Pith. sign in

REVIEW 4 major objections 5 minor 61 references

Adapt Once, Thrive with Updates: Transferable Parameter-Efficient Fine-Tuning on Evolving Base Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that parameter-efficient fine-tuning modules trained on an old base model can be transferred to a continually updated base model without re-tuning, by training them to rely on stable attention patterns rather than…

desk verdict The empirical core—masking and dropping FFN outputs during PEFT fine-tuning to make modules transferable across base-model versions—looks solid and practically useful, but the theoretical guarantee is not: it rests on an unmeasured attention-stability assumption and a bound with a tautological term. read the letter →

arxiv 2506.06844 v1 pith:UJ45PD6R submitted 2025-06-07 cs.CL

Naibin Gu , Peng Fu , Xiyu Liu , Ke Ma , Zheng Lin , Weiping Wang This is my paper
classification cs.CL
keywords parameter-efficientfine-tuningPEFTtransfercontinualpre-trainingbasemodelupdatesattentionstabilityfeed-forwardnetworksLoRAknowledgemasking
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

This paper asks whether a small task-tuned module, trained once on an older version of a large language model, can be dropped onto a newer, continually updated version without retraining. The authors find that base-model updates mainly rewrite task-specific knowledge stored in feed-forward network (FFN) sub-layers, while the task-specific patterns in attention sub-layers stay nearly the same across versions. On that basis they introduce Trans-PEFT, a training-time procedure that randomly masks FFN dimensions and drops whole FFN layers while fine-tuning, pushing the PEFT module to rely on stable attention patterns. Experiments across seven base models and twelve datasets show that transferred Trans-PEFT modules roughly match freshly retuned ones, beating naive direct transfer by up to 30 percentage points. If the finding holds, users and platforms would no longer need to retune thousands of adapters every time the base model is updated.

What carries the argument

The carrying mechanism is a pair of stochastic regularizers applied during fine-tuning on the old base model: Intra-layer Knowledge Masking zeroes random intermediate FFN dimensions with probability $p_i$ per forward pass, and Cross-layer Knowledge Dropping sets whole FFN layer outputs to zero with probability $p_c$. These regularizers force the PEFT parameters to avoid trusting version-specific FFN knowledge and to anchor in attention patterns that are stable across versions. The theoretical engine is a set of assumptions (attention stability $\|W_{\mathrm{att}}^{(1)} - W_{\mathrm{att}}^{(0)}\|_2 \leq \epsilon_{\mathrm{att}} \ll 1$, bounded FFN perturbation, smooth loss, decomposable loss) that yield the Theorem 1 bound; the attention-stability assumption is what removes the attention loss term from the transfer gap.

What would settle it

Measure, for a pair of base model versions, the per-layer weight or activation distance between old and new attention sub-layers, for example $\|W_{\mathrm{att}}^{(\mathrm{new})} - W_{\mathrm{att}}^{(\mathrm{old})}\|_2 / \|W_{\mathrm{att}}^{(\mathrm{old})}\|_2$, and check whether it stays small; if Trans-PEFT's transfer gains disappear on a version pair where this distance is large, the central attention-stability mechanism is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the obstacle to transferring PEFT modules across continually updated base models is not the attention mechanism but the changing knowledge storage in FFN sub-layers. The paper supports this with activation visualizations showing that, after task fine-tuning, attention sub-layer activation distributions look similar across versions while FFN sub-layer distributions differ and even reverse their influence across layers. Trans-PEFT then trains the module on the old version using Intra-layer Knowledge Masking, which randomly zeroes intermediate FFN dimensions, and Cross-layer Knowledge Dropping, which randomly zeroes whole FFN layer outputs, so the module cannot overfit to version-specific knowledge. The paper proves a bound on the transfer loss discrepancy of the form $|L(\theta^{*(0)}; M_1) - L(\theta^{*(0)}; M_0)| \leq L\rho + 2\beta\|\theta_{\mathrm{ffn}}^{*(0)} - \theta_{\mathrm{ffn}}^{*(1)}\| + C(p_i + p_c)$, where the attention term vanishes by assumption and the FFN and regularization terms are controlled. On the tested LoRA, Adapter, and DoRA settings, transferred modules reach performance comparable to retuning on new versions and can even exploit upgraded math and code capabilities in the new base model.

