Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

ICM-Fusion: In-Context Meta-Optimized LoRA Fusion for Multi-Task Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Guiding LoRA fusion with task-vector orientation in latent space produces one adapter that outperforms both the individual adapters and prior merging methods.

desk verdict A plausible follow-up to ICM-LoRA whose core meta-optimization equations are missing from the manuscript, and whose experimental gains are within noise. read the letter →

arxiv 2508.04153 v1 pith:NI4DYD2E submitted 2025-08-06 cs.CV

Yihua Shao , Xiaofeng Lin , Xinwei Long , Siyu Chen , Minxi Yan , Yang Liu , Ziyang Yan , Ao Ma
show 2 more authors
Hao Tang Jingcai Guo
This is my paper
classification cs.CV
keywords LoRAfusionmulti-taskadaptationtaskvectorsvariationalautoencodermeta-learningin-contextlearningmodelmergingfew-shot
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

The paper tries to solve a practical problem: after a base model has been fine-tuned on several separate tasks, the many LoRA adapters (small low-rank weight modifications) are usually merged by averaging or by decomposing weight matrices, which can cancel out task-specific knowledge. ICM-Fusion proposes instead to describe each task by a 'task vector' — the element-wise difference between the fine-tuned and pre-trained model's final-layer outputs — and to feed that vector, together with the flattened LoRA weights, into a variational autoencoder. The VAE maps each adapter to a latent distribution, and the paper argues that adjusting the orientation of these task vectors in latent space resolves conflicting optimization directions before a decoder rebuilds a single fused LoRA. If the claim is right, one merged adapter can serve many tasks, and in few-shot settings the fused model can even beat the individual adapters it was built from, without retraining.

What carries the argument

Task vectors: the difference between last-layer output tokens of the fine-tuned and pre-trained models, $\Delta_{\mathbf{T}_i}=\mathbf{z}^*_{\mathbf{T}_i}-\mathbf{z}^{(0)}$, used as a semantic conditioning signal. Fusion VAE (F-VAE): a variational autoencoder whose encoder takes $[\mathbf{w}^{(i)};\mathbf{T}_i]$ and outputs $\mu_\phi,\sigma^2_\phi$, and whose decoder takes $[\mathbf{z};\mathbf{T}_i]$ and reconstructs LoRA parameters, trained by maximizing the ELBO of Eq. (4). In-context meta-learning: the adaptation loop that treats LoRA parameters and task vectors as context and optimizes latent representations so the fused model can adapt to new tasks with small data. The load-bearing mech

What would settle it

Compute each task vector from a randomly shuffled or deliberately misleading sample of a task's inputs; if the fused adapter's accuracy stays the same, the task-vector signal is not doing the claimed work. Conversely, build two tasks with near-identical task vectors but very different LoRA weights; if the fused model fails on one task, orientation alone is not sufficient.

Watch

Extended reading notes

Core claim

ICM-Fusion's central claim is that LoRA fusion should happen in a learned latent space under task-vector guidance, not in raw weight space. For each task, the task vector $\Delta_{\mathbf{T}_i} = \mathbf{z}^*_{\mathbf{T}_i} - \mathbf{z}^{(0)}$ (Eq. 1) captures the semantic shift induced by fine-tuning; the Fusion VAE encodes the concatenation of the flattened LoRA parameters and this task vector into a Gaussian latent code, and its decoder reconstructs LoRA parameters from the latent code plus the task vector. Fusion is performed by sampling from the latent space so that the merged adapter's task-vector orientation is the one the meta-learned manifold considers optimal, and the decoder then

Load-bearing premise

The method breaks if the per-task difference in final-layer outputs — one vector per task — is not a faithful semantic description of what fine-tuning changed, because that vector is what tells the VAE how to orient the merged LoRA in latent space.

