., Lee, K.M.: Learning to forget for meta-learning via task- and-layer-wise attenuation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    2022

  10. [10]

    Available on https://www

    Bansal, P.: Intel image classification. Available on https://www. kaggle. com/puneet6060/intel-image-classification, Online (2019)

    2019

  11. [11]

    In: WACV (2022)

    Bateni, P., Barber, J., Van de Meent, J.W., Wood, F.: Enhancing few-shot image classification with unlabelled examples. In: WACV (2022)

    2022

  12. [12]

    In: CVPR (2020)

    Bateni, P., Goyal, R., Masrani, V., Wood, F., Sigal, L.: Improved few-shot visual classification. In: CVPR (2020)

    2020

  13. [13]

    In: ECCV (2014)

    Bossard, L., Guillaumin, M., Gool, L.V.: Food-101 – mining discriminative com- ponents with random forests. In: ECCV (2014)

    2014

  14. [14]

    https://doi.org/10.34740/KAGGLE/DS/81794

    CCHANG: Garbage classification.https://www.kaggle.com/ds/81794(2018). https://doi.org/10.34740/KAGGLE/DS/81794

    work page doi:10.34740/kaggle/ds/81794 2018

  15. [15]

    Proceedings of the IEEE (2017)

    Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE (2017)

    2017

  16. [16]

    In: CVPR

    Chi, Y., Porikli, F.: Connecting the dots in multi-class classification: From nearest subspace to collaborative representation. In: CVPR. pp. 3602–3609 (2012)

    2012

  17. [17]

    IEEE Access (2022)

    Choi, J., Baik, S., Choi, M., Kwon, J., Lee, K.M.: Visual tracking by adaptive continual meta-learning. IEEE Access (2022)

    2022

  18. [18]

    In: CVPR (2020)

    Choi, M., Choi, J., Baik, S., Kim, T.H., Lee, K.M.: Scene-adaptive video frame interpolation via meta-learning. In: CVPR (2020)

    2020

  19. [19]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    Choi, M., Choi, J., Baik, S., Kim, T.H., Lee, K.M.: Test-time adaptation for video frame interpolation via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    2022

  20. [20]

    In: CVPR (2020)

    Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: CVPR (2020)

    2020

  21. [21]

    In: CVPR (2014)

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR (2014)

    2014

  22. [22]

    In: NeurIPS (2018)

    Clanuwat, T., Bober-Irizar, M., Kitamoto, A., Lamb, A., Yamamoto, K., Ha, D.: Deep learning for classical japanese literature. In: NeurIPS (2018)

    2018

  23. [23]

    In: AISTATS (2011)

    Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: AISTATS (2011)

    2011

  24. [24]

    In: IJCNN (2017)

    Cohen, G., Afshar, S., Tapson, J., van Schaik, A.: Emnist: Extending mnist to handwritten letters. In: IJCNN (2017)

    2017

  25. [25]

    Journal of Diabetes Science and Technology (2009)

    Cuadros, J., Bresnick, G.: Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology (2009)

    2009

  26. [26]

    cats, 2013

    Cukierski, W.: Dogs vs. cats, 2013. URL https://kaggle.com/competitions/dogs- vs-cats

    2013

  27. [27]

    In: ECCV (2024)

    Davari, M., Belilovsky, E.: Model breadcrumbs: Scaling multi-task model merging with sparse masks. In: ECCV (2024)

    2024

  28. [28]

    DeepNets: Landscape recognition.https : / / www . kaggle . com / datasets / utkarshsaxenadn/landscape-recognition-image-dataset-12k-images

  29. [29]

    IEEE signal processing magazine (2012)

    Deng, L.: The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine (2012)

    2012

  30. [30]

    Nature Machine Intelligence (2023)

    Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence (2023)

    2023

  31. [31]

    Computer Vision and Image Understanding (2007)

    Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object cate- gories. Computer Vision and Image Understanding (2007)

    2007

  32. [32]

    In: ICLR (2017)

    Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: ICLR (2017)

    2017

  33. [33]

    In: CVPR (2025)

    Gargiulo, A.A., Crisostomi, D., Bucarelli, M.S., Scardapane, S., Silvestri, F., Rodolà, E.: Task singular vectors: Reducing task interference in model merging. In: CVPR (2025)

    2025

  34. [34]

    In: ICONIP (2013)

    Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., Zhou, Y., Ramaiah, C., Feng, F., Li, R., Wang, X., Athanasakis, D., Shawe-Taylor, J., Milakov, M., Park, J., Ionescu, R., Popescu, M., Grozea, C., Bergstra, J., Xie, J., Romaszko, L., Xu, B., Chuang, Z., Bengio, Y.: Challenges ...

