Pith. sign in

REVIEW 4 major objections 5 minor 71 references

Composable Cross-prompt Essay Scoring by Merging 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 claims that an essay scorer for an unseen prompt can be composed by linearly mixing per-prompt adapters, and that the mix outperforms joint training on all source prompts.

desk verdict Useful merging-based source-free AES with an overstated headline; cross-dataset results are solid, but the Llama in-dataset row contradicts the 'consistently outperforms joint training' claim. read the letter →

arxiv 2505.18548 v1 pith:LFE3QPEQ submitted 2025-05-24 cs.CL

Sanwoo Lee , Kun Liang , Yunfang Wu This is my paper · ORCID
classification cs.CL
keywords cross-promptautomatedessayscoringsource-freedomainadaptationmodelmergingtaskvectorsLoRABayesianoptimizationinformationmaximizationquadraticweightedkappa
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that cross-prompt automated essay scoring can be done without the labeled source essays or labels on the target prompt: take several scoring adapters, each fine-tuned on a different prompt, and mix their parameter updates linearly. The mixing coefficients are chosen to maximize a new unsupervised criterion, PIM, which asks the merged model to make confident score predictions whose score distribution resembles a prior estimated from the sources. Bayesian optimization searches the coefficients without training. The authors report that this procedure consistently beats the standard practice of jointly training on all source prompts, and holds up better than recent cross-prompt methods when the dataset changes sharply. A sympathetic reader would care because it removes the main privacy and scalability obstacles to deploying prompt-adaptive essay scoring.

What carries the argument

The load-bearing object is the task vector $\tau_j = \theta_j - \theta_{\mathrm{pre}}$, the LoRA parameter update obtained by fine-tuning on source prompt $j$; merging forms $\theta_{\mathrm{pre}} + \sum_j \lambda_j \tau_j$, so the whole adaptation is a coefficient search. The coefficients are scored by PIM, $f(\lambda) = -\mathrm{KL}(p(y|\lambda)\,\|\,q(y)) - H(p(y|x,\lambda))$: the first term rewards a merged model whose marginal score distribution matches the source-derived prior $q(y)$, and the second rewards sharp per-essay score predictions. The prior $q(y)$ itself comes from fitting $\mathrm{Beta}(\alpha_j, \beta_j)$ to each source's rescaled scores and pooling them by moment matching. Bayesian optimization with expected improvement treats $f$ as a black box, so no gradients or target labels are needed.

What would settle it

Take a held-out target prompt and run the same Bayesian-optimization search with coefficients allowed in a wider box, for example $\lambda_j \in [-2, 2]$, and compare QWK with the paper's $[0,1]$ restriction; if wider coefficients improve the score, the convex-hull assumption is load-bearing, and if a target prompt with a genuinely novel rubric degrades under both, the reliance on at least one reasonable source model is the binding constraint.

Watch

Extended reading notes

Core claim

The paper's central claim is that source-free adaptation to a new essay prompt reduces to choosing $M$ scalar weights: the adapted model is the pre-trained LLM plus a weighted sum of LoRA task vectors, and the weights that maximize PIM yield higher quadratic-weighted-kappa agreement than joint training on all sources. PIM replaces the uniform label distribution of classical mutual-information maximization with a categorical prior $q(y)$ built by fitting Beta distributions to each source prompt's scaled score histogram and pooling them. With Bayesian optimization searching the weights in $[0,1]^M$, the method is reported to beat other merging strategies on in-dataset transfer (ASAP to ASAP) and to remain the strongest option when the distribution shifts severely across datasets (ASAP to PERSUADE2.0), while using less adaptation time than a joint-training baseline.

Load-bearing premise

The central load-bearing premise is that the ideal target-prompt model lies inside the family given by the pre-trained model plus a weighted average of the source adapters, with weights between zero and one, and that at least one source adapter already gives sensible score predictions on target essays; if the true ideal scorer needs a parameter change outside that family, no coefficient search can recover it and PIM merely picks the best model in the wrong set.