Load-bearing premise

The method assumes that when a base model is continually updated, its attention sub-layers' task-specific behavior stays nearly the same across versions; the paper's support for this is visual inspection of activation heatmaps rather than a quantitative distance measurement, so an update that materially rewrites attention patterns would undermine the whole procedure.

Editorial extensions

If this is right

  • Deploying a PEFT service no longer requires retuning each user module after every base-model update; one training pass on the old version suffices in the tested continual-update regime.
  • Because transferred modules keep their attention anchors, they can inherit base-model improvements, so in math and code tasks they can outperform the old-version fine-tuned module they came from.
  • Since Trans-PEFT only changes the training-time forward pass, it can be layered onto LoRA, Adapter, or DoRA without extra inference cost or architectural changes.
  • Eliminating retuning removes the need to keep users' raw data for future training, directly addressing the privacy and storage overhead of large PEFT deployments.
  • The method is scoped to continually pre-trained updates: re-pretraining that changes initialization or architecture, such as LLaMA2 to LLaMA3, lies outside its guarantees because parameter spaces no longer align.

Reading between the lines

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

  • If attention stability is a general property of continual pre-training, Trans-PEFT could become a routine deploy-time feature for long-lived model families that release frequent small updates.
  • A quantitative attention-stability diagnostic, measuring weight or activation distance between versions, could predict how much transfer gain a given update pair will deliver; the paper does not provide such a metric.
  • The dual regularization suggests a bias-variance trade-off for adaptation, so tuning $p_i$ and $p_c$ per layer or per task, or annealing them during training, might push transfer performance even closer to retuning.
  • Combining Trans-PEFT with data-free or synthetic-data distillation would let a platform update user modules without ever storing user data, closing a remaining privacy gap in current deployments.
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 studies the transfer of parameter-efficient fine-tuning (PEFT) modules from an old base model to a newer continuously pretrained version, without retuning the modules. Based on an analysis of activation distributions, the authors claim that continual pretraining mainly changes task-specific knowledge stored in FFN sub-layers while the task-specific patterns in attention sub-layers remain stable. They propose Trans-PEFT, which applies random intra-layer knowledge masking and cross-layer knowledge dropping to FFN sub-layers during fine-tuning on the old base model, and provide a theorem bounding the loss discrepancy when transferring to the new version. Experiments across seven base models (Qwen, InternLM, DeepSeek families), twelve datasets, and three PEFT types (LoRA, Adapter, DoRA) show that Trans-PEFT substantially outperforms direct transfer and approaches the performance of retuning on the new version, with additional t-tests and multi-seed experiments in the appendix.

Significance. If the empirical finding holds, Trans-PEFT is practically valuable: it removes the need to retune numerous PEFT modules when a base model is updated, reducing computational cost and avoiding long-term storage of user data. The method is simple, applies to mainstream PEFT methods without architectural changes, and the code is released. The evaluation is a genuine strength: seven base models, twelve datasets, three PEFT types, multi-seed runs, and statistical significance tests in the appendix give the central empirical claim substantial support. The theoretical analysis, however, is not sound as presented: the zero-mean perturbation claim in the proof is incorrect, the theorem contains an unquantified term involving the unknown optimal parameters on the new model, and the key assumptions (attention stability and loss decomposability) are not adequately justified. These issues weaken the paper's claim of a theoretical guarantee but do not, by themselves, invalidate the empirical contribution.

