Pith. sign in

REVIEW 4 major objections 5 minor 38 references

SSMLoRA: Enhancing Low-Rank Adaptation with State Space Model

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

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

Pith's one-line read SSMLoRA matches LoRA's GLUE performance with half the parameters.

desk verdict A novel but unproven PEFT variant: the sparse-SSM adapter idea is worth a serious look, but the paper's efficiency claims rest on batch-statistics-dependent normalization and single-seed experiments. read the letter →

arxiv 2502.04958 v1 pith:357F6G3S submitted 2025-02-07 cs.CL

Jiayang Yu , Yihang Zhang , Bin Wang , Peiqin Lin , Yongkang Liu , Shi Feng This is my paper · ORCID
classification cs.CL
keywords SSMLoRAparameter-efficientfine-tuningLostatespacemodellow-rankadaptationGLUEbenchmarklong-contextsparseinsertion
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

SSMLoRA is a parameter-efficient fine-tuning method that builds on LoRA by adding a state-space recurrence between the low-rank adapter matrices placed in a pretrained transformer. The paper claims that this interconnection lets the model keep most of LoRA's accuracy while using roughly half the trainable parameters, because adapters can be inserted more sparsely without losing performance. The claim matters because LoRA's parameter budget is often spent unevenly across layers, and a cheaper substitute would make large-model adaptation easier to store, deploy, and scale. The paper also reports gains on longer-text and sentence-pair tasks, suggesting the state connection contributes more than a parameter discount.

What carries the argument

The load-bearing object is the Time Module: a LoRA pair $W_a \in \mathbb{R}^{d \times r}$, $W_b \in \mathbb{R}^{r \times d}$ augmented by two $r \times r$ matrices $W_c$ and $W_d$ that carry a state vector $h_t$ from one adapter to the next along the time axis. The state update $h_{t+1} = h_t W_c + x_{\mathrm{new}} W_d + h_t$ is a Taylor-expansion discretization of the classical state-space equation, avoiding the explicit discretization of structured state-space models while keeping their FFT-based parallel training. Before the state modulates the output, it is min-max normalized over batch statistics, $h^{\mathrm{norm}}_{t+1} = (h_{t+1} - \min)/(\max - \min + \epsilon)$, and the adapter output becomes $y = x W_0 + (x W_a + h^{\mathrm{norm}}_{t+1}) W_b$. Separate time axes are used for query, key, and value projections, and sparse alternating insertion into Q and V is what cuts the parameter count roughly in half.

What would settle it

Fine-tune SSMLoRA and LoRA on a fixed GLUE task with batch size 32, then evaluate both at batch size 1 and at batch size 64. If SSMLoRA's score drops several points relative to LoRA under that mismatch, the central half-the-parameters, same-performance claim fails in practical deployment; if the scores stay stable, the acknowledged limitation is not load-bearing.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the low-rank update matrices of LoRA need not be independent islands: by chaining them along a time axis with a state vector $h_t$ that is updated from one adapter to the next, the model can support much sparser insertion, for example alternating between query and value projections in attention, and still match or beat LoRA on GLUE, SuperGLUE, SQuAD, NarrativeQA, and RACE. With roughly half of LoRA's trainable parameters at matched rank, SSMLoRA reports comparable aggregate accuracy and specifically stronger results on MRPC, QNLI, BoolQ, MultiRC, and the high-difficulty RACE split. The intended conclusion is that SSMLoRA is a more parameter-efficient substitute for LoRA that also inherits state-space strengths on longer sequences.

Load-bearing premise

The practical claim stands or falls on the assumption that the model can be trained and evaluated with the same batch size, because the min-max normalization of the state vector uses batch statistics and the paper's Limitations section concedes that a train/inference batch-size mismatch causes a performance drop.