Editorial extensions

If this is right

  • New target prompts can be served by a short black-box search over already-trained adapters, without re-running multi-source training.
  • Source essays never need to be redistributed, because only the adapters and their score statistics leave the data holder.
  • The method adapts to score ranges unseen in any single source, since the prior is computed from rescaled Beta fits rather than from target labels.
  • Under severe cross-dataset shifts the reported gap over joint training widens, so the approach is positioned for exactly the settings where recent cross-prompt methods degrade.
  • Because each additional target prompt reuses the same source adapters, the marginal adaptation cost stays small compared with retraining a joint model.

Reading between the lines

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

  • An extension implicit in the paper: the same PIM-plus-Bayesian-search recipe should transfer to other discrete-label LLM tasks, such as rubric-based trait scoring or document grading, where the only needed signals are next-token probabilities over the label set and a source-derived prior.
  • A testable boundary: if the coefficients are allowed to go negative or to exceed 1, and performance on some target improves, then the convex-hull restriction is what limits the method; if it never improves, the reported gains come from coefficient search within that family.
  • The Beta prior could be made adaptive rather than fixed, using the merged model's own confident predictions on target essays to update $q(y)$ across Bayesian-optimization rounds, turning the source prior into a warm start.
  • Since the method's success on the widest score range (the 0 to 60 prompt) is the clearest stress case, a targeted study of how range mismatch between source and target propagates through the Beta fit would show when the prior becomes a liability.
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 a source-free cross-prompt automated essay scoring method that merges LoRA adapters fine-tuned on individual source prompts. The merged model is formed as a linear combination of task vectors, and the mixing coefficients are optimized by Bayesian optimization to maximize a proposed Prior-encoded Information Maximization (PIM) objective computed on unlabeled target essays. Experiments with Llama-3.1-8B-Instruct and Phi-4-mini-instruct compare the method against merging baselines, joint training on all sources, and two cross-prompt AES methods, in both in-dataset (ASAP to ASAP) and cross-dataset (ASAP to PERSUADE2.0) settings.

Significance. If the claims are supported, the paper offers a useful privacy-preserving and computationally efficient alternative to joint training for cross-prompt AES, with the additional benefit of source selection. The study is substantial in scope: it uses two LLMs, two transfer settings, eight target prompts per setting, several merging baselines, an ablation, a cost analysis, and a comparison with established cross-prompt methods. The proposed PIM criterion is a principled attempt to inject source-derived priors into an unsupervised objective, and the use of Bayesian optimization to avoid backpropagation through the merged model is pragmatic.

