Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

IterIS: Iterative Inference-Solving Alignment for LoRA Merging

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read IterIS merges multiple task-specific LoRAs into a single unified adapter by iteratively re-estimating the unified model's own input features, using only about 50 unlabeled samples per task.

desk verdict The iterative inference-solving idea is genuinely new and worth engaging, but the few-sample efficiency claim leans on a per-dataset-tuned regularizer that is not derived from the stated objective. read the letter →

arxiv 2411.15231 v2 pith:WHHPFFNT submitted 2024-11-21 cs.LG cs.AI

Hongxu Chen , Runshi Li , Bowei Zhu , Zhen Wang , Long Chen This is my paper · ORCID
classification cs.LGcs.AI
keywords LoRAmerginglow-rankadaptationparameter-efficientfine-tuningiterativeinference-solvingclosed-formoptimizationmulti-taskmodelcompositionsampleefficiencytext-to-imagediffusion
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 proposes IterIS, a way to merge several task-specific LoRA adapters into one unified adapter without access to labeled training data or gradient-based fine-tuning. Where earlier optimization-based mergers assume the unified adapter sees the same input features as each individual LoRA, IterIS instead runs inference with the current unified model, extracts the actual input features at every layer, and then solves a closed-form alignment objective that matches each LoRA's outputs. This inference-solving loop is repeated until convergence, and a regularization term plus adaptive weights let it work from roughly 50 unlabeled samples per task. The authors report that IterIS outperforms linear merging and RegMean on multi-concept text-to-image customization, positive and negative style captioning with BLIP, and in-domain and GLUE task integration with FLAN-T5, while using 1-5% of the samples earlier methods need. If correct, the method offers a practical, privacy-preserving way to compose separately fine-tuned models.

What carries the argument

The mechanism is an iterative inference-solving loop for layer-wise LoRA merging. At each iteration the current unified model is run on unlabeled samples to obtain $\tilde{X}_i$, the input features of the unified adapter, which replace the rough approximation $X_i$ in the alignment objective. The objective is solved in closed form, so each layer's merged weight is updated without gradients; convergence follows from the directed acyclic structure of the network, requiring at most $J-1$ iterations for $J$ adapted layers. Two additions make the few-sample regime work: a regularization term $\alpha \|M\|_F I$ added to the inner-product matrices, which prevents the closed-form solution from overfitting or becoming singular, and adaptive per-task weights $\lambda_i$ that rescale each term so that larger feature magnitudes do not dominate the merged result.

What would settle it

Run IterIS on a fixed pair of LoRAs with 50, 200, and 1000 unlabeled samples, and also run one-shot RegMean with the same samples. If the iterative re-estimation of $\tilde{X}_i$ fails to reduce the layer-wise feature discrepancy over iterations, or if reducing that discrepancy does not track improved task accuracy, the central mechanism is not doing the claimed work.

Watch

Extended reading notes

Core claim

The central discovery is that the rough assumption in real-distribution-based LoRA merging, namely that the unified adapter's input features equal the individual LoRAs' input features, can be relaxed by an iterative inference-solving scheme. Starting from the LoRA features, IterIS alternates between running the current unified model on unlabeled samples to extract $\tilde{X}_i$, the actual input features seen by the unified adapter at each layer, and updating the unified adapter weights $W^*$ by the closed-form solution $W^* = (\sum_i \lambda_i \tilde{X}_i \tilde{X}_i^\top)^{-1}(\sum_i \lambda_i \tilde{X}_i X_i^\top W_i)$. The paper argues that this progressively refined objective, together with a Frobenius-norm regularizer added to the inner-product matrices and adaptive weights $\lambda_i = \|W_i\|_F^2 / \|W_i^\top X_i\|_F^2$, resolves the three limitations it identifies: rough feature assumption, massive sample requirements, and unbalanced optimization. The authors demonstrate the resulting algorithm across three model families and report gains over linear merging and RegMean, with the largest margins in style-control accuracy for vision-language captioning.