Editorial extensions

If this is right

  • A single fused adapter can replace multiple task-specific LoRA adapters, cutting parameter storage and inference cost while preserving or improving task performance.
  • The fusion recipe transfers across model architectures and task types, including detection, visual question answering, and pure language modeling.
  • In few-shot long-tail settings, even a small amount of additional data lets the fused model reach markedly higher scores than prior merging methods.
  • The method resists the opposite-sign task-vector interference that degrades SVD-based and averaging-based merging, because orientation is optimized before reconstruction.

Reading between the lines

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

  • The method is called data-free, but computing a task vector requires running representative inputs through both the pre-trained and fine-tuned models, so the practical requirement is a small sample of task data rather than the full training set.
  • If task-vector orientation is the real driver, then two tasks with nearly identical task vectors but divergent LoRA weights should fuse cleanly, while near-orthogonal task vectors should expose conflicts; this is directly testable on synthetic task pairs.
  • The same latent-orientation idea could be extended to sequential fusion: instead of merging all adapters at once, a new task's LoRA could be encoded into the already-fused latent code, turning multi-task fusion into an incremental update.
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 ICM-Fusion, a framework for merging multiple LoRA adapters via a Fusion VAE conditioned on task vectors, with the stated goal of optimizing task-vector orientation in latent space through in-context meta-learning. The authors claim that this yields better multi-task fusion than existing methods and even improves few-shot performance, without relying on original datasets. The manuscript describes a VAE objective (Eq. 4), defines task-vector extraction (Eq. 1), and begins a meta-learning section, but the central meta-optimization mechanism is never specified: after Eq. (5), the method section stops and experimental tables follow. Experiments are reported on object detection, ScienceQA, language modeling, and few-shot long-tail classes, with small average improvements over baselines (e.g., Table 1: 0.90 vs. 0.89 MAP50 for original LoRA; Table 2: 89.39 vs. 89.38 average accuracy).

Significance. If fully specified and validated, the idea of using task vectors to guide latent-space LoRA fusion is potentially useful for efficient multi-task adaptation. The paper covers relevant baselines and demonstrates the approach across vision and language tasks, which is a strength. However, as submitted, the core contribution is not described in enough detail to be implemented, replicated, or falsified, and the reported empirical gains are marginal. The paper does not provide code, pseudo-code, or a complete algorithm, so its significance cannot currently be assessed.

major comments (4)
  1. [Sec. 3.3–3.4, Eq. (5)] The central mechanism of the paper is missing. Section 3.3 ('In-Context Meta Learning') defines only the encoder output (Eq. 5) and then the experimental section follows. There is no equation or description of the 'meta-parameter updating' phase, no objective for 'adjusting the orientation of task vectors', no loss for the learned manifold projections, and no decoder/fusion operation that produces the fused LoRA from latent codes. The abstract and contributions promise 'optimal task vector orientation' and 'learned manifold projections', but these are never formalized. As submitted, the method cannot be implemented or scientifically evaluated.
  2. [Sec. 3.2, Eq. (1); Contributions] The claim that the method works 'without relying on original datasets' is contradicted by the experimental protocol. Computing the task vector in Eq. (1) requires running inputs through both the pre-trained and fine-tuned models, so some data or representative examples are needed. Furthermore, the few-shot experiments explicitly add +10%, +20%, and +30% of training data (Tables 3 and 6). The paper should clarify what data is used for task-vector extraction and whether the method assumes access to unlabeled task-distribution inputs; as written, the data-free claim is unsupported.
  3. [Eq. (2) vs. Eq. (5)] There is an internal inconsistency in the encoder definition. Eq. (2) defines the encoder as taking the concatenation of LoRA parameters and the task vector, i.e., q_phi(z | tau^(i), Δ_Ti). Eq. (5), however, writes the encoder as E_phi(theta_i), which omits the task vector. This discrepancy makes even the described VAE component ambiguous, and the subsequent fusion/interpolation of multiple latent codes is never defined. This is a load-bearing issue because the claimed task-vector guidance depends on this input.
  4. [Tables 1, 2, and 4] The empirical support for the main claims is thin. Table 1 shows an average MAP50 gain of only 0.01 over the original LoRA (0.90 vs. 0.89). Table 2 shows an average accuracy gain of 0.01 over the original model (89.39 vs. 89.38). Table 4 shows a PPL reduction of about 0.02 on average. No standard deviations, multiple seeds, or significance tests are reported. The abstract's phrase 'significantly reduce the multi-tasking loss' is therefore not supported by the evidence presented.
minor comments (5)
  1. [Table 3] The heading 'CM-Fusion (Ours)I' appears to be a typo; it should be 'ICM-Fusion (Ours)'.
  2. [Table 5] The rank column is corrupted (e.g., '� � � 241,241'), and the actual rank values are not readable. Please fix the formatting.
  3. [Table 2] The table formatting is broken: for example, '89.34 96.1085.30' should be separated into distinct columns. This makes the results difficult to verify.
  4. [Figure 3] The text refers to 'Figure 3' and a t-SNE visualization, but no such figure appears in the submitted manuscript. This claim cannot be checked.
  5. [Sec. 3.1] The problem setting mentions that each task-specific LoRA is associated with a VAE, but it is unclear whether one VAE is trained per task or a single VAE is trained on all tasks. The notation VAE_i is introduced but not used consistently.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the fused LoRA is not forced by the paper's equations (the VAE decoder output is not algebraically the input), and the main experimental claims are checked against external baselines. The load-bearing meta-parameter updating is, however, missing from the manuscript, which is a completeness problem, not a circular one.