Editorial extensions

If this is right

  • At the same rank, SSMLoRA uses roughly half of LoRA's trainable parameters on RoBERTa-large and GPT-2, and under 80% on RoBERTa-base and LLaMA2, while keeping GLUE and SuperGLUE scores comparable or better.
  • The alternating sparse-insertion pattern shows that a unified adapter slot is not necessary: the state connection between adapters supplies enough capacity to let whole query or value projections be skipped.
  • On sequence-length-binned NarrativeQA, SSMLoRA's ROUGE-L improves relative to LoRA in the over-1000-token bins, which the paper reads as evidence that the state-space structure helps with longer inputs.
  • Because $W_b$, $W_c$, and $W_d$ start at zero, SSMLoRA begins as a sparsified LoRA and gradually learns cross-layer state dependencies, so it can be inserted into existing LoRA training pipelines without changing the pretrained weights.

Reading between the lines

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

  • The recurrence runs over transformer layers rather than over token positions, so the long-context benefit is better described as cross-layer state propagation than as long-range temporal memory; testing on input-length scaling with a fixed layer count would separate those two mechanisms.
  • If the batch-statistics sensitivity is fixed, for example by replacing min-max normalization with an instance-independent normalization, SSMLoRA's parameter savings could extend to streaming and variable-batch inference, which the current design does not cover.
  • The sparse-insertion principle suggests a search problem: given a fixed parameter budget, which layers and which projections should carry Time Modules to maximize downstream accuracy; the alternating pattern in the paper is one feasible point in that space.
  • A natural extension is to stack the state update over both layers and token positions, turning the Time Module into a two-dimensional state-space adapter; the paper does not explore this, but its equations extend directly.
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 SSMLoRA, an extension of LoRA in which low-rank adapters inserted into the same type of pretrained layer are connected by a recurrent state update based on state-space-model equations. The method is evaluated on GLUE, SuperGLUE, SQuAD, NarrativeQA, and RACE using RoBERTa-base/large, DeBERTaV3-base, GPT-2, and LLaMA2-7B/13B. The main claims are that SSMLoRA matches LoRA's performance with roughly half the trainable parameters, is more robust than LoRA to sparse insertion patterns, and shows promise for long-context tasks.

Significance. If the central claim were established, SSMLoRA would be a practically useful parameter-efficient fine-tuning method: it offers a drop-in alternative to LoRA with larger parameter reduction and, potentially, better handling of long inputs. The paper also provides a code repository and covers a wide range of models and benchmarks, which is commendable. However, the current evidence does not establish the central claim because the method's output depends on min-max batch statistics, the experiments lack statistical rigor, and the long-context narrative is not supported by the architecture.