Load-bearing premise

The load-bearing premise is that aligning, layer by layer, the current unified model's output features with each task LoRA's output features, using only about 50 unlabeled samples and a few iterations, is a faithful proxy for preserving each task on the real data distribution.

Editorial extensions

If this is right

  • Multi-concept text-to-image models can be composed from individually trained LoRAs without mixing training images, preserving data privacy.
  • The same algorithm transfers to vision-language style control and LLM task integration, so a single merging procedure covers diffusion, vision-language, and language backbones.
  • Only about 50 unlabeled samples per task are needed, cutting the feature-extraction inference budget to 1-5% of RegMean's sample requirement.
  • Layer-wise closed-form updates and the convergence bound keep memory and computation low enough to run on a single GPU.

Reading between the lines

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

  • Beyond the reported domains, this iterative feature-alignment scheme should extend to other parameter-efficient adapters, such as adapter modules or prompt tokens, whose layer-wise outputs can be aligned in the same closed-form way.
  • The per-dataset regularization coefficient $\alpha$ is chosen empirically, so a principled selection rule, for example cross-validation on held-out unlabeled inputs, is a natural next step before deployment.
  • The convergence argument suggests the number of iterations can be tied to network depth rather than tuned on validation data, yielding a fully training-free recipe for LoRA composition.
  • If intermediate-layer feature alignment is the true driver of the gains, IterIS should also improve multi-LoRA composition in non-generative settings such as retrieval or recommendation towers, a transfer that would be straightforward to test.
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

5 major / 6 minor

Summary. The paper proposes IterIS, a training-free LoRA merging algorithm that composes multiple task-specific LoRAs into a single unified adapter by iteratively (i) inferring the input features of the current unified adapter, (ii) solving a layer-wise feature-alignment objective in closed form, and (iii) updating each adapter. The method is applied to multi-concept customization in text-to-image diffusion models, multi-style captioning in vision-language models, and multi-task integration in large language models. The paper claims that IterIS mitigates prior methods' rough feature assumption, reduces the unlabeled sample requirement to 1-5% of prior methods, balances the optimization objective with adaptive weights, and significantly outperforms baselines such as linear merging, RegMean, custom diffusion, and task arithmetic.

Significance. If the claims hold, IterIS would be a practically valuable LoRA composition method: it is training-free, uses closed-form updates, and appears broadly applicable across diffusion, vision-language, and language models. The paper includes correct closed-form derivations for the unregularized objective (Eq. 6), and the component ablations in the appendix indicate that each proposed mechanism contributes to the reported behavior. The central idea of re-estimating the unified adapter's input features iteratively rather than approximating them with the individual LoRA's features is a plausible advance over RegMean. However, the load-bearing few-sample and 'significant improvements' claims are not yet established: the regularization term that makes few-shot performance possible is introduced without a derived objective or a selection rule, and the empirical comparisons lack error bars or significance tests.

