Pith. sign in

REVIEW 4 major objections 6 minor 26 references

REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing

T0 review · 4 major / 6 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 REACT claims that LLM knowledge-editing overfitting can be eliminated by extracting a compact 'belief shift' vector from latent representations and applying it only when a pre-trained classifier decides the context is in scope.

desk verdict The gating objective is self-defeating as written and the headline numbers are overstated, but the pipeline has some value. read the letter →

arxiv 2505.18933 v1 pith:WNAEQBS4 submitted 2025-05-25 cs.AI cs.CL

Haitian Zhong , Yuhuan Liu , Ziyang Xu , Guofan Liu , Qiang Liu , Shu Wu , Zhe Zhao , Liang Wang
show 1 more author
Tieniu Tan
This is my paper
classification cs.AIcs.CL
keywords knowledgeeditingoverfittingrepresentationengineeringhidden-stateperturbationbeliefshiftvectorcontrollableEVOKECOUNTERFACT
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

REACT argues that overfitting in LLM knowledge editing is a representation-level error: after an edit, the new object's probability is high even in contexts that do not call for it. The proposed cure is to extract, for each edit, a compact directional vector that captures the model's latent 'belief shift', then perturb hidden states with that vector only when a pre-trained classifier decides the context is in scope. The paper reports that this two-phase procedure produces markedly lower Direct Probability scores on the EVOKE overfitting benchmark across both tested models, while keeping reliability, locality, generality, and portability competitive on COUNTERFACT and MQuAKE. If the reported behavior holds, knowledge editing could become controllable enough for deployment without the usual leakage of edited facts into unrelated queries.

What carries the argument

The load-bearing object is the per-layer 'belief shift' vector $r^{(l)} = W\,[h_+^{(l)}; h_-^{(l)}] + b$, distilled from $N=512$ stimulus pairs by PCA and a learned linear transformation. It is complemented by a pre-trained binary gate built from cosine similarity between prompted and unprompted hidden states; the gate decides whether a hidden state deserves perturbation. The mechanism is then the update rule $h' = h + \alpha\,\mathrm{sign}(h^\top r^{(l)})\,r^{(l)}$ when the gate fires, and $h' = h$ otherwise. The same vector and gate are used across all transformer layers, with one global learnable magnitude $\alpha$, so the method's controllability rests on the vector's direction and the gate's decisions rather than on per-layer parameter edits.

What would settle it

Take EVOKE or a held-out set of prompts, have human annotators label each prompt as in-scope or out-of-scope for a given edit, and compare those labels with the pre-trained gate's decisions. If gate accuracy is near chance, or if the gate opens on out-of-scope prompts about the edited subject, the low Direct Probability result would not be attributable to the claimed gating mechanism.

Watch

Extended reading notes

Core claim

The central discovery REACT tries to establish is that a fact edit in an LLM can be represented as a single low-dimensional direction in hidden-state space, and that moving activations along that direction with a fixed magnitude is enough to install the new fact without letting it leak. The paper constructs the direction by feeding 512 positive/negative stimulus pairs through the model, projecting the layer-wise representation differences with PCA, and training a small linear map to output a 'belief shift' vector per layer. Editing then adds a signed scalar multiple of that vector to every transformer decoder hidden state, but only when a separately trained classifier scores the prompt as related to the edit. The reported results on EVOKE are the load-bearing evidence: lower Direct Probability than all baselines across prefix distraction, multi-hop reasoning, subject specificity, and relation specificity, together with high Editing Overfit Score and Answer Modify Score, which the paper reads as proof that the edit target is no longer over-recalled.

Load-bearing premise

Everything rests on the trained classifier: it is trained only on COUNTERFACT prompted and unprompted template pairs, yet it must decide in new datasets like MQuAKE and EVOKE whether an arbitrary context is in scope for the edit; if the gate misfires on unseen phrasing, REACT either fails to apply the edit or re-introduces overfitting, and its central advantage disappears.

Editorial extensions

If this is right

  • If REACT works as reported, editing an LLM becomes a per-query controlled operation: the new fact can be recalled under its own prompt while neighboring prompts keep their original answer distributions.
  • The low Direct Probability reported on EVOKE means the four overfitting failure modes—multi-hop reasoning, prefix distraction, subject specificity, and relation specificity—would no longer be dominated by the edit target.
  • Because the same COUNTERFACT-trained classifier and edit vector transfer to MQuAKE and EVOKE, a single trained REACT instance could serve edits across benchmarks without retraining per dataset.
  • The method offers an alternative to weight-modifying editors: since only hidden states are perturbed, the original weights remain untouched and reversible per query.