major comments (4)
  1. [Section 3.3, Eq. (8)] The min-max normalization of h_{t+1} is written without specifying the reduction axes. If the min and max are computed over the batch or over the sequence dimension, the normalized state, and therefore the adapter output y = (x W_a + h_norm) W_b in Eqs. (9)-(10), changes with evaluation batch composition and size. This is not a cosmetic issue: the Limitations section concedes that when training and test batch sizes differ, 'this inevitably results in a performance drop,' and no LoRA-style adapter has this failure mode. The headline claim that SSMLoRA is a practical, parameter-efficient drop-in replacement for LoRA therefore requires either a batch-independent normalization or a systematic study of sensitivity to batch size and sequence length at inference.
  2. [Tables 6-8] All reported results are single runs with no seeds or error bars, and many entries are identical across methods or across ranks, e.g., Table 6 shows RTE 81.23 for both LoRA and SSMLoRA at r=1 and r=2, and WNLI 56.34 is repeated in most rows; Table 8 repeats WSC 63.46 throughout. Large gaps in the opposite direction also appear, such as Table 7 CoLA r=16 with 1.81 for SSMLoRA versus 8.16 for LoRA, and Table 6 QQP r=8 with 79.74/85.72 for SSMLoRA versus 87.05/90.19 for LoRA. Without multiple seeds, variance estimates, and a statement that training and evaluation batch sizes matched, the central claim of 'comparable performance across most datasets' is not supported.
  3. [Section 3.2 and Figure 1] The recurrence h_{t+1} = h_t W_c + x_new W_d + h_t runs over Time Modules attached to the same type of layer in successive transformer layers, not over token positions. The long-context rationale in Section 4.5 and the abstract's statement that the method is 'promising in handling tasks with longer input sequences' therefore do not follow from the SSM formalism. The length-binned NarrativeQA evidence in Table 5 and Figure 2 shows small differences without error bars and cannot carry this claim. The authors should either formulate a recurrence over input positions or remove the long-context claims.
  4. [Section 3.2] The text states that h_t is 'detached from the computational graph and does not participate in training.' This means the state is not a differentiable function of previous modules' activations, so it is unclear in what sense SSMLoRA 'leverages the computations from the previous low-rank space' beyond a fixed forward recurrence with learned W_c and W_d. The authors should clarify the gradient flow through the state and provide an ablation that isolates the contribution of the state recurrence, e.g., setting W_c and W_d to zero.
minor comments (5)
  1. [Table 6 caption] The caption contains the typo 'modle:RoBERTa-large'; it should read 'model:RoBERTa-large.'
  2. [Section 4.1] The WSC entry is listed as 'WSC; ()' with no citation; a proper reference should be supplied.
  3. [Appendix B] The text 'Table 5, corresponding to , presents...' has an empty cross-reference, and the citations to 'Section 4.2.3' and 'Section 4.3.3' do not match any section in the paper.
  4. [Section 4.7 and Table 12] The claim that SSMLoRA's memory advantage 'becomes increasingly pronounced as batch sizes grow' is not supported by Table 12, which reports only batch size 1; the text also refers to an inference-time overhead of 0.95 seconds at 7000 tokens (5.200s vs. 4.050s) as 'no significant inference latency overhead,' which should be quantified and justified.
  5. [Table 4] The column header reads 'All Middle High' but omits 'Low'; please state whether the low-difficulty subset was included and report its score.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: SSMLoRA's performance claims rest on externally anchored benchmark comparisons, not on fitted inputs or load-bearing self-citations.

full rationale

SSMLoRA's central claims—comparable GLUE performance with roughly half of LoRA's parameters and gains on sentence-pair and long-text tasks—are empirical benchmark comparisons against externally specified baselines such as LoRA, BitFit, QLoRA, and MixLoRA, with reported parameter counts and scores. I found no derivation chain in which a prediction is obtained from a fitted parameter, no uniqueness theorem imported from the authors' prior work, and no ansatz whose only support is a self-citation. The only author self-citation, Liu et al. (2024b, HiFT), appears in the introduction as background and is not load-bearing for SSMLoRA's validity. The min-max normalization in Eq. (8), combined with the Limitations concession that differing train/test batch sizes cause a performance drop, is a real robustness and generalization concern, but it does not make the reported benchmark numbers equivalent to the method's assumptions by construction; it is an empirical limitation, not a circular step. The sparse-insertion strategy is justified partly by the authors' own findings and partly by prior LoRA insertion studies, but a post-hoc rationale for a design choice is not circularity because the benchmark comparisons themselves remain externally anchored. Overall, the paper's derivation and evaluation chain is self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claims depend on a set of hand-chosen hyperparameters (rank, insertion pattern, alpha) and on unproven architectural assumptions about layer-wise state propagation and batch-dependent normalization.

free parameters (5)
  • Rank r = 8 for main results; scanned 1,2,4,8,16
    The low-rank dimension of Wa, Wb, Wc, and Wd; central to parameter count and capacity.
  • Scaling factor alpha = 16
    LoRA scaling hyperparameter, fixed for all experiments.
  • Sparse insertion pattern = alternating Q/V skip-one
    Chosen as a simple policy; parameter counts depend on it, and performance could vary with the pattern.
  • Learning rate range = 5e-4 to 1e-6
    Varied per task; no per-task selection criterion reported.
  • Dropout = 0.1
    Applied to all LoRA-derived techniques.
assumptions (4)
  • standard math Taylor expansion h_{t+1} = h'_t + h_t approximates the continuous SSM discretization (Eq. 3-4).
    Assumed to be a valid discretization without formal derivation.
  • domain assumption Layer-wise state propagation improves long-input handling.
    The state recurrence is over layers, not token positions, yet the paper attributes long-context gains to SSM properties (Section 4.5).
  • ad hoc to paper Min-max normalization of h does not harm gradient flow or training stability.
    The normalization is non-standard and no analysis of its effect is provided (Eq. 8).
  • ad hoc to paper Detaching h_t from the computational graph preserves the SSM benefit.
    Section 3.2 states h_t does not participate in training, which means no backprop through time; the temporal coupling is not optimized end-to-end.
invented entities (2)
  • Time Module
    purpose: Connects low-rank adapters across layers via a state-space recurrence with matrices Wc and Wd.
    New architectural component; no external validation outside the paper.
  • Time axis over layers
    purpose: Defines the ordering of states propagated across transformer layers.
    Introduced concept; not established by prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SSMLoRA: Enhancing Low-Rank Adaptation with State Space Model." pith.science (2026). https://pith.science/paper/357F6G3S