major comments (5)
  1. [Section 3.2, Eq. (7)] The regularized problem actually solved is not the minimizer of Eq. (5) under any stated regularized objective. A ridge-style regularization of Eq. (6) would modify the denominator as Σ λ_i (\tilde X_i \tilde X_i^T + α I), but Eq. (7) adds norm-scaled identity matrices to both the denominator and the numerator cross-covariance term. This biases the solution toward linear merging in a way that is not a minimum of a written objective. Because Tables 8-9 show that removing this term degrades performance sharply (CIDEr from 0.794 to 0.534; some GLUE task pairs collapsing to 0.0), the 'only 1-5% of samples' claim currently rests on an unprincipled modification rather than on a derived regularized objective.
  2. [Section 4.2, Section 4.3] The regularization hyperparameter α is tuned per experimental setting (8e-7 for the vision-language model, 1e-7 for in-domain NLP, 1e-4 for GLUE), but no selection rule, validation procedure, or sensitivity analysis is given. Since Eq. (7) is essential to the few-sample performance, the reader cannot judge whether the reported 50-sample results transfer to new tasks. Please report how α is chosen and provide an ablation over at least an order of magnitude around each selected value.
  3. [Tables 1, 2, 3 and appendix Tables 12-14] All comparisons are single-run point estimates without error bars, number of seeds, or significance tests. Several reported gains are very small (e.g., image-alignment1 mean 0.6889 vs. linear's 0.6811 in Table 1; SST-2 0.946 vs. 0.945 in Table 3), and the abstract's 'significant improvements' is therefore not statistically supported. Multi-seed runs with paired or per-composition significance tests are needed.
  4. [Section B.3, Section E, Figure 3(b)] The '1-5% of unlabeled samples' claim is not defined against a consistent baseline budget. Section B.3 states that RegMean is reproduced with 100-200 inference samples, while the efficiency analysis in Section E charges RegMean 1600 samples and Figure 3(b) appears to use 2400; IterIS uses 50 samples. Against the Section B.3 budget, 50 samples is 25-50% of RegMean, not 1-5%. The sample-efficiency advantage must be reported against matching baseline sample budgets.
  5. [Section 3.2 and Appendix C] The paper never demonstrates that the layer-wise alignment error measured on the 50 inference samples is predictive of held-out task performance. The essentiality of the regularizer (Tables 8-9) makes this a load-bearing assumption rather than an evident one. Please provide a transfer test or a plot of alignment error versus accuracy/CIDEr across sample counts and α values, or otherwise validate the proxy.
minor comments (6)
  1. [Section 4.1] There is a typo in 'text-to-tmage diffusion model' in the experimental setup paragraph; it should read 'text-to-image'.
  2. [Section 3.1] The text refers to 'Eq. (9)' as though it appears in the main text, but Eq. (9) is only introduced in Appendix A.1; please renumber or provide a main-text cross-reference.
  3. [Section 4.3, Table 3] The sentence 'our method consistently outperforms both Linear and RegMean' is contradicted by the COLA row, where linear merging achieves 0.678 while IterIS achieves 0.666; the wording should be softened or the COLA result discussed.
  4. [Figure 3] The horizontal axis labeled 'score' in Figure 3 is not defined in the caption or the text, which makes the comparison difficult to interpret.
  5. [Appendix A.3] The convergence proof gives an upper bound on the number of iterations, but the experiments simply cap iterations at 10 or 20; no stopping criterion based on the bound is used or discussed.
  6. [Notation throughout] The notation X_i and \tilde X_i is overloaded: these symbols denote random variables in Section 3.1, sample-feature matrices in Eq. (5), and per-layer feature tensors in Algorithm 1; please distinguish these consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: IterIS extends RegMean with iterative feature re-estimation, regularization, and adaptive weights, and is evaluated on external benchmarks.

full rationale

The derivation chain is self-contained. The optimization objective in Eq. (5) is a well-posed least-squares alignment problem, and the closed-form solution in Eq. (6) is the actual minimizer of that objective (verified by differentiating the Frobenius norm). The iterative scheme replaces the rough assumption by re-estimating the unified adapter features Xtilde_i through inference on the current merged model; this is a fixed-point refinement, not an identity. The regularization in Eq. (7) adds a scaled identity to Gram matrices, which is a standard shrinkage device; the per-dataset choice of alpha is ordinary hyperparameter selection, and the reported metrics (CLIP alignment, CIDEr, BLEU, F1, accuracy) are computed on held-out external benchmarks rather than being fitted or defined by the method's own objective. The adaptive weights in Eq. (8) are derived from parameter and feature norms, not from the evaluation metrics, so no reported score is forced by construction. The convergence proof is an internal correctness argument about fixed points on a DAG, not a circular appeal. The only self-citations in the reference list are related-work context on PEFT and are not load-bearing for the central claims. The main weakness, namely that few-sample performance depends heavily on the tuned regularization term, is a robustness and generalization risk rather than a circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's central claims rest on three hyperparameters (alpha, sample count, iteration count) that are tuned per experiment, and on the domain assumption that feature alignment on a small unlabeled set transfers to task performance. No new physical entities or conserved quantities are introduced.