full rationale

The paper's central derivation is a conditional VAE: LoRA parameters θ(i) and task vectors Δτ_Ti are fed to an encoder (Eq. 2), a latent z is sampled, and the decoder (Eq. 3) reconstructs θ(i); fusion is claimed to happen by combining latents and decoding. Although the task vector in Eq. (1) is computed from the very fine-tuned checkpoints whose LoRA weights are the VAE's reconstruction targets, this does not make the fused result equal to its inputs by construction: the decoder is a learned mapping with a stochastic latent, and the reported multi-task performance is measured on external benchmarks (Tables 1-4), not derived from the equations. Thus no prediction is statistically forced by a fitted parameter. There is a self-citation to the authors' prior ICM-LoRA (Shao et al. 2025c), but it is used as a prior method and motivation, not as the justification for the correctness of ICM-Fusion, so it is not load-bearing. The more serious issue is that Sec. 3.3 ('In-Context Meta Learning') promises 'meta-parameter updating' and 'optimal task vector orientation' but never provides the update rule, loss, or decoding/fusion algorithm; and the abstract's 'without relying on original datasets' conflicts with Eq. (1)'s need to run inputs through both models. These are omitted-specification and consistency defects that prevent verification, but they are not themselves a circular reduction of outputs to inputs. Hence a low circularity score is appropriate.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The framework rests on the assumed meaningfulness of task-vector guidance, the assumed reconstructability of flattened LoRA weights through a low-dimensional Gaussian latent, and an unstated meta-learning objective that does not appear in the received text. No free parameters are numerically reported, but the VAE capacity, latent dimension, and the few-shot data proportions are hand-chosen protocol values that the central claims depend on.