major comments (4)
  1. [Abstract / Table 3] The abstract and Section 6 claim that PIM 'consistently outperforms training jointly on all sources.' Table 3 (Llama-3.1-8B-it, ASAP to ASAP) does not support this: PIM averages 0.602 QWK versus 0.601 for joint training, neither row carries a significance marker, and joint training is numerically higher on P4 (0.656 vs 0.647), P6 (0.760 vs 0.690), and P8 (0.257 vs 0.152). Please report per-prompt confidence intervals or paired significance tests across the five seeds, and revise the wording to 'generally outperforms' or 'outperforms in most settings' unless stronger evidence is provided.
  2. [Eq. (2) / Sec. 4.1] Equation (2) defines the mixing coefficients as λj ∈ R, but the implementation in Section 4.1 restricts each λj to [0, 1]. This search space excludes negative coefficients and values greater than 1, which could be necessary to cancel a harmful source task vector or amplify a relevant adapter. The Limitations section does not mention this convex-hull restriction, and no experiment tests coefficients outside [0, 1]. Please either justify the restriction, allow a wider range, or add an experiment (for example, [−1, 2]) to show that the chosen search space is not the limiting factor for the method's performance.
  3. [Figure 3 / Sec. 4.2] The comparison with PAES and PMAES in Figure 3 is not controlled: those baselines use a different backbone (CNN+LSTM) and training procedure, so the claim that PIM 'excels under severe distribution shifts where recent leading cross-prompt methods struggle' may reflect architecture or scale differences rather than the merging method itself. The in-dataset results already show PIM lagging behind PAES and PMAES on average. Please add a same-backbone or same-scale comparison, or explicitly attribute the cross-dataset gap to the architecture and training differences and temper the claim accordingly.
  4. [Table 5 / Sec. 4.3] The ablation study shows that removing the conditional-entropy term −H(p(y|x,λ)) from Eq. (10) improves Llama-3.1-8B from 0.602 to 0.617 QWK. Thus the sharpness term, which is a core component of the PIM objective, is not beneficial for one of the two models. The discussion acknowledges a 'crossing result' but does not reconcile it with the presentation of PIM as the proposed unsupervised objective. Please provide an analysis of when the sharpness term helps or hurts, or adjust the objective and the central claim accordingly.
minor comments (5)
  1. [Eq. (2)] The summation in Eq. (2) uses λiτj; the coefficient index should be λj to match the summation variable j.
  2. [Tables 3 and 4] The footnote for the significance asterisk should describe the statistical test used (for example, paired bootstrap or paired t-test over the five seeds) and report p-values or confidence intervals; currently the reader cannot determine which differences are reliable.
  3. [Figure 4] The caption states results are from a single run on an A40 GPU; please state explicitly that this cost analysis is not averaged over seeds, unlike the main results.
  4. [Reproducibility] No code or configuration files are released; providing the merging and Bayesian optimization implementation would substantially aid reproducibility.
  5. [Limitations] The limitation that 'at least one source model provides reasonable predictions' is stated qualitatively; the P8 case with score range [0, 60], mentioned in the same paragraph, could serve as a concrete quantification of this failure mode.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PIM is an unsupervised objective evaluated against an external human-rater metric, and no equation reduces to its own inputs by construction.

full rationale

The derivation chain is self-contained. Equation (10) defines the PIM objective as f(λ) = -KL(p(y|λ)||q(y)) - H(p(y|x,λ)); q(y) is a fixed Beta-derived prior precomputed from source label statistics (Eqs. 7-9), the entropy term is computed from the merged model's predictions on unlabeled target inputs, and Bayesian optimization selects λ to maximize this fixed objective. The reported QWK is measured against human raters and does not appear anywhere in the objective, so no fitted quantity is renamed as a prediction and no equation reduces to its input by construction. The model family (linear combination of LoRA task vectors with λ∈[0,1], Eq. 2) is an explicit ansatz rather than a derived theorem, and the authors' Limitations section appropriately concedes that if no source model captures the target prompt, PIM can degrade arbitrarily. There are no load-bearing self-citations: the task-vector and information-maximization components cite prior external work (Ilharco et al. 2023; Liang et al. 2020; Krause et al. 2010) and are implemented as stated. The main caveat is not circularity but consistency of presentation: the abstract's claim that the method 'consistently outperforms training jointly on all sources' is contradicted by Table 3 for Llama-3.1-8B-it, where PIM averages 0.602 vs. joint-train 0.601 and joint-train is better on P4, P6, and P8; that is a correctness/claim-calibration issue, not a circularity issue.

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

The method relies on standard model-merging assumptions and on source-derived statistics. No new physical or conceptual entities are introduced. The key free parameters are the mixing coefficients and the fixed hyperparameters of the optimization pipeline.