free parameters (3)
  • alpha (regularization coefficient) = 1e-4 or lower; specific values 8e-7 (V&L), 1e-7 (in-domain LLM), 1e-4 (multi-task LLM)
    Introduced in Eq. (7) to make the few-sample closed-form solution robust; the paper tunes it per experiment and shows in ablation (Table 8) that without it the method fails, so the central sample-efficiency claim depends on this choice.
  • number of unlabeled samples per task = 50
    All experiments use 50 inputs per task; the paper claims this is only 1-5% of prior methods' requirements, but this specific number is chosen by hand and not derived.
  • maximum iterations = 6 (V&L), 10 (LLM), capped at 20
    The algorithm's iteration count is a hyperparameter; even though a theoretical bound is given, the paper caps iterations to prevent overfitting and selects the number per experiment.
assumptions (4)
  • domain assumption The optimization objective in Eq. (5), which aligns the output features of the unified adapter with those of each individual LoRA on unlabeled samples, is a valid proxy for preserving task performance.
    This is the core modeling assumption inherited from RegMean and other feature-alignment merging methods; it is necessary for the closed-form solution to be meaningful (Section 3.2).
  • domain assumption The input features for the unified adapter can be accurately extracted by running the current merged model on the unlabeled samples.
    The iterative framework in Algorithm 1 re-extracts features from fM each iteration; if the 50 samples are not representative, the feature estimates are biased.
  • standard math The dependency graph of LoRA positions is a directed acyclic graph, and a layer's optimal adapter is fully determined once its input features are fixed.
    Used in Appendix A.3 to prove the finite iteration bound; this is a graph-theoretic argument about the algorithm's fixed point.
  • domain assumption The regularized matrix sum (Eq. 7) remains invertible with the chosen alpha.
    The closed-form solution requires inversion of the matrix sum; the paper uses identity regularization to guarantee invertibility, relying on alpha > 0.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IterIS: Iterative Inference-Solving Alignment for LoRA Merging." pith.science (2026). https://pith.science/paper/WHHPFFNT

@misc{pith2026241115231,
  author       = {Pith},
  title        = {Pith review of: IterIS: Iterative Inference-Solving Alignment for LoRA Merging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHHPFFNT}},
  note         = {Machine review of arXiv:2411.15231}
}
read the original abstract

Low-rank adaptations (LoRA) are widely used to fine-tune large models across various domains for specific downstream tasks. While task-specific LoRAs are often available, concerns about data privacy and intellectual property can restrict access to training data, limiting the acquisition of a multi-task model through gradient-based training. In response, LoRA merging presents an effective solution by combining multiple LoRAs into a unified adapter while maintaining data privacy. Prior works on LoRA merging primarily frame it as an optimization problem, yet these approaches face several limitations, including the rough assumption about input features utilized in optimization, massive sample requirements, and the unbalanced optimization objective. These limitations can significantly degrade performance. To address these, we propose a novel optimization-based method, named IterIS: 1) We formulate LoRA merging as an advanced optimization problem to mitigate the rough assumption. Additionally, we employ an iterative inference-solving framework in our algorithm. It can progressively refine the optimization objective for improved performance. 2) We introduce an efficient regularization term to reduce the need for massive sample requirements (requiring only 1-5% of the unlabeled samples compared to prior methods). 3) We utilize adaptive weights in the optimization objective to mitigate potential unbalances in LoRA merging process. Our method demonstrates significant improvements over multiple baselines and state-of-the-art methods in composing tasks for text-to-image diffusion, vision-language models, and large language models. Furthermore, our layer-wise algorithm can achieve convergence with minimal steps, ensuring efficiency in both memory and computation.