    2013

  35. [35]

    Griffin, G., Holub, A., Perona, P.: Caltech-256 object category dataset. Tech. rep., California Institute of Technology (2007)

    2007

  36. [36]

    In: ICLR (2017)

    Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: ICLR (2017)

    2017

  37. [37]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2019)

    Helber, P., Bischke, B., Dengel, A., 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 (2019)

    2019

  38. [38]

    In: CVPR (2015)

    Horn, G.V., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: CVPR (2015)

    2015

  39. [39]

    Howard, J.: Imagenette: A smaller subset of 10 easily classified classes from ima- genet.https://github.com/fastai/imagenette(2019)

    2019

  40. [40]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  41. [41]

    In: NeurIPS (2024)

    Huang, C., Ye, P., Chen, T., He, T., Yue, X., Ouyang, W.: Emr-merging: Tuning- free high-performance model merging. In: NeurIPS (2024)

    2024

  42. [42]

    In: ICLR (2023)

    Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023)

    2023

  43. [43]

    In: ICLR (2025)

    Iurada, L., Ciccone, M., Tommasi, T.: Efficient model editing with task-localized sparse fine-tuning. In: ICLR (2025)

    2025

  44. [44]

    In: ICLR (2023)

    Jin, X., Ren, X., Preotiuc-Pietro, D., Cheng, P.: Dataless knowledge fusion by merging weights of language models. In: ICLR (2023)

    2023

  45. [45]

    In: ICLR (2025)

    Kang, J., Karlinsky, L., Luo, H., Wang, Z., Hansen, J., Glass, J., Cox, D., Panda, R., Feris, R., Ritter, A.: Self-moe: Towards compositional large language models with self-specialized experts. In: ICLR (2025)

    2025

  46. [46]

    Cell (2018)

    Kermany,D.S.,Goldbaum,M.,Cai,W.,Valentim,C.C.S.,Liang,H.,Baxter,S.L., McKeown, A., Yang, G., Wu, X., Yan, F., Dong, J., Prasadha, M., Pei, J., Ting, M., Zhu, J., Li, C., Hewett, S., Dong, J., Ziyar, I., Shi, A., Zhang, R., Gupta, K., Wong, R.M.Y., Lam, L.A.T.C., Cheung, J., Tsoi, M., Wu, V., Yan, C., Huang, C., Lee, D.H., Zhang, Y., Wong, J.W., Li, K.M., ...

    2018

  47. [47]

    In: CVPR (2011)

    Khosla, A., Jayadevaprakash, N., Yao, B., Fei-Fei, L.: Novel dataset for fine- grained image categorization. In: CVPR (2011)

    2011

  48. [48]

    In: AAAI (2020)

    Khot, T., Clark, P., Guerquin, M., Jansen, P., Sabharwal, A.: Qasc: A dataset for question answering via sentence composition. In: AAAI (2020)

    2020

  49. [49]

    In: ICLR (2015)

    Kingma, D.P.: Adam: A method for stochastic optimization. In: ICLR (2015)

    2015

  50. [50]

    In: ICCV Workshop (2013)

    Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine- grained categorization. In: ICCV Workshop (2013)

    2013

  51. [51]

    Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)

    2009

  52. [52]

    In: ECCV (2012)

    Kumar, N., Belhumeur, P.N., Biswas, A., Jacobs, D.W., Kress, W.J., Lopez, I.C., Soares, J.V.B.: Leafsnap: A computer vision system for automatic plant species identification. In: ECCV (2012)

    2012

  53. [53]

    com / AI - Lab - Makerere/ibean/

    Lab, M.A.: Bean disease dataset (2020),https : / / github . com / AI - Lab - Makerere/ibean/

    2020

  54. [54]

    In: ICLR (2025)

    Lee, Y., Jung, J., Baik, S.: Mitigating parameter interference in model merging via sharpness-aware fine-tuning. In: ICLR (2025)

    2025

  55. [55]

    In: KR (2012)

    Levesque, H., Davis, E., Morgenstern, L.: The winograd schema challenge. In: KR (2012)

    2012

  56. [56]

    In: ICLR (2024)

    Li, P., Zhang, Z., Yadav, P., Sung, Y.L., Cheng, Y., Bansal, M., Chen, T.: Merge, then compress: Demystify efficient smoe with hints from its routing policy. In: ICLR (2024)

    2024

  57. [57]

    In: CVPR (2017)

    Li, S., Deng, W., Du, J.: Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In: CVPR (2017)

    2017

  58. [58]

    In: CVPR (2022)