major comments (4)
  1. [Appendix A, Eqs. (20)-(23)] The proof claims E_{m,z}[δ(m,z)] = 0, but this is incorrect for the proposed masking and dropping mechanisms. Since m ~ Bernoulli(1-p_i) and z ~ Bernoulli(1-p_c), the expected mask values are E[m]=1-p_i and E[z]=1-p_c, so the perturbation introduced by masking and dropping has nonzero mean proportional to p_i and p_c. Consequently, the first-order term in the Taylor expansion in Eq. (20) does not vanish, and the bound R ≤ C·(p_i + p_c) in Eq. (25) does not follow from the stated derivation. A corrected expansion must retain the first-order term; while an O(p_i+p_c) bound may still be recoverable, the proof as written is not valid.
  2. [Theorem 1, Eq. (10) and Eq. (11)] The bound contains the term 2β∥θ*_ffn^(0) − θ*_ffn^(1)∥, where θ*_ffn^(1) is the hypothetical optimal PEFT parameter for the new model M1. The paper asserts that Trans-PEFT reduces this term but provides no bound or characterization of it in terms of known quantities. As stated, the theorem is a decomposition rather than a guarantee: it does not establish that the transferred parameters are close to the optimal parameters on M1. To substantiate the claim of a theoretical guarantee, the authors must either bound this parameter-deviation term or state clearly that the result is conditional on an unverified closeness property.
  3. [Section 4.2, Assumption 1 and Eq. (14)] Assumption 1 (attention stability, ∥W_att^(1) − W_att^(0)∥ ≤ ϵ_att ≪ 1) is load-bearing for both the theorem and the paper's mechanistic explanation, but its only support is visual similarity of activation heatmaps (Figures 1, 3, and 8). Activation similarity does not imply weight-space closeness: the same attention outputs can arise from different weight matrices when combined with residual connections, LayerNorm, and the downstream FFN block. Furthermore, Eq. (14) sets L_att·ϵ_att ≈ 0 with no measurement of either L_att or ϵ_att. The authors should provide quantitative weight-distance measurements (e.g., Frobenius norms of attention weight differences across versions) or explicitly downgrade this assumption to a stated limitation.
  4. [Section 4.2, Assumption 4 and Eq. (13)] Assumption 4 (loss decomposability, L(θ;M) = L_att(θ_att;M) + L_ffn(θ_ffn;M)) is used in the first equality of Eq. (13) to separate the attention and FFN loss contributions. In a transformer, the attention output feeds directly into the FFN sub-layer, so the loss is not additively separable in general. The paper gives no justification for this assumption beyond the observation that attention patterns appear stable. If the theorem is to remain, the assumption must be justified for a concrete simplified model or clearly marked as a strong modeling assumption that limits the scope of the result.
minor comments (5)
  1. [Section 3.2 and Figure 3] The 'influence of each FFN sub-layer on activation magnitudes' is not formally defined. Please specify the exact computation used to obtain the 'Changes in Activations' values (for example, the norm of the difference between hidden states before and after each FFN layer), so the analysis is reproducible.
  2. [Appendix E.1, Table 3] The t-test table reports only p-values. For completeness, report the mean differences and standard deviations (or effect sizes) for the three runs underlying Trans-PEFT and Direct Transfer, since the claim of significance cannot be evaluated without the magnitude and variance of the improvements.
  3. [Section 5.4, Figure 6(a)] The text says that when p_c = 0, neither of the proposed strategies is employed, but this is true only because Figure 6(a) fixes p_i = 0 as well. Please phrase the description to avoid implying that p_c = 0 alone deactivates both strategies.
  4. [Throughout] The notation 'Fine-tune_n', 'Fine-tune_o', and 'Direct Transfer' is typeset inconsistently (e.g., 'Fine-tunen', 'Fine-tuneo', 'Direct T.'). Please use consistent subscript formatting and define all abbreviations in table captions.
  5. [Appendix A.2, Eq. (19)] The definition of θ*_ffn^(1) as the 'hypothetical optimal PEFT parameter for M1' is ambiguous: it is not stated whether this is the optimum under standard fine-tuning on M1 or under Trans-PEFT's stochastic perturbation. Clarifying this is necessary for the parameter-deviation term to be meaningful.

Circularity Check

2 steps flagged · score 4.0 of 10

Theorem 1's transfer bound is partially tautological: its key term is the distance to the unknown optimal module on M1, which the paper asserts rather than proves, and the appendix proof redefines M0 as the masked model.