Figures

Figures reproduced from arXiv: 2411.15231 by the authors.

Figure 1
Figure 1. Overview of the application of our method (IterIS) across multiple domains. Our general method is adaptable for merging LoRAs in various contexts. IterIS can be applied to (a) text-to-image diffusion models for multi-concept customization, (b) vision-language models for multi-style caption generation, and (c) large language models for multiple NLP tasks integration. Abstract Low-rank adaptations (LoRA) are widely us… view at source ↗
Figure 2
Figure 2. Overview of methods for multi-task application (a) Retain all LoRAs fine-tuned on task-specific datasets. (b) Train unified adapters using gradient-based methods on mixed datasets for multi-tasking. (c) Create each unified adapter via LoRA merg￾ing without labeled data or gradient-based training. Most methods formulate LoRA merging as an optimization problem to align fea￾tures and solve for each unified adapter. ful… view at source ↗
Figure 3
Figure 3. Three key limitations of real-distribution-based merging methods and our improvements. In the scenario of combining the COLA [49] and MNLI [50] tasks in NLP, we com￾pare the representative real-distribution-based merging method, RegMean [18], with our proposed method: (a) RegMean exhibits increasing discrepancies with deeper encoder layers, while IterIS can fully resolve discrepancies. The value of the “score” metri… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Comparison of linear merging, real-distribution-based merging, and IterIS for LoRA Merging. “OPT.” denotes the optimization problem introduced by IterIS. “PTM” denotes the pre-trained model. We define B A to represent A −1B. (a) Linear merging combines each individual …
Figure 5
Figure 5. Figure 5: Qualitative results for multi-concept customization. Target images illustrate single concepts used in composition. (a) Single concept generated by IterIS after composing Cat + Barn. (b) Comparison of pairwise composition across methods. (c) Triple composition examples …
Figure 6
Figure 6. Figure 6: Examples of style caption generated by IterIS. taining both caption generation quality and diversity. Addi￾tional examples and detailed results are in the appendix. 4.3. IterIS for Large Language Model Experimental Setup. We applied IterIS to large language models for …
Figure 7
Figure 7. Figure 7: Illustration of the maximum iteration count in graph GM. For a transformer composed of one encoder and one decoder, the abstracted GM after LoRA fine-tuning on the k and v matrices allows IterIS to converge within two iterations. B-1 B-2 B-3 B-4 RougeL CIDEr ACCP OS AC…
Figure 8
Figure 8. Figure 8: Performance across different maximum iterations alpha was set to 1. During the textual inversion phase, we used a learning rate of 5 × 10−4 , with each new concept represented by a single learnable token, all initialized uniformly. In the sub￾sequent fine-tuning phase,…
Figure 9
Figure 9. Figure 9: Additional qualitative results for multi-concept customization. Target images represent individual concepts used in the compositions. (a) Examples of pairwise compositions generated by our method. (b) Examples of individual concepts generated by IterIS [PITH_FULL_IMAG…
Figure 10
Figure 10. Figure 10: More examples of style caption generated by our algorithm. Composition (MRPC, SST2) (COLA, MNLI) (RTE, MRPC) Ours w/o Reg (0.272, 0.177) (0.0, 0.0) (0.671, 0.667) Ours w/ Reg (0.824, 0.951) (0.299, 0.780) (0.805, 0.814) [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: The workflow diagram of our algorithm. More Results on LLM. A detailed version of [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. MultLFG: Training-free Multi-LoRA composition using Frequency-domain Guidance

    cs.CV 2025-05 conditional novelty 6.0 of 10

    MultLFG merges multiple LoRA adapters by adaptively weighting them in wavelet frequency subbands per denoising timestep, improving multi-concept composition on the ComposLoRA benchmark compared to prior training-free methods.