free parameters (5)
  • Mixing coefficients lambda_j = Optimized via Bayesian optimization; no fixed values
    The core free parameters; search space [0,1]^M, selected by maximizing PIM on 64 target samples.
  • Bayesian optimization iterations = 40 total (10 initial random + 30 refinement)
    Hand-set budget; results may vary with more iterations.
  • Target adaptation sample size = 64
    Fixed subset of target essays used in the PIM objective; chosen by hand without explicit justification.
  • LoRA rank = r=16
    Hyperparameter for fine-tuning source models; affects capacity of task vectors.
  • EI exploration parameter xi = 0.01
    Acquisition function parameter for Bayesian optimization; set without sensitivity analysis.
assumptions (5)
  • domain assumption Linear combination of task vectors approximates joint training (Eq. 2)
    The method assumes merged LoRA adapters behave like a jointly trained model; known to be approximate (Ilharco et al., 2023).
  • domain assumption Source-derived Beta prior q(y) approximates target score distribution after scaling (Eq. 7-9)
    Assumes score distributions are comparable after affine normalization; the authors note extreme ranges can fail in Limitations.
  • ad hoc to paper Maximizing PIM (Eq. 10) improves scoring accuracy
    The objective is a heuristic combining marginal match to prior with sharpness; no proof that it tracks QWK.
  • domain assumption LLM next-token probabilities are calibrated score probabilities
    Normalized next-token probabilities over score tokens are used as p(y|x); calibration is not verified.
  • domain assumption A fixed set of 64 target essays is representative for adaptation
    Coefficients are chosen on 64 samples; small and potentially biased for prompts with wide score ranges (e.g., P8 with range 0-60).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Composable Cross-prompt Essay Scoring by Merging Models." pith.science (2026). https://pith.science/paper/LFE3QPEQ

@misc{pith2026250518548,
  author       = {Pith},
  title        = {Pith review of: Composable Cross-prompt Essay Scoring by Merging Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFE3QPEQ}},
  note         = {Machine review of arXiv:2505.18548}
}
read the original abstract

Recent advances in cross-prompt automated essay scoring (AES) typically train models jointly on all source prompts, often requiring additional access to unlabeled target prompt essays simultaneously. However, using all sources is suboptimal in our pilot study, and re-accessing source datasets during adaptation raises privacy concerns. We propose a source-free adaptation approach that selectively merges individually trained source models' parameters instead of datasets. In particular, we simulate joint training through linear combinations of task vectors -- the parameter updates from fine-tuning. To optimize the combination's coefficients, we propose Prior-encoded Information Maximization (PIM), an unsupervised objective which promotes the model's score discriminability regularized by priors pre-computed from the sources. We employ Bayesian optimization as an efficient optimizer of PIM. Experimental results with LLMs on in-dataset and cross-dataset adaptation show that our method (1) consistently outperforms training jointly on all sources, (2) maintains superior robustness compared to other merging methods, (3) excels under severe distribution shifts where recent leading cross-prompt methods struggle, all while retaining computational efficiency.

Figures

Figures reproduced from arXiv: 2505.18548 by the authors.

Figure 1
Figure 1. Agreement with human raters (QWK) on target prompt (P7 of ASAP), using BERT (Devlin et al., 2019) fine-tuned jointly on varying number of source prompt datasets. Notably, training on all sources leads to suboptimal performance. Similar trends are observed on other target prompts, see Appendix A. on new prompts (Phandi et al., 2015), accelerating efforts on cross-prompt AES with domain adap￾tation or generalization t… view at source ↗
Figure 2
Figure 2. An illustration of our method for source-free cross-prompt AES. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of PIM (phi-4-mini-it) with top [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Log-scale time (y-axis) for pre-training source [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Agreement with human raters (QWK) on target domains using BERT ( [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Comparison of PIM (llama-3.1-8b-it) with top [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 38 canonical work pages

  1. [1]

    Kazuki Adachi, Shin'ya Yamaguchi, Atsutoshi Kumagai, and Tomoki Hamagami. 2025. https://openreview.net/forum?id=SXtl7NRyE5 Test-time adaptation for regression by subspace alignment . In The Thirteenth International Conference on Learning Representations

    work page 2025

  2. [2]

    Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. 2023. https://openreview.net/forum?id=CQsmMYmlP5T Git re-basin: Merging models modulo permutation symmetries . In The Eleventh International Conference on Learning Representations

    2023

  3. [3]

    John Bridle, Anthony Heading, and David MacKay. 1991. https://proceedings.neurips.cc/paper_files/paper/1991/file/a8abb4bb284b5b27aa7cb790dc20f80b-Paper.pdf Unsupervised classifiers, mutual information and phantom targets . In Advances in Neural Information Processing Systems, volume 4. Morgan-Kaufmann

    work page 1991

  4. [4]