    Li, W.H., Liu, X., Bilen, H.: Cross-domain few-shot learning with task-specific adapters. In: CVPR (2022)

    2022

  59. [59]

    arXiv preprint arXiv:2206.11404 (2022)

    Liao, P., Li, X., Liu, X., Keutzer, K.: The artbench dataset: Benchmarking gen- erative models with artworks. arXiv preprint arXiv:2206.11404 (2022)

    arXiv 2022

  60. [60]

    In: NeurIPS (2024)

    Lu, Z., Fan, C., Wei, W., Qu, X., Chen, D., Cheng, Y.: Twin-merging: Dynamic integration of modular expertise in model merging. In: NeurIPS (2024)

    2024

  61. [61]

    arXiv preprint arXiv:1306.5151 (2013)

    Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)

    Pith/arXiv arXiv 2013

  62. [62]

    and Raffel, C.A.: Merging models with fisher-weighted aver- aging

    Matena, Michael S. and Raffel, C.A.: Merging models with fisher-weighted aver- aging. In: NeurIPS (2022)

    2022

  63. [63]

    Frontiers in Plant Science (2016)

    Mohanty, S.P., Hughes, D.P., Salathé, M.: Using deep learning for image-based plant disease detection. Frontiers in Plant Science (2016)

    2016

  64. [64]

    Transactions on Machine Learning Research (2023)

    Muqeeth, M., Liu, H., Raffel, C.: Soft merging of experts with adaptive routing. Transactions on Machine Learning Research (2023)

    2023

  65. [65]

    Acta Universitatis Sapientiae, Informatica (2018)

    Muresan, H., Oltean, M.: Fruit recognition from images using deep learning. Acta Universitatis Sapientiae, Informatica (2018)

    2018

  66. [66]

    Nagaraj, A.: Asl alphabet.https://www.kaggle.com/datasets/grassknoted/ asl-alphabet

  67. [67]

    In: NIPS Workshop (2011)

    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: NIPS Workshop (2011)

    2011

  68. [68]

    In: ICVGIP (2008)

    Nilsback, M.E., Zisserman, A.: Automated flower classification over a large num- ber of classes. In: ICVGIP (2008)

    2008

  69. [69]

    In: ICLR (2025)

    Oh, C., Li, Y., Song, K., Yun, S., Han, D.: Dawin: Training-free dynamic weight interpolation for robust adaptation. In: ICLR (2025)

    2025

  70. [70]

    Scientific Reports (2019)

    Olsen, A., Konovalov, D.A., Philippa, B., Ridd, P., Wood, J.C., Johns, J., Banks, W.,Girgenti,B.,Kenny,O.,Whinney,J.,Calvert,B.,RahimiAzghadi,M.,White, R.D.: Deepweeds: A multiclass weed species image dataset for deep learning. Scientific Reports (2019)

    2019

  71. [71]

    In: NeurIPS (2023)

    Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. In: NeurIPS (2023)

    2023

  72. [72]

    In: CVPR (2012)

    Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.V.: Cats and dogs. In: CVPR (2012)

    2012

  73. [73]

    In: AAAI (2018)

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018)

    2018

  74. [74]

    In: ACM MMSys (2017)

    Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: ACM MMSys (2017)

    2017

  75. [75]

    In: CVPR (2009)

    Quattoni, A., Torralba, A.: Recognizing indoor scenes. In: CVPR (2009)

    2009

  76. [76]

    In: ICML (2021)

    Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G.,Agarwal,S.,Sastry,G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  77. [77]

    In: NeurIPS (2019)

    Requeima, J., Gordon, J., Bronskill, J., Nowozin, S., Turner, R.E.: Fast and flexible multi-task classification using conditional neural adaptive processes. In: NeurIPS (2019)

    2019

  78. [78]

    arXiv preprint arXiv:2404.18416 (2024)

    Saab, K., Tu, T., Weng, W.H., Tanno, R., Stutz, D., Wulczyn, E., Zhang, F., Strother,T.,Park,C.,Vedadi,E.,etal.:Capabilitiesofgeminimodelsinmedicine. arXiv preprint arXiv:2404.18416 (2024)

    Pith/arXiv arXiv 2024

  79. [79]

    In: AAAI (2020)

    Sakaguchi, K., Bras, R.L., Bhagavatula, C., Choi, Y.: Winogrande: An adversarial winograd schema challenge at scale. In: AAAI (2020)

    2020

  80. [80]

    In: ACL (2018)

    Sharma, R., Allen, J., Bakhshandeh, O., Mostafazadeh, N.: Tackling the story ending biases in the story cloze test. In: ACL (2018)

    2018

Showing first 80 references.