  1. self definitional [Section 4.2, Theorem 1, Eq. (10); Appendix A.2, Eqs. (16)-(18)]
    "the parameter deviation term measures the proximity between our transferred PEFT parameters and the optimal PEFT parameters for M1, which can be reduced as Trans-PEFT encourages PEFT modules to capture persistent patterns in attention... |L(θ∗(0);M1)−L(θ∗(0);M0)| ≤ Lρ+2β∥θ∗(0)ffn−θ∗(1)ffn∥+C·(pi+pc), where θ∗(1)ffn is the hypothetical optimal PEFT parameter for M1."

    The headline quantity in the bound is ∥θ*^(0)_ffn − θ*^(1)_ffn∥, the distance to the unobserved optimum on M1. The proof (Eq. 18) only inserts this term via triangle inequality: A1+A2 ≤ 2β∥θ*^(0)_ffn−θ*^(1)_ffn∥. Nothing in the masking/dropping construction is shown to make this distance small; the paper instead asserts that Trans-PEFT reduces it. That is the very claim the theorem is offered to support. Any method whose transferred parameters happen to be close to M1's optimum satisfies the same bound, so as a 'guarantee for Trans-PEFT' the bound is a tautological decomposition rather than a derivation from the method.

  2. other [Appendix A.2, Eq. (12)]
    "Since Trans-PEFT introduces a mask in the FFN module W̃(0)_ffn = W(0)_ffn + δ(m,z), we can therefore derive: |L(θ∗(0);M1)−L(θ∗(0);M0)| = |L(θ∗(0);[W(1)_att,W(1)_ffn])−L(θ∗(0);[W(0)_att,W̃(0)_ffn])|."

    The right-hand side evaluates M0 at W̃^(0)_ffn, the masked model used during Trans-PEFT training, not at the actual W^(0)_ffn appearing in L(θ*;M0). The omitted difference between the two is exactly the method's own stochastic perturbation δ, later bounded as C(pi+pc). The equality is therefore false as written; when repaired, the 'bound' must include a term introduced by the proof's redefinition of the comparison object. This makes the stated theoretical guarantee partially an artifact of construction: the theorem is not bounding the loss on the true old base model but on the method's masked model.

full rationale

The circularity found here is confined to the theory section. The empirical core of the paper—Tables 1-2, Figure 5, and the significance/stability tests—compares Trans-PEFT with Direct Transfer and Fine-tunen on public benchmarks, and those numbers do not depend on Theorem 1. There is no load-bearing self-citation chain: citations to the authors' own prior PEFT papers appear only in related work, and no uniqueness theorem is imported from the authors. The attention-stability assumption is indeed supported only by activation heatmaps (Figures 1-3), with no quantitative weight-distance estimate; that is a missing-support and correctness risk, not itself a circular reduction. The Taylor-expansion step also assumes δ(m,z) is zero-mean even though Bernoulli masking and dropping are not centered, another non-circular correctness flaw. Because Theorem 1's bound is partly tautological and Eq. (12) redefines M0 as the masked model, the 'theoretical support for Trans-PEFT' is partially circular; nevertheless, the central empirical claim stands independently, so the overall score is 4 rather than higher.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method's transfer guarantee rests on the attention-stability and FFN-boundedness assumptions (qualitatively motivated), plus strong loss-decomposability and zero-mean perturbation assumptions that are not established. The empirical results provide independent support for the central claim, which is why the circularity burden is moderate rather than high.

