Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Winner-take-all merging beats averaging in medical LVLMs by keeping each LoRA parameter from its strongest expert.

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 ↗

A per-parameter winner-take-all merge beats averaging and subspace methods on a new 16-model, 8-modality medical LVLM benchmark.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuinely useful medical-LoRA merging benchmark wrapped around a simple winner-take-all rule that wins on average but whose stated rationale — and 'consistently' — outruns the evidence. the 3 major comments →

arxiv 2607.15661 v1 pith:T4JRGUD4 submitted 2026-07-17 cs.CV

Model Merging for Medical LVLMs: A Benchmark and a Winner-Take-All Approach

Lichao Mou , Shilan Zhang , Chunlei Li , Bingcong Yan , Jingliang Hu , Yilei Shi , Shengwu Xiong , Xiao Xiang Zhu
show 2 more authors
Lei Li Yaxiong Chen
This is my paper
classification cs.CV
keywords model mergingLoRAmedical vision-language modelsparameter-efficient fine-tuningmulti-task learningwinner-take-allbenchmarkMergeMedBench
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 reading

The paper argues that when several medical vision-language models are fine-tuned with LoRA on different imaging modalities, the best way to merge them into one model is not to average their updates but, at every position in the low-rank matrices, to keep the value from whichever expert has the largest normalized magnitude. It supports this with MergeMedBench, a new benchmark of 16 LoRA-fine-tuned experts across eight imaging modalities and two model families, where winner-take-all outperforms averaging, sign-consensus, and subspace-alignment baselines on both architectures. The reason offered is that a small set of dominant LoRA parameters governs each expert's behavior and is highly sensitive to perturbation, so preserving them exactly avoids the information dilution caused by averaging. If correct, this gives a hyperparameter-free, computationally cheap merging recipe that makes deploying a single multi-task medical LVLM practical.

Core claim

On its own terms, the paper establishes winner-take-all as the strongest LoRA merge rule on MergeMedBench. For each entry of the LoRA A and B matrices, the merged value is taken directly from the expert whose z-scored magnitude is largest; no averaging, sign voting, or subspace transform is applied. Reported results are 90.88% average normalized accuracy on Qwen3-VL-2B-Instruct and 84.64% on InternVL2-1B, ahead of all listed baselines and at lower runtime than most. The paper also shows that the rule degrades as k rises when taking the k-th ranked expert instead of the top one, that it clearly beats plain averaging (90.88% vs. 72.18%), and that it can be plugged into subspace methods such as

What carries the argument

The winner-take-all rule itself: a scoring function based on z-score normalization of each LoRA matrix, a coordinate-wise argmax over experts, and direct assignment of the winning value, yielding the merged low-rank update ΔW = B_merged · A_merged. The mechanism it exploits is the empirically demonstrated dominant-parameter phenomenon: rescaling high-magnitude LoRA entries severely degrades task performance while rescaling low-magnitude entries barely matters, implying that merging should preserve the former exactly rather than average them.

Load-bearing premise

The method assumes that after per-expert normalization, the entry with the largest magnitude at each position is the one most worth keeping, and that picking A and B winners independently still yields a coherent low-rank update.

What would settle it

A direct test: shuffle the coordinate-wise winner choices, or replace the z-score with a random permutation of the same scores, while keeping the same merged parameter values; if accuracy stays near 90%, the dominance ordering is not what drives the result. Alternatively, on the same benchmark, swap the z-score for rank-based selection; if rank-based selection performs as well, the assumption that normalized magnitude measures task relevance is not the operative cause.

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

If this is right

  • A single merged medical LVLM can plausibly replace a set of modality-specific experts, cutting deployment cost while retaining most per-modality accuracy.
  • The method is hyperparameter-free and fast (about 0.2 seconds in the paper's timing), making it a practical default baseline for future LoRA-merging research.
  • The dominant-parameter principle transfers to subspace-based merging: replacing averaging with winner-take-all inside Core and KnOTS improves their accuracy.
  • Because the rule operates on LoRA deltas, it applies to any LoRA fine-tuned model family, not only the two architectures tested.