    Yue Cao, Hanqi Jin, Xiaojun Wan, and Zhiwei Yu. 2020. Domain-adaptive neural automated essay scoring. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pages 1011--1020

    work page 2020

  5. [5]

    Hongbo Chen and Ben He. 2013. https://aclanthology.org/D13-1180/ Automated essay scoring by maximizing human-machine agreement . In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1741--1752, Seattle, Washington, USA. Association for Computational Linguistics

    work page 2013

  6. [6]

    Yuan Chen and Xia Li. 2023. https://doi.org/10.18653/v1/2023.acl-long.83 PMAES : Prompt-mapping contrastive learning for cross-prompt automated essay scoring . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489--1503, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.acl-long.83 2023

  7. [7]

    Yuan Chen and Xia Li. 2024. https://aclanthology.org/2024.lrec-main.1118/ PLAES : Prompt-generalized and level-aware learning framework for cross-prompt automated essay scoring . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12775--12786, Torino, Italia. ...

    work page 2024

  8. [8]

    SeongYeub Chu, Jong Woo Kim, Bryan Wong, and Mun Yong Yi. 2025. https://aclanthology.org/2025.findings-naacl.322/ Rationale behind essay scores: Enhancing S - LLM `s multi-trait essay scoring with rationale generated by LLM s . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5796--5814, Albuquerque, New Mexico. Association ...

    work page 2025

Show all 71 references
  1. [9]

    M a d a lina Cozma, Andrei Butnaru, and Radu Tudor Ionescu. 2018. https://doi.org/10.18653/v1/P18-2080 Automated essay scoring with string kernels and word embeddings . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short ...

    2018 doi

  2. [10]

    Crossley, Y

    S.A. Crossley, Y. Tian, P. Baffour, A. Franklin, M. Benner, and U. Boser. 2024. https://doi.org/10.1016/j.asw.2024.100865 A large-scale corpus for assessing written argumentation: Persuade 2.0 . Assessing Writing, 61:100865

    2024

  3. [11]

    Ronan Cummins, Meng Zhang, and Ted Briscoe. 2016. https://doi.org/10.18653/v1/P16-1075 Constrained Multi-Task Learning for Automated Essay Scoring . In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics ( Volume 1: Long Papers ) , pages 789...

    2016 doi

  4. [12]

    Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. 2024. https://doi.org/10.48550/arXiv.2406.11617 Della-merging: Reducing interference in model merging through magnitude-based sampling . CoRR, abs/2406.11617

    2024 doi arXiv

  5. [13]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  6. [14]

    Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. 2022. https://doi.org/10.1109/CVPR52688.2022.00707 Source-free domain adaptation via distribution estimation . In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7202--7212

    2022

  7. [15]

    Heejin Do, Yunsu Kim, and Gary Geunbae Lee. 2023. https://doi.org/10.18653/v1/2023.findings-acl.98 Prompt- and trait relation-aware cross-prompt essay trait scoring . In Findings of the Association for Computational Linguistics: ACL 2023, pages 1538--1551, Toronto, Canada. Ass...

    2023 doi

  8. [16]

    Fei Dong and Yue Zhang. 2016. https://doi.org/10.18653/v1/D16-1115 Automatic features for essay scoring -- an empirical study . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1072--1077, Austin, Texas. Association for Computati...

    2016 doi

  9. [17]

    Fei Dong, Yue Zhang, and Jie Yang. 2017. https://doi.org/10.18653/v1/K17-1017 Attention-based recurrent convolutional neural network for automatic essay scoring . In Proceedings of the 21st Conference on Computational Natural Language Learning ( C o NLL 2017) , pages 153--162,...

    2017 doi

  10. [18]

    Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/168908dd3227b8358eababa07fcaf091-Paper.pdf Confident anchor-induced multi-source free domain adaptation . In Advances in Neural Information Processin...