free parameters (3)
  • p_i (intra-layer masking probability) = 0.1, 0.2, 0.3 (commonsense); 0.01, 0.05, 0.1 (math/code)
    Chosen by grid search per task (Appendix C, Section 5.4); controls strength of knowledge masking, central to the method's regularization.
  • p_c (cross-layer dropping probability) = 0.1, 0.2, 0.3
    Chosen by grid search per task; controls how often entire FFN layers are dropped during fine-tuning.
  • LoRA/Adapter rank = 32 or 64
    Standard PEFT capacity hyperparameter; affects the module's expressiveness and transfer, but not a novel free parameter.
assumptions (5)
  • domain assumption Attention Stability: ∥W(1)_att − W(0)_att∥2 ≤ ϵatt ≪ 1
    Assumption 1, Section 4.2. Needed so that attention patterns learned on M0 transfer to M1. Supported only by qualitative activation visualizations.
  • domain assumption FFN Perturbation Boundedness: ∥W(1)_ffn − W(0)_ffn∥2 ≤ ρ
    Assumption 2, Section 4.2. Bounds the FFN weight shift; ρ is never estimated.
  • standard math Loss Smoothness: L is L-Lipschitz in W_att, W_ffn and β-Lipschitz in θ_ffn
    Assumption 3, Section 4.2. Standard but unquantified; L and β are not computed.
  • ad hoc to paper Loss Decomposability: L(θ;M) = L_att(θ_att;M) + L_ffn(θ_ffn;M)
    Assumption 4, Section 4.2. The total loss of a transformer is not generally additive over attention and FFN sub-layers; this is a strong simplification used to split the discrepancy bound.
  • ad hoc to paper Zero-mean perturbation: E[δ(m,z)] = 0
    Appendix A, Eq. (20)-(23). Stated without proof; with Bernoulli(1-p) masking, m−1 has mean −p, so the first-order Taylor term does not obviously vanish. This is a gap in the proof of the regularization term C(p_i+p_c).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adapt Once, Thrive with Updates: Transferable Parameter-Efficient Fine-Tuning on Evolving Base Models." pith.science (2026). https://pith.science/paper/UJ45PD6R

@misc{pith2026250606844,
  author       = {Pith},
  title        = {Pith review of: Adapt Once, Thrive with Updates: Transferable Parameter-Efficient Fine-Tuning on Evolving Base Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UJ45PD6R}},
  note         = {Machine review of arXiv:2506.06844}
}
read the original abstract

Parameter-efficient fine-tuning (PEFT) has become a common method for fine-tuning large language models, where a base model can serve multiple users through PEFT module switching. To enhance user experience, base models require periodic updates. However, once updated, PEFT modules fine-tuned on previous versions often suffer substantial performance degradation on newer versions. Re-tuning these numerous modules to restore performance would incur significant computational costs. Through a comprehensive analysis of the changes that occur during base model updates, we uncover an interesting phenomenon: continual training primarily affects task-specific knowledge stored in Feed-Forward Networks (FFN), while having less impact on the task-specific pattern in the Attention mechanism. Based on these findings, we introduce Trans-PEFT, a novel approach that enhances the PEFT module by focusing on the task-specific pattern while reducing its dependence on certain knowledge in the base model. Further theoretical analysis supports our approach. Extensive experiments across 7 base models and 12 datasets demonstrate that Trans-PEFT trained modules can maintain performance on updated base models without re-tuning, significantly reducing maintenance overhead in real-world applications.

Figures

Figures reproduced from arXiv: 2506.06844 by the authors.

Figure 1
Figure 1. Comparison of activation distributions within [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Comparison of each FFN sub-layer’s influence [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Illustration of Trans-PEFT. The Trans-PEFT approach mainly includes two strategies: intra-layer [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Effects of pc and pi in our proposed strategies. The results are obtained on commonsense reasoning tasks using LoRA. In (a), we fix pi = 0 and vary pc. In (b), we use the optimal setting pc = 0.2 and vary pi . which are generally considered more challenging. In commons…
Figure 7
Figure 7. Figure 7: Fine-tuning time cost and performance of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of activation distributions within [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 10
Figure 10. Figure 10: Results on mathematical reasoning tasks with applying dropping on the attention sub-layer. We evaluate the transfer performance on DeepSeek base models with Adapter. D.2 Further Discussion on the Impact of Model Updates To further validate our findings in Section 3 th…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 16 canonical work pages

  1. [1]

    Armen Aghajanyan, Sonal Gupta, and Luke Zettlemoyer. 2021. https://doi.org/10.18653/v1/2021.acl-long.568 Intrinsic dimensionality explains the effectiveness of language model fine-tuning . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (V...

    doi:10.18653/v1/2021.acl-long.568 2021

  2. [2]

    Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J

    Jacob Austin, Augustus Odena, Maxwell I. Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. 2021. https://arxiv.org/abs/2108.07732 Program synthesis with large language models . CoRR, abs/2108.07732

    arXiv 2021

  3. [3]

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2019. https://arxiv.org/abs/1911.11641 PIQA: reasoning about physical commonsense in natural language . CoRR, abs/1911.11641

    arXiv 2019

  4. [4]