Where Pith is reading between the lines

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

  • If the dominance phenomenon is general, the same winner-take-all rule may extend to fully fine-tuned delta parameters (task vectors), not just low-rank ones, though normalization would need careful design.
  • The z-score normalization embeds a strong assumption that expert scales are comparable; a natural stress test is replacing it with rank-based scores, which would reveal whether winner identity or magnitude scale drives the result.
  • The paper defines dominance by magnitude; a gradient-weighted score could identify different winners and potentially improve accuracy further, at the cost of needing task data.
  • The benchmark evaluates per-modality accuracy on a fixed split; extending to other medical tasks such as report generation would test whether the result generalizes beyond visual question answering.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The manuscript introduces MergeMedBench, a benchmark for merging LoRA-fine-tuned medical vision-language models. It curates 88,995 image-question-answer triplets from OmniMedVQA across eight imaging modalities and builds 16 expert LoRA models on two backbones (Qwen3-VL-2B-Instruct and InternVL2-1B). It evaluates a wide range of existing merging methods and proposes winner-take-all (WTA), which merges each LoRA A/B factor coordinate by taking the value from the expert with the largest z-scored magnitude. The paper reports that WTA achieves the best average accuracy on both backbones (90.88% and 84.64%) while being fast and hyperparameter-free.

Significance. If the empirical claims hold, MergeMedBench is a useful resource for the community and WTA is a strong, simple baseline for medical LoRA merging. The paper uses a fixed train/test split, compares many baselines, and reports runtime, which are strengths. However, the central mechanistic claim — that preserving the largest normalized factor entries preserves task-dominant parameters of the merged update — is not established, the 'consistently outperforms' statement is contradicted by per-modality results, and the design choices (k=1, normalization scheme) are selected on the same evaluation split used for the headline numbers. Without error bars or significance tests, the reported average-accuracy margins are not statistically supported. These issues are load-bearing for the paper's central claims.

major comments (3)
  1. [§3.3, Eqs. (2)–(5)] The method's premise is that the coordinate with the largest z-scored magnitude in A or B is the most task-relevant, so preserving it yields a better merged ΔW = B A. This is not entailed by the equations: a selected A[k,j] contributes only through the column B_merged[:,k]; because A and B winners are chosen independently, a coordinate can win as an outlier in an expert with small overall factor norms and contribute negligibly to the product, while a moderate coordinate from a large-norm expert that actually dominates ΔW loses. The z-score normalization explicitly removes per-expert scale. The §3.1/Fig. 2 'dominant parameter' analysis rescales entries of the factors, not entries of the product BA, so it does not establish that factor-level magnitude dominance tracks product-level dominance. Please add evidence directly about the merged update, e.g., compare WTA with selection based on en
  2. [§4.2, Tables 1–2; Abstract] The abstract and §4.2 state that WTA 'consistently outperforms' all baselines, but this is contradicted by per-modality numbers. For Qwen-VL, WTA is below DARE-TIES on ultrasound (91.51 vs. 96.05), below RobustMerge on MRI (96.30 vs. 96.39), and below STF on OCT (86.08 vs. 87.50). For InternVL2, WTA is below STF on CT (88.71 vs. 89.42), below Core-DARE-TIES on ultrasound (62.15 vs. 92.14), and below STF on fundus (92.26 vs. 93.26). The average-accuracy lead is real (90.88 vs. 89.81 for STF; 84.64 vs. 82.91 for STF), but no error bars, multiple seeds, or significance tests are reported. Please qualify the claim as average-accuracy superiority and add uncertainty quantification before making a 'consistently outperforms' statement.
  3. [§3.3, Tables 3 and 5] The paper calls WTA 'hyperparameter-free,' but the winner rank k=1 and the z-score normalization are selected using the same evaluation split used in the final Tables 1–2. Table 3 chooses k=1 after sweeping k=1..8, and Table 5 selects the normalization scheme. Because the final numbers come from the same split used for model selection, the reported margins are optimistic and the 'hyperparameter-free' label is misleading. Please either use a separate validation split or present these sweeps as exploratory, with the final method fixed independently of the test set.
minor comments (5)
  1. [§4.1] The text says 'We use normalized per-modality accuracy and average normalized accuracy,' but Tables 1–2 report raw values labeled 'accuracy (%)'. Please clarify which quantity is actually reported and whether the averages are computed on normalized or raw accuracies.
  2. [§3.1, footnote 1] The footnote promises to 'discuss more general definitions of dominance later,' but no such discussion appears in the paper. Either add it or remove the promise.
  3. [Tables 1–2] The tables contain formatting artifacts: 'T able 1', 'F und.' for Fundus, and inconsistent column alignment. Clean up the LaTeX/camera-ready rendering.
  4. [Eq. (2)] State explicitly that E and Var are computed over all entries of a single factor matrix for a given expert, layer, and weight matrix, and specify how ε is chosen. This matters because z-scoring across different matrix sizes could behave differently.
  5. [Table 5] Table 5 reports only InternVL2 results. State why Qwen-VL results are omitted and whether the conclusion is backbone-dependent.