  2. Decouple and Orthogonalize: A Data-Free Framework for LoRA Merging

    cs.CV 2025-05 conditional novelty 6.0 of 10

    A data-free LoRA merging framework that decouples weight magnitude from direction and orthogonalizes directions to reduce task interference, outperforming existing merging methods across vision, language and multimoda...

  3. DP-FedLoRA: Privacy-Enhanced Federated Fine-Tuning for On-Device Large Language Models

    cs.CR 2025-09 reject novelty 3.0 of 10

    DP-FedLoRA clips and adds Gaussian noise to per-client LoRA matrices in federated LLM fine-tuning, claiming unbiased updates and bounded variance, but the privacy calibration and experiments have significant gaps.

Reference graph

Works this paper leans on

58 extracted references · 33 canonical work pages · cited by 3 Pith papers

  1. [1]

    An analysis of annotated corpora for emotion classification in text

    Laura Ana Maria Bostan and Roman Klinger. An analysis of annotated corpora for emotion classification in text. 2018. 2

    work page 2018

  2. [2]

    Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa De- hghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dea...

    work page 2022

  3. [3]

    Sherl: Synthesizing high accuracy and efficient memory for resource-limited transfer learning

    Haiwen Diao, Bo Wan, Xu Jia, Yunzhi Zhuge, Ying Zhang, Huchuan Lu, and Long Chen. Sherl: Synthesizing high accuracy and efficient memory for resource-limited transfer learning. In ECCV, pages 75–95, 2024. 3

    work page 2024

  4. [4]

    Unipt: Universal parallel tuning for trans- fer learning with efficient parameter and memory

    Haiwen Diao, Bo Wan, Ying Zhang, Xu Jia, Huchuan Lu, and Long Chen. Unipt: Universal parallel tuning for trans- fer learning with efficient parameter and memory. In CVPR, pages 28729–28740, 2024. 3

    work page 2024

  5. [5]

    Automatically constructing a corpus of sentential paraphrases

    Bill Dolan and Chris Brockett. Automatically constructing a corpus of sentential paraphrases. In IWP2005, 2005. 8, 2

    work page 2005

  6. [6]

    An image is worth one word: Personalizing text-to- image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3, 6, 7

    arXiv 2022

  7. [7]

    Stylenet: Generating attractive visual captions with styles

    Chuang Gan, Zhe Gan, Xiaodong He, Jianfeng Gao, and Li Deng. Stylenet: Generating attractive visual captions with styles. In CVPR, pages 3137–3146, 2017. 2, 5

    work page 2017

  8. [8]

    Detect- ing emotion stimuli in emotion-bearing sentences

    Diman Ghazi, Diana Inkpen, and Stan Szpakowicz. Detect- ing emotion stimuli in emotion-bearing sentences. In CI- CLing, pages 152–165. Springer, 2015. 8, 2

    work page 2015

Show all 58 references
  1. [9]

    The third pascal recognizing textual en- tailment challenge

    Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and William B Dolan. The third pascal recognizing textual en- tailment challenge. In ACL-PASCAL, pages 1–9, 2007. 8, 2

    2007

  2. [10]

    Mix-of-show: Decentralized low- rank adaptation for multi-concept customization of diffusion models

    Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yun- peng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low- rank adaptation for multi-concept customization of diffusion models. NeurIPS, 36, 2024. 2, 3, 4

    2024

  3. [11]

    Parameter-efficient fine-tuning for large models: A comprehensive survey

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608 ,

    arXiv

  4. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 3

    2016

  5. [13]

    On the effectiveness of adapter-based tuning for pretrained lan- guage model adaptation

    Ruidan He, Linlin Liu, Hai Ye, Qingyu Tan, Bosheng Ding, Liying Cheng, Jia-Wei Low, Lidong Bing, and Luo Si. On the effectiveness of adapter-based tuning for pretrained lan- guage model adaptation. arXiv preprint arXiv:2106.03164,

    arXiv

  6. [14]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3

    2021 arXiv

  7. [15]

    Lorahub: Efficient cross-task gen- eralization via dynamic lora composition

    Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task gen- eralization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 2, 3