    Lucas Caccia, Alan Ansell, Edoardo Ponti, Ivan Vulić, and Alessandro Sordoni. 2025. https://arxiv.org/abs/2503.08727 Training plug-n-play knowledge modules with deep context distillation . Preprint, arXiv:2503.08727

    arXiv 2025

  5. [5]

    Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, S...

    arXiv 2024

  6. [6]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pond \' e de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bava...

    arXiv 2021

  7. [7]

    Yilong Chen, Junyuan Shang, Zhengyu Zhang, Jiawei Sheng, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, and Haifeng Wang. 2024. https://arxiv.org/abs/2412.05644 Mixture of hidden-dimensions transformer . Preprint, arXiv:2412.05644

    work page Pith review arXiv 2024

  8. [8]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1300 B ool Q : Exploring the surprising difficulty of natural yes/no questions . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language...

    doi:10.18653/v1/n19-1300 2019

Show all 61 references
  1. [9]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. https://arxiv.org/abs/1803.05457 Think you have solved question answering? try arc, the AI2 reasoning challenge . CoRR, abs/1803.05457

    2018 arXiv

  2. [10]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. https://arxiv.org/abs/2110.14168 Training verifiers to solve math word problems . ...

    2021 arXiv

  3. [11]

    DeepSeek-AI, :, Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, Huazuo Gao, Kaige Gao, Wenjun Gao, Ruiqi Ge, Kang Guan, Daya Guo, Jianzhong Guo, Guangbo Hao, Zhewen Hao, Ying He, Wenjie Hu, Panpan Huang, Er...

    2024 arXiv

  4. [12]

    Weilong Dong, Xinwei Wu, Renren Jin, Shaoyang Xu, and Deyi Xiong. 2025. https://aclanthology.org/2025.coling-main.279/ CONTRANS : Weak-to-strong alignment engineering via concept transplantation . In Proceedings of the 31st International Conference on Computational Linguistics...

    2025

  5. [13]

    Jessica Echterhoff, Fartash Faghri, Raviteja Vemulapalli, Ting-Yao Hu, Chun-Liang Li, Oncel Tuzel, and Hadi Pouransari. 2024. https://arxiv.org/abs/2407.09435 Muscle: A model update strategy for compatible llm evolution . Preprint, arXiv:2407.09435

    2024 arXiv

  6. [14]

    Angela Fan, Edouard Grave, and Armand Joulin. 2020. https://openreview.net/forum?id=SylO2yStDr Reducing transformer depth on demand with structured dropout . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . Open...

    2020

  7. [15]

    William Fleshman and Benjamin Van Durme. 2024. https://arxiv.org/abs/2405.15007 Re-adapt: Reverse engineered adaptation of large language models . Preprint, arXiv:2405.15007

    2024 arXiv

  8. [16]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.446 Transformer feed-forward layers are key-value memories . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484--5495, ...

    2021 doi

  9. [17]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  10. [18]

    Naibin Gu, Peng Fu, Xiyu Liu, Bowen Shen, Zheng Lin, and Weiping Wang. 2024. https://doi.org/10.18653/v1/2024.findings-acl.447 Light- PEFT : Lightening parameter-efficient fine-tuning via early pruning . In Findings of the Association for Computational Linguistics: ACL 2024, p...

    2024 doi

  11. [19]

    Naibin Gu, Zhenyu Zhang, Xiyu Liu, Peng Fu, Zheng Lin, Shuohuan Wang, Yu Sun, Hua Wu, Weiping Wang, and Haifeng Wang. 2025. https://arxiv.org/abs/2502.13604 Beamlora: Beam-constraint low-rank adaptation . Preprint, arXiv:2502.13604

    2025 arXiv

  12. [20]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. 2024. https://arxiv.org/abs/2401.14196 Deepseek-coder: When the large language model meets programming -- the rise of code...