Circularity Check

0 steps flagged

No circularity: Winner-take-all is an empirically evaluated merge rule, not a result forced by its own inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. The winning expert in Eq. (3) is selected from the expert LoRA matrices by a fixed normalized-magnitude score (Eq. 2), and Eq. (4)-(5) construct the merged update directly from those matrices. No benchmark accuracy, evaluation label, or fitted quantity is fed back into the method's equations, so the reported accuracy (90.88% / 84.64%) is not equivalent to an input by construction. The 'dominant parameters' motivation in Sec. 3.1 is supported by an independent rescaling experiment, not by the merging result. Tables 3 and 5 report ablations over k and the normalization scheme; this is a model-selection-on-eval-set validity concern rather than a circular step, and it does not change the fact that the central comparison against baselines is an external measurement. The paper contains no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in by citation. The skeptic's A/B-factor-level versus product-level concern is a soundness gap in the method's motivation, but it is not a circularity: the paper never defines 'dominant' in terms of the benchmark outcome and never uses the benchmark outcome to construct the merged model.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The method is hyperparameter-free in name, but two implicit design choices (winner rank k=1 and the z-score normalization family) are validated on the evaluation set. The central assumption is that normalized magnitude tracks functional importance and that independently selected cross-expert A/B products remain coherent. No new physical entities are introduced; 'dominant parameters' is a motivation construct with only in-benchmark evidence.

free parameters (2)
  • winner rank k = 1
    Table 3 shows k=1 gives the highest average on the evaluation set; the method's name fixes this choice, but it is validated on the same split used for the final results.
  • normalization scheme = z-score (Eq. 2)
    Table 5 compares no normalization, min-max, RMS, and z-score on the evaluation set and selects the best; this is test-set-based model selection.
axioms (3)
  • domain assumption Z-score normalized LoRA entries are commensurable across experts, so coordinate-wise argmax is a valid selection rule.
    Section 3.3, Eqs. (2)-(5): merged A and B entries are taken from different experts based on normalized magnitude; if scales are not comparable, selection is arbitrary.
  • domain assumption Parameter magnitude (after normalization) is a valid proxy for functional importance.
    Section 3.1: the rescaling experiment underpins the method; footnote 1 promises broader definitions of dominance that are not delivered.
  • ad hoc to paper The fixed evaluation split used for final reporting is also used for choosing design variants without materially inflating results.
    Section 2.1 fixes the split, and Section 4.3 Tables 3/5 compare method variants on that eval split; the unbiasedness of this procedure is asserted, not tested.
invented entities (1)
  • dominant parameters no independent evidence
    purpose: To justify preserving a small set of LoRA weights during merging.
    Defined by magnitude in Section 3.1; evidence is the in-benchmark rescaling plot (Fig. 2) and the downstream benchmark, with no external falsifiable handle.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Model Merging for Medical LVLMs: A Benchmark and a Winner-Take-All Approach." pith.science (2026). https://pith.science/paper/T4JRGUD4

@misc{pith2026260715661,
  author       = {Pith},
  title        = {Pith review of: Model Merging for Medical LVLMs: A Benchmark and a Winner-Take-All Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4JRGUD4}},
  note         = {Machine review of arXiv:2607.15661}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large vision-language models (LVLMs) can be adapted to specialized medical imaging tasks via parameter-efficient fine-tuning approaches such as low-rank adaptation (LoRA), leading to a growing ecosystem of expert models tailored to specific imaging modalities and clinical scenarios. However, deploying multiple expert LVLMs in practice incurs substantial computational and operational overhead. Model merging provides a promising solution by consolidating multiple experts into a single model without retraining, yet it remains largely unexplored in the medical domain. In this work, we present the first systematic study of model merging for medical LVLMs. We introduce MergeMedBench, a comprehensive benchmark spanning eight imaging modalities and diverse clinical task types, comprising 16 LoRA fine-tuned models built upon two mainstream architectures. We conduct an extensive evaluation of existing merging methods and further propose winner-take-all, a simple and hyperparameter-free approach that retains only the most dominant parameters across expert models. By preserving the critical parameters that govern model behavior and discarding weaker ones, our method avoids the information dilution inherent in averaging- or alignment-based strategies. Despite its simplicity, winner-take-all consistently outperforms existing approaches, offering both a new perspective on LoRA merging and a strong practical baseline for future research.