Reading between the lines

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

  • The paper does not test borderline prompts near the classifier's 0.5 threshold; an implicit prediction is that REACT's behavior is bimodal, with nearby paraphrases abruptly switching between 'edited' and 'unedited' responses.
  • The global scalar $\alpha$ ignores layer and instance variation; a natural extension the paper leaves open is per-layer magnitudes or threshold calibration, which could fine-tune the trade-off between Direct Probability and Correct Answer Probability.
  • Because the edit is applied at forward-pass time rather than written into weights, serving systems would have to recompute the perturbation for every generated token or cache edited hidden states; the paper does not discuss this deployment cost.
  • The same belief-shift vector could be sign-flipped to perform unlearning or fact negation, since the extraction procedure is symmetric in positive and negative stimuli; this is an extension, not a claim in the paper.
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 / 6 minor

Summary. The paper proposes REACT, a two-phase knowledge editing method for large language models. In Phase I, it extracts a per-instance 'belief shift' vector from positive/negative stimulus pairs using PCA and a learnable linear transformation. In Phase II, it perturbs hidden states by a scalar magnitude times this vector, gated by a pretrained classifier that is supposed to permit the edit only when the input is contextually relevant. The method is evaluated on COUNTERFACT, MQuAKE, and EVOKE using Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, and the paper claims balanced editing performance plus significantly reduced overfitting relative to FT, MEND, MEMIT, MELO, and GRACE.

Significance. If the claims were supported, REACT would be a useful contribution to knowledge editing, since it directly addresses overfitting through explicit gating of hidden-state perturbations rather than weight modifications. The paper includes ablations, detailed hyperparameters, and a clear exposition of the intended pipeline. However, the central gating mechanism is internally inconsistent as written, and the two strongest headline empirical claims are contradicted by the paper's own tables. The absence of error bars further weakens confidence in the small differences that support the overfitting-reduction claims. The core idea of classifier-gated perturbation is interesting and potentially salvageable, but the current manuscript does not establish it.

major comments (4)
  1. [§3.3, Eqs. (12)–(14)] The classifier training objective is inconsistent with its intended gating behavior. For a fixed nonzero Δh^(l), the combined loss c_edit,cls ||γ^(l) Δh^(l)||² + c_loc,cls ||(1−γ^(l)) Δh^(l)||² is minimized at γ^(l) = c_loc,cls / (c_edit,cls + c_loc,cls). With the settings in Appendix D.1.1 (c_edit,cls=1, c_loc,cls=0.1), this is γ* ≈ 0.09, far below the 0.5 threshold in Eq. (11). The stated loss therefore drives γ to a constant low value, making Φ output 0 almost everywhere and preventing the perturbation in Eq. (2) from ever being applied. The text's explanation that L_edit,cls 'encourages large Δh when γ is high' is the opposite of what the equation does, since the squared term penalizes large γΔh. This is a load-bearing flaw: the central controllable-perturbation mechanism is not realized by the equations as written.
  2. [§5.1, Finding 1 and Table 6] The claim that REACT 'outperforms the second-best baseline by at least 20 percentage points in terms of average score' is directly contradicted by Table 6. On Llama3.1-8B the average-score gap to the second-best method (MEND) is 81.86 − 75.17 = 6.69 points; on Qwen2.5-7B the gap to the second-best method (MEMIT) is 81.52 − 63.58 = 17.94 points. Neither reaches 20 points. The statement overstates the results and must be corrected to match the table.
  3. [§5.2, Finding 1 and Table 7] The claim that REACT yields 'markedly lower Direct Probability (DP) scores across all evaluation settings' is contradicted by Table 7 on Llama3.1. In Prefix Distraction, REACT's DP is 5.44 while MELO achieves 2.57; in Multi-hop Reasoning, REACT's DP is 0.96 while MELO achieves 0.58. Thus REACT does not have lower DP than all baselines in all settings. The empirical basis for the central overfitting-reduction claim is therefore not supported by the reported data.
  4. [Tables 6 and 7] The experimental results are reported as single point estimates without error bars, confidence intervals, or statistical significance tests. Many of the differences that underpin the paper's claims—for example, the DP gaps of a few points on some EVOKE tasks—are small enough that they could be within run-to-run variance. Given that the headline findings depend on comparably small margins, the absence of any uncertainty quantification is a major limitation.
minor comments (6)
  1. [Abstract] There is a typo: 'learnbale' should be 'learnable', and 'REACTsignificantly' should have a space.
  2. [§3.3] The paragraph after Eq. (10) contains the typo 'quries' which should be 'queries'.
  3. [§4.3.3] The phrase 'To evulate the impact' contains a typo: 'evulate' should be 'evaluate'.
  4. [Figure 2 caption] The caption begins with 'We First construct' where 'First' is capitalized mid-sentence; it should be lowercase.
  5. [Table 6 caption] The caption says 'A radar chart for the table is created at 3' but should refer to 'Figure 3' explicitly for clarity.
  6. [§5.1, Finding 1] The phrase 'by at least 20 percentage points' is not merely an overstatement but is inconsistent with the data in Table 6; it should be revised to reflect the actual margins.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the editing operator is trained on COUNTERFACT-train and evaluated on held-out benchmarks; the self-cited EVOKE benchmark is public and its numbers are reportable.