free parameters (3)
  • VAE latent dimension d and network architecture
    The encoder/decoder sizes and latent dimensionality are not reported in the received text; these hyperparameters control the capacity of the fusion bottleneck and were presumably chosen by hand (Sec. 3.3).
  • few-shot data proportions (+10%, +20%, +30%) = 0.10, 0.20, 0.30
    The few-shot claim in Tables 3 and 6 depends on an additional sampling proportion that is a protocol choice made by the authors; at 0% all methods score 0.00, so the headline 'task enhancement' only appears once extra data is supplied.
  • task vector input examples
    Eq. (1) requires output tokens h*_Ti; the input set used to elicit those tokens is unspecified, and its choice determines the task vector and therefore the fused weights.
assumptions (3)
  • domain assumption The final-layer output-token difference between fine-tuned and pretrained models is a valid task descriptor that can guide LoRA parameter fusion.
    Eq. (1) defines the task vector; the entire fusion guidance relies on this premise without theoretical or empirical validation (Sec. 3.2).
  • domain assumption Flattened LoRA weight vectors lie on a manifold where conditional VAE encoding, latent interpolation, and decoding preserve and combine task knowledge.
    Sec. 3.3 assumes LoRA parameters can be reconstructed from a low-dimensional Gaussian latent conditioned on the task vector; no reconstruction error analysis is provided.
  • ad hoc to paper Meta-learning in the latent space transfers to unseen and few-shot tasks without task boundaries or retraining.
    Sec. 3.4 asserts in-context meta-learning enables rapid adaptation and generalization, but the meta-objective and update rule do not appear in the provided text, so the transfer claim is unsubstantiated.
invented entities (2)
  • Fusion VAE (F-VAE)
    purpose: Encode task-specific LoRA parameters into a shared latent space and decode a fused LoRA.
    A new architectural component introduced by the paper; no external validation or theoretical guarantee beyond its own experiments.
  • optimized task-vector orientation in latent space
    purpose: Dynamically balance conflicting optimization directions across domains.
    The paper claims this learned adjustment of task-vector direction is the key innovation, but no equations, algorithm, or evaluation of the orientation mechanism appear in the provided text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ICM-Fusion: In-Context Meta-Optimized LoRA Fusion for Multi-Task Adaptation." pith.science (2026). https://pith.science/paper/NI4DYD2E

@misc{pith2026250804153,
  author       = {Pith},
  title        = {Pith review of: ICM-Fusion: In-Context Meta-Optimized LoRA Fusion for Multi-Task Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI4DYD2E}},
  note         = {Machine review of arXiv:2508.04153}
}
read the original abstract

Enabling multi-task adaptation in pre-trained Low-Rank Adaptation (LoRA) models is crucial for enhancing their generalization capabilities. Most existing pre-trained LoRA fusion methods decompose weight matrices, sharing similar parameters while merging divergent ones. However, this paradigm inevitably induces inter-weight conflicts and leads to catastrophic domain forgetting. While incremental learning enables adaptation to multiple tasks, it struggles to achieve generalization in few-shot scenarios. Consequently, when the weight data follows a long-tailed distribution, it can lead to forgetting in the fused weights. To address this issue, we propose In-Context Meta LoRA Fusion (ICM-Fusion), a novel framework that synergizes meta-learning with in-context adaptation. The key innovation lies in our task vector arithmetic, which dynamically balances conflicting optimization directions across domains through learned manifold projections. ICM-Fusion obtains the optimal task vector orientation for the fused model in the latent space by adjusting the orientation of the task vectors. Subsequently, the fused LoRA is reconstructed by a self-designed Fusion VAE (F-VAE) to realize multi-task LoRA generation. We have conducted extensive experiments on visual and linguistic tasks, and the experimental results demonstrate that ICM-Fusion can be adapted to a wide range of architectural models and applied to various tasks. Compared to the current pre-trained LoRA fusion method, ICM-Fusion fused LoRA can significantly reduce the multi-tasking loss and can even achieve task enhancement in few-shot scenarios.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. DA-MergeLoRA: Hypernetwork-Based LoRA Merging for Few-Shot Test-Time Domain Adaptation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A hypernetwork generates per-column merging weights to combine source LoRA modules on CLIP, achieving state-of-the-art few-shot test-time domain adaptation.

  2. SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass

    cs.CL 2026-02 unverdicted novelty 6.0 of 10

    SHINE trains a scalable in-context hypernetwork to generate high-quality LoRA adapters from contexts in one pass, enabling efficient LLM adaptation that saves time and compute compared to standard fine-tuning.