    2021

  11. [19]

    Youmna Farag, Helen Yannakoudakis, and Ted Briscoe. 2018. https://doi.org/10.18653/v1/N18-1024 Neural automated essay scoring and coherence modeling for adversarially crafted input . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Co...

    2018 doi

  12. [20]

    Peter I Frazier. 2018. A tutorial on bayesian optimization. arXiv preprint arXiv:1807.02811

    2018 arXiv

  13. [21]

    Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, and Emanuele Rodol \`a . 2025. https://doi.org/10.48550/arXiv.2412.00081 Task Singular Vectors : Reducing Task Interference in Model Merging . Preprint, arXiv:2412.00081

    2025 doi arXiv

  14. [22]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  15. [23]

    Ben Hamner, Jaison Morgan, lynnvandev, Mark Shermis, and Tom Vander Ark. 2012. https://kaggle.com/competitions/asap-aes The hewlett foundation: Automated essay scoring

    2012

  16. [24]

    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

  17. [25]

    Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/1dba5eed8838571e1c80af145184e515-Paper.pdf Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data . In Adva...

    2021

  18. [26]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. https://openreview.net/forum?id=6t0Kwf8-jrj Editing models with task arithmetic . In The Eleventh International Conference on Learning Representations

    2023

  19. [27]

    Chaeyun Jang, Hyungi Lee, Jungtaek Kim, and Juho Lee. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/34d3cf97696022b179171e5abda42c0b-Paper-Conference.pdf Model fusion through bayesian optimization in language model fine-tuning . In Advances in Neural Informa...

    2024

  20. [28]

    Zhiwei Jiang, Tianyi Gao, Yafeng Yin, Meng Liu, Hua Yu, Zifeng Cheng, and Qing Gu. 2023. https://doi.org/10.18653/v1/2023.acl-long.696 Improving domain generalization for prompt-aware essay scoring via disentangled representation learning . In Proceedings of the 61st Annual Me...

    2023 doi

  21. [29]

    Cancan Jin, Ben He, Kai Hui, and Le Sun. 2018. https://doi.org/10.18653/v1/P18-1100 TDNN : A two-stage deep neural network for prompt-independent automated essay scoring . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2018 doi

  22. [30]

    Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. 2023. https://openreview.net/forum?id=FCnohuR6AnM Dataless knowledge fusion by merging weights of language models . In The Eleventh International Conference on Learning Representations

    2023

  23. [31]

    Andreas Krause, Pietro Perona, and Ryan Gomes. 2010. https://proceedings.neurips.cc/paper_files/paper/2010/file/42998cf32d552343bc8e460416382dca-Paper.pdf Discriminative clustering by regularized information maximization . In Advances in Neural Information Processing Systems, ...

    2010

  24. [32]

    Shengjie Li and Vincent Ng. 2024. https://doi.org/10.18653/v1/2024.acl-long.414 Conundrums in cross-prompt automated essay scoring: Making sense of the state of the art . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long...

    2024 doi

  25. [33]

    Jian Liang, Dapeng Hu, and Jiashi Feng. 2020. https://proceedings.mlr.press/v119/liang20a.html Do we really need to access the source data? S ource hypothesis transfer for unsupervised domain adaptation . In Proceedings of the 37th International Conference on Machine Learning,...

    2020

  26. [34]

    Deyuan Liu, Zecheng Wang, Bingning Wang, Weipeng Chen, Chunshan Li, Zhiying Tu, Dianhui Chu, Bo Li, and Dianbo Sui. 2024. Checkpoint merging via bayesian optimization in llm pretraining. arXiv preprint arXiv:2403.19390

    2024 arXiv

  27. [35]

    Ilya Loshchilov and Frank Hutter. 2019. https://openreview.net/forum?id=Bkg6RiCqY7 Decoupled weight decay regularization . In International Conference on Learning Representations

    2019

  28. [36]

    Anastassia Loukina, Nitin Madnani, and Klaus Zechner. 2019. https://doi.org/10.18653/v1/W19-4401 The many dimensions of algorithmic fairness in educational applications . In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, ...