    2023 arXiv

  8. [16]

    Editing models with task arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 7

    2022 arXiv

  9. [17]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. In ECCV, pages 709–727. Springer,

  10. [18]

    Dataless knowledge fusion by merging weights of language models

    Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849 , 2022. 2, 3, 4, 7, 8

    2022 arXiv

  11. [19]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 3

    2014 arXiv

  12. [20]

    Diverse im- age captioning with grounded style

    Franz Klein, Shweta Mahajan, and Stefan Roth. Diverse im- age captioning with grounded style. In DAGM German Con- ference on Pattern Recognition , pages 421–436. Springer,

  13. [21]

    Multi-concept customiza- tion of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customiza- tion of text-to-image diffusion. In CVPR, pages 1931–1941,

    1931

  14. [22]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888–12900. PMLR, 2022. 1, 7, 3

    2022

  15. [23]

    On the limited memory bfgs method for large scale optimization.Mathematical program- ming, 45(1):503–528, 1989

    Dong C Liu and Jorge Nocedal. On the limited memory bfgs method for large scale optimization.Mathematical program- ming, 45(1):503–528, 1989. 3

    1989

  16. [24]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1

    2024

  17. [25]

    Multi-task deep neural networks for natural language understanding

    Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. Multi-task deep neural networks for natural language understanding. arXiv preprint arXiv:1901.11504, 2019. 3

    1901 arXiv

  18. [26]

    P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengx- iao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 3

    arXiv

  19. [27]

    Senticap: Generating image descriptions with sentiments

    Alexander Mathews, Lexing Xie, and Xuming He. Senticap: Generating image descriptions with sentiments. In AAAI,

  20. [28]

    Comparison of the predicted and ob- served secondary structure of t4 phage lysozyme

    Brian W Matthews. Comparison of the predicted and ob- served secondary structure of t4 phage lysozyme. Biochim- ica et Biophysica Acta (BBA)-Protein Structure, 405(2):442– 451, 1975. 8

    1975

  21. [29]

    # emotional tweets

    Saif Mohammad. # emotional tweets. In SEM, pages 246– 255, 2012. 8, 2

    2012

  22. [30]

    Wassa- 2017 shared task on emotion intensity

    Saif M Mohammad and Felipe Bravo-Marquez. Wassa- 2017 shared task on emotion intensity. arXiv preprint arXiv:1708.03700, 2017. 8, 2

    2017 arXiv

  23. [31]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311–318, 2002. 7

    2002

  24. [32]

    Flickr30k entities: Collecting region-to-phrase corre- spondences for richer image-to-sentence models

    Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazeb- nik. Flickr30k entities: Collecting region-to-phrase corre- spondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 2

    2015

  25. [33]

    Lora soups: Merg- ing loras for practical skill composition tasks

    Akshara Prabhakar, Yuanzhi Li, Karthik Narasimhan, Sham Kakade, Eran Malach, and Samy Jelassi. Lora soups: Merg- ing loras for practical skill composition tasks. arXiv preprint arXiv:2410.13025, 2024. 2, 3

    2024 arXiv

  26. [34]

    Improving language understanding by gener- ative pre-training

    Alec Radford. Improving language understanding by gener- ative pre-training. 2018. 1

    2018

  27. [35]

    Language models are unsu- pervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsu- pervised multitask learners. OpenAI blog, 1(8):9, 2019. 1

    2019

  28. [36]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, pages 8748–8763. PMLR, 2021. 6

    2021

  29. [37]

    Squad: 100,000+ questions for machine com- prehension of text

    P Rajpurkar. Squad: 100,000+ questions for machine com- prehension of text. arXiv preprint arXiv:1606.05250, 2016. 8, 2

    2016 arXiv

  30. [38]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10684– 10695, 2022. 1, 6, 2

    2022

  31. [39]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500–22510, 2023. 6, 2

    2023

  32. [40]