Figures

Figures reproduced from arXiv: 2607.15661 by Bingcong Yan, Chunlei Li, Jingliang Hu, Lei Li, Lichao Mou, Shengwu Xiong, Shilan Zhang, Xiao Xiang Zhu, Yaxiong Chen, Yilei Shi.

Figure 1
Figure 1. Figure 1: Modality (left) and task (right) distributions of the evaluation set. – We introduce MergeMedBench, a comprehensive benchmark spanning eight imaging modalities and multiple clinical task types, with 16 LoRA fine-tuned models on two mainstream LVLM architectures, to facilitate future research on medical model merging. – We extensively benchmark existing model merging approaches on the pro￾posed benchmark. –… view at source ↗
Figure 2
Figure 2. Figure 2: Effect of rescaling dominant vs. non-dominant parameters. Dominant pa￾rameters are identified using a simple magnitude-based criterion for illustration. This motivates our winner-take-all merging strategy that prioritizes preserving domi￾nant parameters. Before introducing our method, we present a key observation that mo￾tivates our approach. Within learned LoRA parameters, a small subset dis￾proportionate… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of our merged model on multiple modalities and tasks. 5 Conclusion We present the first systematic study of model merging for medical LVLMs. Through extensive evaluation, we benchmark existing merging methods in the medical domain. Furthermore, we propose winner-take-all, a simple yet effec￾tive hyperparameter-free strategy that preserves dominant parameters during merging and consistently outperf… 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

26 extracted references · 3 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, ...

    Pith/arXiv arXiv 2025

  2. [2]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, H., Wang, Z., Li, R., Zhu, B., Chen, L.: IterIS: Iterative inference-solving alignment for LoRA merging. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4829–4838. (2025)

    2025

  3. [3]

    arXiv preprint arXiv:2312.14238 (2023)

    Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu,X.,Lu,L.,Li,B.,Luo,P.,Lu,T.,Qiao,Y.,Dai,J.:InternVL:Scalingupvision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    In: Advances in Neural Information Processing Systems

    Du, G., Lee, J., Li, J., Jiang, R., Guo, Y., Yu, S., Liu, H., Goh, S.K., Tang, H., He, D., Zhang, M.: Parameter competition balancing for model merging. In: Advances in Neural Information Processing Systems. pp. 84746–84776. (2024)

    2024

  5. [5]

    In: International Conference on Learning Representations, (2022) 10 L

    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: International Conference on Learning Representations, (2022) 10 L. Mou et al

    2022

  6. [6]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Hu, Y., Li, T., Lu, Q., Shao, W., He, J., Qiao, Y., Luo, P.: OmniMedVQA: A new large-scale comprehensive evaluation benchmark for medical LVLM. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22170– 22183. (2024)

    2024

  7. [7]

    arXiv preprint arXiv:2307.13269 (2023)

    Huang, C., Liu, Q., Lin, B.Y., Pang, T., Du, C., Lin, M.: LoraHub: Effi- cient cross-task generalization via dynamic LoRA composition. arXiv preprint arXiv:2307.13269 (2023)

    Pith/arXiv arXiv 2023

  8. [8]

    In: International Conference on Learning Representations, (2023)

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

    2023

  9. [9]

    In: International Conference on Learning Representations, (2023)

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

    2023

  10. [10]

    IEEE Transactions on Medical Imaging pp

    Lai, Y., Zhong, J., Li, M., Zhao, S., Li, Y., Psounis, K., Yang, X.: Med-R1: Rein- forcement learning for generalizable medical reasoning in vision-language models. IEEE Transactions on Medical Imaging pp. 1–1 (2026)

    2026

  11. [11]

    In: International Conference on Machine Learning, (2025)

    Marczak, D., Magistri, S., Cygert, S., Twardowski, B., Bagdanov, A.D., van de Weijer, J.: No task left behind: Isotropic model merging with common and task- specific subspaces. In: International Conference on Machine Learning, (2025)

    2025

  12. [12]

    In: Ad- vances in Neural Information Processing Systems

    Matena, M., Raffel, C.: Merging models with fisher-weighted averaging. In: Ad- vances in Neural Information Processing Systems. pp. 17703–17716. (2022)

    2022

  13. [13]