Reference graph

Works this paper leans on

75 extracted references · 33 canonical work pages · cited by 2 Pith papers

  1. [1]

    G.; Soares, C.; and Vilalta, R

    Brazdil, P.; Carrier, C. G.; Soares, C.; and Vilalta, R. 2008. Metalearning: Applications to data mining. Springer science & business media

    work page 2008

  2. [2]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  3. [3]

    Caruana, R. 1997. Multitask learning. Machine learning, 28(1): 41--75

    1997

  4. [4]

    Chen, Y.; Zhong, R.; Zha, S.; Karypis, G.; and He, H. 2021. Meta-learning via language model in-context tuning. arXiv preprint arXiv:2110.07814

    arXiv 2021

  5. [5]

    Dong, Q.; Li, L.; Dai, D.; Zheng, C.; Ma, J.; Li, R.; Xia, H.; Xu, J.; Wu, Z.; Liu, T.; et al. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

    arXiv 2022

  6. [6]

    K.; Winn, J.; and Zisserman, A

    Everingham, M.; Van Gool, L.; Williams, C. K.; Winn, J.; and Zisserman, A. 2010. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html

    work page 2010

  7. [7]

    Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, 1126--1135. PMLR

    2017

  8. [8]

    Gao, L.; Biderman, S.; Black, S.; Golding, L.; Hoppe, T.; Foster, C.; Phang, J.; He, H.; Thite, A.; Nabeshima, N.; et al. 2020. The Pile: An 800GB Dataset of Diverse Text for Language Modeling. arXiv preprint arXiv:2101.00027

    arXiv 2020

Show all 75 references
  1. [9]

    Gepperth, A.; and Hammer, B. 2016. Incremental learning algorithms and applications. In European symposium on artificial neural networks (ESANN)

    2016

  2. [10]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  3. [11]

    Guo, Z.; Su, J.; Wang, S.; Fan, J.; Zhang, J.; Zhou, W.; Amirpour, H.; Zhao, Y.; Han, L.; and Wang, P. 2025. GP-GS: Gaussian Processes for Enhanced Gaussian Splatting. arXiv preprint arXiv:2502.02283

    2025

  4. [12]

    Hendel, R.; Geva, M.; and Globerson, A. 2023. In-context learning creates task vectors. arXiv preprint arXiv:2310.15916

    2023 arXiv

  5. [13]

    Hospedales, T.; Antoniou, A.; Micaelli, P.; and Storkey, A. 2021. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9): 5149--5169

    2021

  6. [14]

    Hsu, K.; Levine, S.; and Finn, C. 2018. Unsupervised learning via meta-learning. arXiv preprint arXiv:1810.02334

    2018 arXiv

  7. [15]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

    2022

  8. [16]

    Huang, L.; Wang, W.; Wu, Z.-F.; Shi, Y.; Dou, H.; Liang, C.; Feng, Y.; Liu, Y.; and Zhou, J. 2024. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775

    2024 arXiv

  9. [17]

    Huang, Y.; Lin, Z.; Liu, H.; et al. 2023. ComPEFT : Computationally Efficient LoRA Fusion via Shared Sparse Adapters. In EMNLP

    2023

  10. [18]

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

    Ilharco, G.; Ribeiro, M. T.; Wortsman, M.; Gururangan, S.; Schmidt, L.; Hajishirzi, H.; and Farhadi, A. 2022. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089

    2022 arXiv

  11. [19]

    Jang, D.-H.; Yun, S.; and Han, D. 2024. Model stock: All we need is just a few fine-tuned models. In European Conference on Computer Vision, 207--223. Springer

    2024

  12. [20]

    Jiang, W.; Li, D.; Hu, M.; Zhai, G.; Yang, X.; and Zhang, X.-P. 2024. Few-shot class-incremental learning with prior knowledge. arXiv preprint arXiv:2402.01201

    2024 arXiv

  13. [21]