@misc{pith2026250204958,
  author       = {Pith},
  title        = {Pith review of: SSMLoRA: Enhancing Low-Rank Adaptation with State Space Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/357F6G3S}},
  note         = {Machine review of arXiv:2502.04958}
}
read the original abstract

Fine-tuning is a key approach for adapting language models to specific downstream tasks, but updating all model parameters becomes impractical as model sizes increase. Parameter-Efficient Fine-Tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address this challenge by introducing additional adaptation parameters into pre-trained weight matrices. However, LoRA's performance varies across different insertion points within the model, highlighting potential parameter inefficiency due to unnecessary insertions. To this end, we propose SSMLoRA (State Space Model Low-Rank Adaptation), an extension of LoRA that incorporates a State Space Model (SSM) to interconnect low-rank matrices. SSMLoRA ensures that performance is maintained even with sparser insertions. SSMLoRA allows the model to not only map inputs to a low-rank space for better feature extraction but also leverage the computations from the previous low-rank space. Our method achieves comparable performance to LoRA on the General Language Understanding Evaluation (GLUE) benchmark while using only half the parameters. Additionally, due to its structure, SSMLoRA shows promise in handling tasks with longer input sequences. .You can find our code here:https://github.com/yuhkalhic/SSMLoRA.

Figures

Figures reproduced from arXiv: 2502.04958 by the authors.

Figure 1
Figure 1. This image illustrates how our technique is applied to the original pre-trained model. In the lower part of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. This figure shows the results of the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 5 canonical work pages

  1. [1]

    Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. 2015. http://arxiv.org/abs/1512.08571 Structured pruning of deep convolutional neural networks

    work page Pith review arXiv 2015

  2. [2]

    Daniel Cer, Mona Diab, Eneko Agirre, I \ n igo Lopez-Gazpio, and Lucia Specia. 2017. https://doi.org/10.18653/v1/S17-2001 S em E val-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation . In Proceedings of the 11th International Workshop on Semantic Evaluation ( S em E val-2017) , pages 1--14, Vancouver, Canada. Associ...

    doi:10.18653/v1/s17-2001 2017

  3. [3]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. http://arxiv.org/abs/1905.10044 Boolq: Exploring the surprising difficulty of natural yes/no questions

    arXiv 2019

  4. [4]

    Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Machine learning challenges workshop, pages 177--190. Springer

    2005

  5. [5]

    Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107--124

    2019

  6. [7]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023 b . Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314

    arXiv 2023

  7. [8]

    Ning Ding, Xingtai Lv, Qiaosen Wang, Yulin Chen, Bowen Zhou, Zhiyuan Liu, and Maosong Sun. 2023. Sparse low-rank adaptation of pre-trained language models. arXiv preprint arXiv:2311.11696

    arXiv 2023

  8. [9]

    Dolan and Chris Brockett

    William B. Dolan and Chris Brockett. 2005. https://aclanthology.org/I05-5002 Automatically constructing a corpus of sentential paraphrases . In Proceedings of the Third International Workshop on Paraphrasing ( IWP 2005)

    2005

Show all 38 references
  1. [10]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    2023 arXiv

  2. [11]

    Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher R\' e . 2020. Hippo: Recurrent memory with optimal polynomial projections. arXiv preprint arXiv:2008.07669

    2020 arXiv

  3. [12]