    2024 arXiv

  13. [21]

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024. https://openreview.net/forum?id=lIsCS8b6zj Parameter-efficient fine-tuning for large models: A comprehensive survey . Transactions on Machine Learning Research

    2024

  14. [22]

    Yaru Hao, Li Dong, Furu Wei, and Ke Xu. 2021. https://arxiv.org/abs/2004.11207 Self-attention attribution: Interpreting information interactions inside transformer . Preprint, arXiv:2004.11207

    2021 arXiv

  15. [23]

    Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2022. https://openreview.net/forum?id=0RDcd5Axok Towards a unified view of parameter-efficient transfer learning . In International Conference on Learning Representations

    2022

  16. [24]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. https://arxiv.org/abs/2103.03874 Measuring mathematical problem solving with the MATH dataset . CoRR, abs/2103.03874

    2021 arXiv

  17. [25]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. http://proceedings.mlr.press/v97/houlsby19a.html Parameter-efficient transfer learning for NLP . In Proceedings of the 36th Int...

    2019

  18. [26]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    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 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representat...

    2022

  19. [27]

    Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.319 LLM -adapters: An adapter family for parameter-efficient fine-tuning of large language models . In Proceedings of ...

    2023 doi

  20. [28]

    Shih-Cheng Huang, Pin-Zu Li, Yu-chi Hsu, Kuang-Ming Chen, Yu Tung Lin, Shih-Kai Hsiao, Richard Tsai, and Hung-yi Lee. 2024. https://doi.org/10.18653/v1/2024.acl-long.590 Chat vector: A simple approach to equip LLM s with instruction following and model alignment in new languag...

    2024 doi

  21. [29]

    Moritz Imfeld, Jacopo Graldi, Marco Giordano, Thomas Hofmann, Sotiris Anagnostidis, and Sidak Pal Singh. 2024. https://openreview.net/forum?id=LjeqMvQpen Transformer fusion with optimal transport . In The Twelfth International Conference on Learning Representations

    2024

  22. [30]

    Ting Jiang, Shaohan Huang, Shengyue Luo, Zihan Zhang, Haizhen Huang, Furu Wei, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. 2024. https://arxiv.org/abs/2405.12130 Mora: High-rank updating for parameter-efficient fine-tuning . Preprint, arXiv:2405.12130

    2024 arXiv

  23. [31]

    Feihu Jin, Jiajun Zhang, and Chengqing Zong. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.22 Parameter-efficient tuning for large language model without calculating its gradients . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, ...

    2023 doi

  24. [32]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.243 The power of scale for parameter-efficient prompt tuning . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045--3059, Online a...

    2021 doi

  25. [33]

    Brian Lester, Joshua Yurtsever, Siamak Shakeri, and Noah Constant. 2022. https://arxiv.org/abs/2208.05577 Reducing retraining by recycling parameter-efficient prompts . Preprint, arXiv:2208.05577

    2022 arXiv

  26. [34]

    Xiang Lisa Li and Percy Liang. 2021. https://doi.org/10.18653/v1/2021.acl-long.353 Prefix-tuning: Optimizing continuous prompts for generation . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conferen...

    2021 doi

  27. [35]

    Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. 2023. https://openreview.net/forum?id=1qvx610Cu7 Is your code generated by chat GPT really correct? rigorous evaluation of large language models for code generation . In Thirty-seventh Conference on Neural Informa...

    2023

  28. [36]

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. https://proceedings.mlr.press/v235/liu24bn.html D o RA : Weight-decomposed low-rank adaptation . In Proceedings of the 41st International Conference on M...

    2024

  29. [37]

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. https://doi.org/10.18653/v1/2022.acl-short.8 P -tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks . In Proceedings of the 60th Annual Meeting of the Associat...

    2022 doi

  30. [38]

    Fanxu Meng, Zhaohui Wang, and Muhan Zhang. 2024. https://openreview.net/forum?id=6ZBHIEtdP4 Pi SSA : Principal singular values and singular vectors adaptation of large language models . In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  31. [39]

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. https://doi.org/10.18653/v1/D18-1260 Can a suit of armor conduct electricity? a new dataset for open book question answering . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language P...