    Jin, X.; Ren, X.; Preotiuc-Pietro, D.; and Cheng, P. 2022. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849

    2022 arXiv

  14. [22]

    P.; Welling, M.; et al

    Kingma, D. P.; Welling, M.; et al. 2013. Auto-encoding variational bayes

    2013

  15. [23]

    Koren, Y. 2008. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, 426--434

    2008

  16. [24]

    Lee, K.; Maji, S.; Ravichandran, A.; and Soatto, S. 2019. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10657--10665

    2019

  17. [25]

    Li, H.; Zhang, Y.; Zhang, S.; Chen, P.-Y.; Liu, S.; and Wang, M. 2025 a . When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers. In The Thirteenth International Conference on Learning Representations

    2025

  18. [26]

    Li, K.; Santiago, J. A. C.; Samarasinghe, S. D.; Liu, G.; and Shah, M. 2025 b . GVD: Guiding Video Diffusion Model for Scalable Video Distillation. arXiv preprint arXiv:2507.22360

    2025 arXiv

  19. [27]

    Li, W.; Peng, Y.; Zhang, M.; Ding, L.; Hu, H.; and Shen, L. 2023. Deep model fusion: A survey. arXiv preprint arXiv:2309.15698

    2023 arXiv

  20. [28]

    Li, W.; Zou, L.; Tang, M.; Yu, Q.; Li, W.; and Li, C. 2025 c . META-LORA: Memory-Efficient Sample Reweighting for Fine-Tuning Large Language Models. In Proceedings of the 31st International Conference on Computational Linguistics, 8504--8517

    2025

  21. [29]

    Li, Z.; Zhou, F.; Chen, F.; and Li, H. 2017. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835

    2017 arXiv

  22. [30]

    B.; Wang, X.; Yan, Z.; and Shao, Y

    Liao, M.; Dong, H. B.; Wang, X.; Yan, Z.; and Shao, Y. 2025. GM-MoE: Low-Light Enhancement with Gated-Mechanism Mixture-of-Experts. arXiv preprint arXiv:2503.07417

    2025

  23. [31]

    Liu, H.; Li, C.; Li, Y.; and Lee, Y. J. 2024. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 26296--26306

    2024

  24. [32]

    Lu, P.; Mishra, S.; Xia, T.; Qiu, L.; Chang, K.-W.; Zhu, S.-C.; Tafjord, O.; Clark, P.; and Kalyan, A. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35: 2507--2521

    2022

  25. [33]

    Maaten, L. v. d.; and Hinton, G. 2008. Visualizing data using t-SNE. Journal of machine learning research, 9(Nov): 2579--2605

    2008

  26. [34]

    D.; and Van De Weijer, J

    Masana, M.; Liu, X.; Twardowski, B.; Menta, M.; Bagdanov, A. D.; and Van De Weijer, J. 2022. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5513--5533

    2022

  27. [35]

    Min, S.; Lewis, M.; Zettlemoyer, L.; and Hajishirzi, H. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943

    2021 arXiv

  28. [36]

    Nichol, A.; Achiam, J.; and Schulman, J. 2018. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999

    2018 arXiv

  29. [37]

    A.; and Lewis, M

    Press, O.; Zhang, M.; Min, S.; Schmidt, L.; Smith, N. A.; and Lewis, M. 2022. Measuring and narrowing the compositionality gap in language models. arXiv preprint arXiv:2210.03350

    2022 arXiv

  30. [38]

    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. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

    2021

  31. [39]

    Remondino, F.; Karami, A.; Yan, Z.; Mazzacca, G.; Rigon, S.; and Qin, R. 2023. A critical analysis of nerf-based 3d reconstruction. Remote Sensing, 15(14): 3585

    2023

  32. [40]

    Rubin, O.; Herzig, J.; and Berant, J. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633

    2021 arXiv

  33. [41]