    2019 doi

  29. [37]

    Michal Lukasik, Zhao Meng, Harikrishna Narasimhan, Yin-Wen Chang, Aditya Krishna Menon, Felix Yu, and Sanjiv Kumar. 2025. https://openreview.net/forum?id=xGs7Ch3Vyo Better autoregressive regression with LLM s via regression-aware fine-tuning . In The Thirteenth International C...

    2025

  30. [38]

    Bagdanov, and Joost van de Weijer

    Daniel Marczak, Simone Magistri, Sebastian Cygert, Bart omiej Twardowski, Andrew D. Bagdanov, and Joost van de Weijer. 2025. https://doi.org/10.48550/arXiv.2502.04959 No Task Left Behind : Isotropic Model Merging with Common and Task-Specific Subspaces . Preprint, arXiv:2502.04959

    2025 doi arXiv

  31. [39]

    Michael S Matena and Colin A Raffel. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/70c26937fbf3d4600b69a129031b66ec-Paper-Conference.pdf Merging models with fisher-weighted averaging . In Advances in Neural Information Processing Systems, volume 35, pages 17...

    2022

  32. [40]

    Microsoft, :, Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, and 57 others. 2025...

    2025 arXiv

  33. [41]

    Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. 2020. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512--523

    2020

  34. [42]

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. 2022. https://proceedings.mlr.press/v162/niu22a.html Efficient test-time model adaptation without forgetting . In Proceedings of the 39th International Conference on Machine Lear...

    2022

  35. [43]

    Fernando Nogueira. 2014--. https://github.com/bayesian-optimization/BayesianOptimization Bayesian Optimization : Open source constrained global optimization tool for Python

    2014

  36. [44]

    Chai, and Hwee Tou Ng

    Peter Phandi, Kian Ming A. Chai, and Hwee Tou Ng. 2015. https://doi.org/10.18653/v1/D15-1049 Flexible Domain Adaptation for Automated Essay Scoring Using Correlated Linear Regression . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , ...

    2015 doi

  37. [45]

    Robert Ridley, Liang He, Xin-yu Dai, Shujian Huang, and Jiajun Chen. 2021. Automated cross-prompt scoring of essay traits. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 13745--13753

    2021

  38. [46]

    Robert Ridley, Liang He, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2020. Prompt agnostic essay scorer: a domain generalization approach to cross-prompt automated essay scoring. arXiv preprint arXiv:2008.01441

    2020 arXiv

  39. [47]

    Nils-Jonathan Schaller, Yuning Ding, Andrea Horbach, Jennifer Meyer, and Thorben Jansen. 2024. https://aclanthology.org/2024.bea-1.18/ Fairness in automated essay scoring: A comparative analysis of algorithms on G erman learner essays from secondary education . In Proceedings ...

    2024

  40. [48]

    Lin Sun, Guangxiang Zhao, Xiaoqi Jian, Yuhan Wu, Weihong Lin, Yongfu Zhu, Change Jia, Linglin Zhang, Jinzhu Wu, Junfeng Ran, Sai-er Hu, Zihan Jiang, Junting Zhou, Wenrui Liu, Bin Cui, Tong Yang, and Xiangzheng Zhang. 2025. https://doi.org/10.48550/arXiv.2503.04872 TinyR1-32B-P...

    2025 doi arXiv

  41. [49]

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. 2020. https://proceedings.mlr.press/v119/sun20b.html Test-time training with self-supervision for generalization under distribution shifts . In Proceedings of the 37th International Conference on M...

    2020

  42. [50]

    Kaveh Taghipour and Hwee Tou Ng. 2016. https://doi.org/10.18653/v1/D16-1193 A neural approach to automated essay scoring . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1882--1891, Austin, Texas. Association for Computational ...

    2016 doi

  43. [51]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, and 75 others. 2025. https://doi...