    In: Medical Image Computing and Computer Assisted Intervention

    Pan, J., Liu, C., Wu, J., Liu, F., Zhu, J., Li, H.B., Chen, C., Ouyang, C., Rueck- ert, D.: MedVLM-R1: Incentivizing medical reasoning capability of vision-language models (VLMs) via reinforcement learning. In: Medical Image Computing and Computer Assisted Intervention. pp. 337–347. (2025)

    2025

  14. [14]

    arXiv preprint arXiv:2509.17786 (2025)

    Panariello, A., Marczak, D., Magistri, S., Porrello, A., Twardowski, B., Bagdanov, A.D., Calderara, S., van de Weijer, J.: Accurate and efficient low-rank model merg- ing in core space. arXiv preprint arXiv:2509.17786 (2025)

    arXiv 2025

  15. [15]

    In: International Con- ference on Computational Linguistics

    Prabhakar, A., Li, Y., Narasimhan, K., Kakade, S.M., Malach, E., Jelassi, S.: LoRA soups: Merging LoRAs for practical skill composition tasks. In: International Con- ference on Computational Linguistics. pp. 644–655. (2025)

    2025

  16. [16]

    In: Empirical Methods in Natural Language Processing

    Qiu, H., Wu, Y., Li, D., Guo, J., Yao, Q.: Superpose task-specific features for model merging. In: Empirical Methods in Natural Language Processing. pp. 4200–4214. (2025)

    2025

  17. [17]

    In: European Conference on Computer Vision

    Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.: ZipLoRA: Any subject in any style by effectively merging LoRAs. In: European Conference on Computer Vision. pp. 422–438. (2024)

    2024

  18. [18]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Shenaj, D., Bohdal, O., Ozay, M., Zanuttigh, P., Michieli, U.: LoRA.rar: Learning to merge LoRAs via hypernetworks for subject-style conditioned image genera- tion. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16132–16142. (2025)

    2025

  19. [19]

    In: International Conference on Learning Representations, (2025)

    Stoica, G., Ramesh, P., Ecsedi, B., Choshen, L., Hoffman, J.: Model merging with SVD to tie the knots. In: International Conference on Learning Representations, (2025)

    2025

  20. [20]

    In: Advances in Neural Information Processing Systems

    Yadav, P., Tam, D., Choshen, L., Raffel, C.A., Bansal, M.: TIES-Merging: Re- solving interference when merging models. In: Advances in Neural Information Processing Systems. pp. 7093–7115. (2023)

    2023

  21. [21]

    In: International Conference on Learning Representations, (2024) Model Merging for Medical LVLMs 11

    Yang, E., Wang, Z., Shen, L., Liu, S., Guo, G., Wang, X., Tao, D.: AdaMerging: Adaptive model merging for multi-task learning. In: International Conference on Learning Representations, (2024) Model Merging for Medical LVLMs 11

    2024

  22. [22]

    In: International Conference on Machine Learning

    Yu,L.,Yu,B.,Yu,H.,Huang,F.,Li,Y.:Languagemodelsaresupermario:Absorb- ing abilities from homologous models as a free lunch. In: International Conference on Machine Learning. pp. 57755–57775. (2024)

    2024

  23. [23]

    arXiv preprint arXiv:2502.17159 (2025)

    Zeng, F., Guo, H., Zhu, F., Shen, L., Tang, H.: RobustMerge: Parameter- efficient model merging for mllms with direction robustness. arXiv preprint arXiv:2502.17159 (2025)

    arXiv 2025

  24. [24]

    arXiv preprint arXiv:2509.25712 (2025)

    Zhang, D., Ma, X., Ni, Z., Wu, Z., Shu, H., Jiang, X., Chen, X.: Expert merging: Model merging with unsupervised expert alignment and importance-guided layer chunking. arXiv preprint arXiv:2509.25712 (2025)

    arXiv 2025

  25. [25]

    In: Advances in Neural Information Processing Systems

    Zhang, J., Chen, S., Liu, J., He, J.: Composing parameter-efficient modules with arithmetic operation. In: Advances in Neural Information Processing Systems. pp. 12589–12610. (2023)

    2023

  26. [26]

    In: International Conference on Learning Representations, (2025)

    Zhao, Z., Shen, T., Zhu, D., Li, Z., Su, J., Wang, X., Wu, F.: Merging LoRAs like playing LEGO: Pushing the modularity of LoRA to extremes through rank-wise clustering. In: International Conference on Learning Representations, (2025)

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.