    Shao, Y.; He, H.; Li, S.; Chen, S.; Long, X.; Zeng, F.; Fan, Y.; Zhang, M.; Yan, Z.; Ma, A.; et al. 2025 a . EventVAD: Training-Free Event-Aware Video Anomaly Detection. arXiv preprint arXiv:2504.13092

    2025 arXiv

  34. [42]

    Shao, Y.; Liang, S.; Lin, X.; Ling, Z.; Zhu, Z.; Yan, M.; Liu, H.; Chen, S.; Yan, Z.; Meng, Y.; et al. 2024 a . GWQ: Gradient-Aware Weight Quantization for Large Language Models. arXiv preprint arXiv:2411.00850

    2024 arXiv

  35. [43]

    Shao, Y.; Lin, D.; Zeng, F.; Yan, M.; Zhang, M.; Chen, S.; Fan, Y.; Yan, Z.; Wang, H.; Guo, J.; et al. 2025 b . TR-DQ: Time-Rotation Diffusion Quantization. arXiv preprint arXiv:2503.06564

    2025 arXiv

  36. [44]

    Shao, Y.; Xu, Y.; Long, X.; Chen, S.; Yan, Z.; Yang, Y.; Liu, H.; Wang, Y.; Tang, H.; and Lei, Z. 2024 b . AccidentBlip: Agent of Accident Warning based on MA-former. arXiv preprint arXiv:2404.12149

    2024 arXiv

  37. [45]

    Shao, Y.; Yan, M.; Liu, Y.; Chen, S.; Chen, W.; Long, X.; Yan, Z.; Li, L.; Zhang, C.; Sebe, N.; et al. 2025 c . In-Context Meta LoRA Generation. arXiv preprint arXiv:2501.17635

    2025 arXiv

  38. [46]

    Sinha, S.; Yue, Y.; Soto, V.; Kulkarni, M.; Lu, J.; and Zhang, A. 2024. Maml-en-llm: Model agnostic meta-training of llms for improved in-context learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2711--2720

    2024

  39. [47]

    Stoica, G.; Ramesh, P.; Ecsedi, B.; Hoffman, J.; and Choshen, L. 2024. KnOTS: Model Merging with SVD to Tie the Knots. arXiv preprint arXiv:2410.19735

    2024 arXiv

  40. [48]

    H.; and Hospedales, T

    Sung, F.; Yang, Y.; Zhang, L.; Xiang, T.; Torr, P. H.; and Hospedales, T. M. 2018. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1199--1208

    2018

  41. [49]

    Tang, A.; Shen, L.; Luo, Y.; Hu, H.; Du, B.; and Tao, D. 2024. Fusionbench: A comprehensive benchmark of deep model fusion. arXiv preprint arXiv:2406.03280

    2024

  42. [50]

    Thoppilan, R.; De Freitas, D.; Hall, J.; Shazeer, N.; Kulshreshtha, A.; Cheng, H.-T.; Jin, A.; Bos, T.; Baker, L.; Du, Y.; et al. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239

    2022 arXiv

  43. [51]

    Vandenhende, S.; Georgoulis, S.; Van Gansbeke, W.; Proesmans, M.; Dai, D.; and Van Gool, L. 2021. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7): 3614--3633

    2021

  44. [52]

    Vanschoren, J. 2018. Meta-learning: A survey. arXiv preprint arXiv:1810.03548

    2018 arXiv

  45. [53]

    Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D.; et al. 2016. Matching networks for one shot learning. Advances in neural information processing systems, 29

    2016

  46. [54]

    Wang, H.; Ping, B.; Wang, S.; Han, X.; Chen, Y.; Liu, Z.; and Sun, M. 2024 a . Lora-flow: Dynamic lora fusion for large language models in generative tasks. arXiv preprint arXiv:2402.11455

    2024 arXiv

  47. [55]

    Wang, N.; Chen, Y.; Xiao, L.; Xiao, W.; Li, B.; Chen, Z.; Ye, C.; Xu, S.; Zhang, S.; Yan, Z.; et al. 2025. Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting. arXiv preprint arXiv:2506.05280