    Evidence for univer- sality and cultural variation of differential emotion response patterning

    Klaus R Scherer and Harald G Wallbott. Evidence for univer- sality and cultural variation of differential emotion response patterning. Journal of personality and social psychology, 66 (2):310, 1994. 8, 2

    1994

  33. [41]

    Sentiwordnet: A pub- licly available lexical resource for opinion mining

    Fabrizio Sebastiani and Andrea Esuli. Sentiwordnet: A pub- licly available lexical resource for opinion mining. In ELRA, pages 417–422, 2006. 7

    2006

  34. [42]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, pages 422–438. Springer, 2025. 2, 3, 4

    2025

  35. [43]

    Recursive deep models for semantic compositional- ity over a sentiment treebank

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositional- ity over a sentiment treebank. In EMNLP, pages 1631–1642,

  36. [44]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1

    2023 arXiv

  37. [45]

    Attention is all you need

    A Vaswani. Attention is all you need. NeurIPS, 2017. 3

    2017

  38. [46]

    Cider: Consensus-based image description evalua- tion

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evalua- tion. In CVPR, pages 4566–4575, 2015. 7

    2015

  39. [47]

    Diffusers: State-of-the-art diffu- sion models

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffu- sion models. https://github.com/huggingface/ diffusers...

    2022

  40. [48]

    Glue: A multi-task benchmark and analysis platform for natural language understanding

    Alex Wang. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 8, 2

    2018 arXiv

  41. [49]

    Cola: The corpus of linguistic acceptability (with added an- notations)

    Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Cola: The corpus of linguistic acceptability (with added an- notations). 2019. 3, 8, 2

    2019

  42. [50]

    A broad-coverage challenge corpus for sentence understanding through inference

    Adina Williams, Nikita Nangia, and Samuel R Bowman. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426, 2017. 3, 8, 2

    2017 arXiv

  43. [51]

    Ties-merging: Resolving interference when merging models

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raf- fel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. NeurIPS, 36:7093–7115, 2023. 7

    2023

  44. [52]

    Compos- ing parameter-efficient modules with arithmetic operation

    Jinghan Zhang, Junteng Liu, Junxian He, et al. Compos- ing parameter-efficient modules with arithmetic operation. NeurIPS, 36:12589–12610, 2023. 2, 3, 4, 7, 8

    2023

  45. [53]

    Llama-adapter: Efficient fine-tuning of language models with zero-init attention

    Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Ao- jun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3

    2023 arXiv

  46. [54]

    Llamafac- tory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac- tory: Unified efficient fine-tuning of 100+ language models. In ACL, Bangkok, Thailand, 2024. Association for Compu- tational Linguistics. 2, 3

    2024

  47. [55]

    Multi-lora composition for image generation

    Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2, 3 IterIS: Iterative Inference-Solving Alignment for LoRA Merging Suppleme...

    2024 arXiv

  48. [56]

    Therefore, g−1(1) is entirely dependent on the input, indicating that these nodes and their corresponding matrices will remain unchanged in subsequent iterations

    Base Case ( k = 0 ): For the initial iteration, it is clear that g−1(1) is non-empty and contains only nodes with Oin as their sole parent. Therefore, g−1(1) is entirely dependent on the input, indicating that these nodes and their corresponding matrices will remain unchanged ...

  49. [57]

    , k0 + 1}) and their associated matrices remain constant

    Inductive Step (Assume true for k = k0): Assume that at iteration k = k0, all nodes in g−1({1, 2, . . . , k0 + 1}) and their associated matrices remain constant. Since each node ing−1(k0 +

  50. [58]

    [adj n] n,

    has its parent nodes contained withing−1({1, 2, . . . , k0+1}), it follows that the nodes ing−1(k0 + 2)are reliant on the unchanged nodes in g−1({1, 2, . . . , k0 + 1}). Consequently, these nodes and their corresponding updated matrices will also remain constant in subsequent ...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.