    Albert Gu, Karan Goel, and Christopher R\'e. 2022. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations ( ICLR )

    2022

  4. [13]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. http://arxiv.org/abs/2006.03654 Deberta: Decoding-enhanced bert with disentangled attention

    2021 arXiv

  5. [14]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. http://arxiv.org/abs/1902.00751 Parameter-efficient transfer learning for nlp

    2019 arXiv

  6. [15]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  7. [16]

    Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. Looking beyond the surface:a challenge set for reading comprehension over multiple sentences. In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL)

    2018

  8. [17]

    Kopiczko, Tijmen Blankevoort, and Yuki M

    Dawid J. Kopiczko, Tijmen Blankevoort, and Yuki M. Asano. 2024. http://arxiv.org/abs/2310.11454 Vera: Vector-based random matrix adaptation

    2024 arXiv

  9. [18]

    Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. 2017. http://arxiv.org/abs/1712.07040 The narrativeqa reading comprehension challenge

    2017 arXiv

  10. [19]

    Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. https://doi.org/10.18653/v1/D17-1082 RACE : Large-scale R e A ding comprehension dataset from examinations . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages...

    2017 doi

  11. [20]

    Dengchun Li, Yingzi Ma, Naizheng Wang, Zhengmao Ye, Zhiyuan Cheng, Yinghao Tang, Yan Zhang, Lei Duan, Jie Zuo, Cal Yang, and Mingjie Tang. 2024. http://arxiv.org/abs/2404.15159 Mixlora: Enhancing large language models fine-tuning with lora-based mixture of experts

    2024 arXiv

  12. [21]

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024 a . Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353

    2024 arXiv

  13. [22]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. http://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach

    2019 arXiv

  14. [23]

    Yongkang Liu, Yiqun Zhang, Qian Li, Tong Liu, Shi Feng, Daling Wang, Yifei Zhang, and Hinrich Schütze. 2024 b . http://arxiv.org/abs/2401.15207 Hift: A hierarchical full parameter fine-tuning strategy

    2024 arXiv

  15. [24]

    Seongmin Park and Jihwa Lee. 2021. https://doi.org/10.18653/v1/2021.insights-1.5 Finetuning pretrained transformers into variational autoencoders . In Proceedings of the Second Workshop on Insights from Negative Results in NLP, pages 29--35, Online and Punta Cana, Dominican Re...

    2021 doi

  16. [25]

    Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. http://arxiv.org/abs/1808.09121 Wic: the word-in-context dataset for evaluating context-sensitive meaning representations

    2019 arXiv

  17. [26]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners

    2019

  18. [27]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://doi.org/10.18653/v1/D16-1264 SQ u AD : 100,000+ questions for machine comprehension of text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-...

    2016 doi

  19. [28]

    Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In 2011 AAAI spring symposium series

    2011

  20. [29]

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language...

    2013

  21. [30]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, W...

    2023 arXiv

  22. [31]

    Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2018. Neural network acceptability judgments. arXiv preprint arXiv:1805.12471

    2018 arXiv

  23. [32]

    Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. 2016. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems

    2016

  24. [33]

    Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/N18-1101 A broad-coverage challenge corpus for sentence understanding through inference . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computation...

    2018 doi

  25. [34]

    Guotian Xie, Jingdong Wang, Ting Zhang, Jianhuang Lai, Richang Hong, and Guo-Jun Qi. 2018. http://arxiv.org/abs/1804.06202 Igcv 2 : Interleaved structured sparse convolutional neural networks

    2018 arXiv

  26. [35]

    Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. 2022. http://arxiv.org/abs/2106.10199 Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models

    2022

  27. [36]

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. http://arxiv.org/abs/2303.10512 Adalora: Adaptive budget allocation for parameter-efficient fine-tuning

    2023 arXiv

  28. [37]

    Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme. 2018. http://arxiv.org/abs/1810.12885 Record: Bridging the gap between human and machine commonsense reading comprehension

    2018 arXiv

  29. [38]

    URL: " 'urlintro :=

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

  30. [39]

    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 8, 2026 · model on record in the stance chip above.