    2025

  48. [56]

    L.; and Pham, M

    Wang, P.; Guo, Z.; Sait, A. L.; and Pham, M. H. 2024 b . Robot Shape and Location Retention in Video Generation Using Diffusion Models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 7375--7382. IEEE

    2024

  49. [57]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  50. [58]

    Wies, N.; Levine, Y.; and Shashua, A. 2023. The learnability of in-context learning. Advances in Neural Information Processing Systems, 36: 36637--36651

    2023

  51. [59]

    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. 2022 a . Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In In...

    2022

  52. [60]

    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. 2022 b . Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7959--7971

    2022

  53. [61]

    Wu, X.; Huang, S.; and Wei, F. 2024. Mixture of lora experts. arXiv preprint arXiv:2404.13628

    2024 arXiv

  54. [62]

    Wu, Y.; Huang, L.-K.; Wang, R.; Meng, D.; and Wei, Y. 2024. Meta continual learning revisited: Implicitly enhancing online hessian approximation via variance reduction. In The Twelfth international conference on learning representations, volume 2

    2024

  55. [63]

    Xiao, B.; Wu, H.; Xu, W.; Dai, X.; Hu, H.; Lu, Y.; Zeng, M.; Liu, C.; and Yuan, L. 2024. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4818--4829

    2024

  56. [64]

    A.; and Bansal, M

    Yadav, P.; Tam, D.; Choshen, L.; Raffel, C. A.; and Bansal, M. 2023. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36: 7093--7115

    2023

  57. [65]

    Yan, Z.; Dong, W.; Shao, Y.; Lu, Y.; Haiyang, L.; Liu, J.; Wang, H.; Wang, Z.; Wang, Y.; Remondino, F.; et al. 2024 a . Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356

    2024 arXiv

  58. [66]

    Yan, Z.; Li, L.; Shao, Y.; Chen, S.; Wu, Z.; Hwang, J.-N.; Zhao, H.; and Remondino, F. 2024 b . 3dsceneeditor: Controllable 3d scene editing with gaussian splatting. arXiv preprint arXiv:2412.01583

    2024

  59. [67]

    M.; Trybala, P.; Remondino, F.; et al

    Yan, Z.; Mazzacca, G.; Rigon, S.; Farella, E. M.; Trybala, P.; Remondino, F.; et al. 2023. NeRFBK: a holistic dataset for benchmarking NeRF-based 3D reconstruction. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 48(1): 219--226

    2023

  60. [68]

    M.; and Remondino, F

    Yan, Z.; Padkan, N.; Tryba a, P.; Farella, E. M.; and Remondino, F. 2025. Learning-Based 3D Reconstruction Methods for Non-Collaborative Surfaces—A Metrological Evaluation. Metrology, 5(2): 20

    2025

  61. [69]

    Yu, L.; Yu, B.; Yu, H.; Huang, F.; and Li, Y. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Forty-first International Conference on Machine Learning

    2024

  62. [70]

    Zhang, Q.; Chen, M.; Bukharin, A.; He, P.; Cheng, Y.; Chen, W.; and Zhao, T. 2023. AdaLoRA : Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning. In ICML

    2023

  63. [71]

    Zhang, W.; Deng, L.; Zhang, L.; and Wu, D. 2022. A survey on negative transfer. IEEE/CAA Journal of Automatica Sinica, 10(2): 305--329

    2022

  64. [72]

    Zhang, Y.; and Li, R. 2024. Dlp-lora: Efficient task-specific lora fusion with a dynamic, lightweight plugin for large language models. arXiv preprint arXiv:2410.01497

    2024 arXiv

  65. [73]

    Zoph, B.; Bello, I.; Kumar, S.; Du, N.; Huang, Y.; Dean, J.; Shazeer, N.; and Fedus, W. 2022. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906

    2022 arXiv

  66. [74]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  67. [75]

    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 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.