    2025 doi arXiv

  44. [52]

    Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. 2022. https://doi.org/10.1109/TCSVT.2021.3111034 Vdm-da: Virtual domain modeling for source data-free domain adaptation . IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3749--3760

    2022

  45. [53]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  46. [54]

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. 2021. https://openreview.net/forum?id=uXl3bZLkr3c Tent: Fully test-time adaptation by entropy minimization . In International Conference on Learning Representations

    2021

  47. [55]

    Jiong Wang and Jie Liu. 2025. https://aclanthology.org/2025.coling-main.81/ T - MES : Trait-aware mix-of-experts representation learning for multi-trait essay scoring . In Proceedings of the 31st International Conference on Computational Linguistics, pages 1224--1236, Abu Dhab...

    2025

  48. [56]

    Christopher KI Williams and Carl Edward Rasmussen. 2006. Gaussian processes for machine learning, volume 2. MIT press Cambridge, MA

    2006

  49. [57]

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and 1 others. 2022. Model soups: averaging weights of multiple fine-tuned models improves accuracy without in...

    2022

  50. [58]

    Jing Xu, Jiazheng Li, and Jingzhao Zhang. 2025. https://doi.org/10.48550/arXiv.2502.12706 Scalable Model Merging with Progressive Layer-wise Distillation . Preprint, arXiv:2502.12706

    2025 doi arXiv

  51. [59]

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/1644c9af28ab7916874f6fd6228a9bcf-Paper-Conference.pdf Ties-merging: Resolving interference when merging models . In Advances in Neural I...

    2023

  52. [60]

    Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. 2024 a . Representation surgery for multi-task model merging. In International Conference on Machine Learning, pages 56332--56356. PMLR

    2024

  53. [61]

    Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xingwei Wang, Xiaocun Cao, Jie Zhang, and Dacheng Tao. 2024 b . https://doi.org/10.48550/arXiv.2410.14389 SurgeryV2 : Bridging the Gap Between Model Merging and Multi-Task Learning with Deep Representation Surgery . Preprint, arX...

    2024 doi arXiv

  54. [62]

    Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. 2024 c . https://openreview.net/forum?id=nZP6NgD3QY Adamerging: Adaptive model merging for multi-task learning . In The Twelfth International Conference on Learning Representations

    2024

  55. [63]

    Shiqi Yang, yaxing wang, kai wang, Shangling Jui, and Joost van de Weijer. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/26300457961c3e056ea61c9d3ebec2a4-Paper-Conference.pdf Attracting and dispersing: A simple approach for source-free domain adaptation . In...

    2022

  56. [64]

    Maxwell Yin, Boyu Wang, and Charles Ling. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.44 Source-free unsupervised domain adaptation for question answering via prompt-assisted self-learning . In Findings of the Association for Computational Linguistics: NAACL 2024, pa...

    2024 doi

  57. [65]

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In International Conference on Machine Learning, pages 57755--57775. PMLR

    2024

  58. [66]

    Torsten Zesch, Michael Wojatzki, and Dirk Scholten-Akoun. 2015. https://doi.org/10.3115/v1/W15-0626 Task-independent features for automated essay grading . In Proceedings of the Tenth Workshop on Innovative Use of NLP for Building Educational Applications , pages 224--232, Den...

    2015 doi

  59. [67]

    Bo Zhang, Xiaoming Zhang, Yun Liu, Lei Cheng, and Zhoujun Li. 2021. https://doi.org/10.18653/v1/2021.acl-long.421 Matching distributions between model and data: Cross-domain knowledge distillation for unsupervised domain adaptation . In Proceedings of the 59th Annual Meeting o...

    2021 doi

  60. [68]

    Aurick Zhou and Sergey Levine. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/07ac7cd13fd0eb1654ccdbd222b81437-Paper.pdf Bayesian adaptation for covariate shift . In Advances in Neural Information Processing Systems, volume 34, pages 914--927. Curran Associates, Inc

    2021

  61. [69]

    Zhanpeng Zhou, Yongyi Yang, Xiaojiang Yang, Junchi Yan, and Wei Hu. 2023. Going beyond linear mode connectivity: The layerwise linear feature connectivity. Advances in neural information processing systems, 36:60853--60877

    2023

  62. [70]

    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...

  63. [71]

    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.