    2018 doi

  32. [40]

    OpenAI. 2023. https://api.semanticscholar.org/CorpusID:266362871 Gpt-4 technical report . ArXiv, abs/2303.08774

    2023 arXiv

  33. [41]

    Yujia Qin, Cheng Qian, Xu Han, Yankai Lin, Huadong Wang, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2023. https://doi.org/10.18653/v1/2023.findings-acl.723 Recyclable tuning for continual pre-training . In Findings of the Association for Computational Linguistics: AC...

    2023 doi

  34. [42]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  35. [43]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. https://arxiv.org/abs/1907.10641 WINOGRANDE: an adversarial winograd schema challenge at scale . CoRR, abs/1907.10641

    2019 arXiv

  36. [44]

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. https://doi.org/10.18653/v1/D19-1454 Social IQ a: Commonsense reasoning about social interactions . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9...

    2019 doi

  37. [45]

    Rana Muhammad Shahroz, Pingzhi Li, Sukwon Yun, Zhenyu Wang, Shahriar Nirjon, Chau - Wai Wong, and Tianlong Chen. 2025. https://openreview.net/forum?id=gyHoR6uFhU Portllm: Personalizing evolving large language models with training-free and portable model patches . In The Thirte...

    2025

  38. [46]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. https://arxiv.org/abs/2402.03300 Deepseekmath: Pushing the limits of mathematical reasoning in open language models . Preprint, arXiv:2402.03300

    2024 arXiv

  39. [47]

    Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Huadong Wang, Kaiyue Wen, Zhiyuan Liu, Peng Li, Juanzi Li, Lei Hou, Maosong Sun, and Jie Zhou. 2022. https://doi.org/10.18653/v1/2022.naacl-main.290 On transferability of prompt tuning for natural language processi...

    2022 doi

  40. [48]

    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,...

    2023 doi arXiv

  41. [49]

    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

  42. [50]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. https://arxiv.org/abs/1706.03762 Attention is all you need . CoRR, abs/1706.03762

    2017 arXiv

  43. [51]

    Runqian Wang, Soumya Ghosh, David Cox, Diego Antognini, Aude Oliva, Rogerio Feris, and Leonid Karlinsky. 2024. https://arxiv.org/abs/2405.17258 Trans-LoRA : towards data-free transferable parameter efficient finetuning . Preprint, arXiv:2405.17258

    2024 arXiv

  44. [52]

    Zijun Wu, Yongkang Wu, and Lili Mou. 2024. https://openreview.net/forum?id=26XphugOcS Zero-shot continuous prompt transfer: Generalizing task semantics across language models . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May...

    2024

  45. [53]

    Guangxuan Xiao, Ji Lin, and Song Han. 2023. https://arxiv.org/abs/2302.04870 Offsite-tuning: Transfer learning without full model . Preprint, arXiv:2302.04870

    2023 arXiv

  46. [54]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze...

    2024 arXiv

  47. [55]

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng YU, Zhengying Liu, Yu Zhang, James Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2024. https://openreview.net/forum?id=N8N0hgNDRt Metamath: Bootstrap your own mathematical questions for large language models . In The Twelfth Inte...

    2024

  48. [56]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. https://doi.org/10.18653/v1/P19-1472 H ella S wag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791--4...

    2019 doi

  49. [57]

    Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Yang You, Guiming Xie, Xuejian Gong, and Kunlong Zhou. 2025. https://openreview.net/forum?id=s7DkcgpRxL Train small, infer large: Memory-efficient lora training for large language models . In The Thirteenth International Confe...

    2025

  50. [58]

    Kaiyan Zhang, Ning Ding, Biqing Qi, Xuekai Zhu, Xinwei Long, and Bowen Zhou. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.597 CR a S h: Clustering, removing, and sharing enhance fine-tuning without full large language model . In Proceedings of the 2023 Conference on Empir...

    2023 doi

  51. [59]

    Tianyu Zheng, Ge Zhang, Tianhao Shen, Xueling Liu, Bill Yuchen Lin, Jie Fu, Wenhu Chen, and Xiang Yue. 2025. https://arxiv.org/abs/2402.14658 Opencodeinterpreter: Integrating code generation with execution and refinement . Preprint, arXiv:2402.14658

    2025 arXiv

  52. [60]

    online" 'onlinestring :=

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

  53. [61]

    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 7, 2026 · model on record in the stance chip above.