full rationale

No significant circularity. The belief-shift vector r^(l) is built from per-edit positive/negative stimuli, and the perturbation parameters W, b, alpha, and the classifier Phi are trained on COUNTERFACT-train and then evaluated on held-out COUNTERFACT-edit, MQuAKE-v2, and EVOKE; no evaluation metric is used as a training target and then reported as a prediction. The use of the edit target o* inside the stimuli is the declared input of the editing operator K(f,e), not a fitted prediction. The classifier gate is trained on prompted/unprompted template distinctions from the training split and applied to new contexts, which is an empirical generalization claim rather than a definitional equivalence. The EVOKE benchmark is cited to prior work sharing two co-authors (Qiang Liu, Shu Wu), but the benchmark is a public dataset and the paper reports its own numbers in Table 7, so this self-citation is not load-bearing in a circular sense. The manuscript does contain serious non-circular defects: minimizing Eqs. 12-14 with the stated weights drives gamma to a constant below the 0.5 threshold, and the Section 5.1/5.2 claims of a 20-point average gap and lower DP across all settings are not supported by Tables 6-7; these are consistency and empirical issues, not circular reductions.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The method relies on several asserted modeling choices: stimulus templates elicit edit-relevant representations, PCA on 512 pairs isolates the true belief shift, a single global scalar and sign heuristic safely control perturbation, and the classifier trained on COUNTERFACT templates transfers to other benchmarks. Each of these is a domain assumption supported only by the paper's own benchmark numbers, which are weakened by missing error bars and some contradicted headline claims.

free parameters (6)
  • N (number of stimulus pairs per layer) = 512
    Selected by ablation on COUNTERFACT (Appendix C.1); N=256 degrades metrics and N=1024 causes OOM.
  • W, b (linear transformation in Eq. 1) = learned on COUNTERFACT-train (values not reported)
    Maps concatenated PCA components to the belief-shift vector; central learned component.
  • alpha (perturbation magnitude scalar) = learned, value not reported
    Global scalar controlling perturbation size in Eq. (2); no per-layer or per-edit variation.
  • Classifier parameters W_Q^(l), W_U^(l) per layer = learned on COUNTERFACT-train
    Project prompted and unprompted hidden states before cosine similarity in Eqs. (8)-(9).
  • Classifier threshold = 0.5
    Fixed cosine-similarity cutoff in Eq. (11); no sensitivity analysis reported.
  • Loss weights c_edit, c_loc, c_edit_cls, c_loc_cls = 1, 0.1, 1, 0.1
    Hand-set in Appendix D.1.1; no grid search reported.
assumptions (6)
  • domain assumption A single linear transformation W,b of PCA-reduced positive/negative stimulus representations yields a valid "belief shift" vector r(l) for each edit.
    Section 3.1 Eq. (1); this is the core representation extraction, asserted and trained, not derived.
  • domain assumption The sign of h^T r(l) determines the correct direction of the edit perturbation for all relevant hidden states.
    Section 3.2 Eq. (2); no analysis of sign accuracy is given.
  • domain assumption The cosine-similarity classifier with threshold 0.5, trained on COUNTERFACT prompted/unprompted inputs, reliably detects edit-relevant contexts on MQuAKE and EVOKE.
    Section 3.3 Eqs. (10)-(11); generalization is claimed but not proven.
  • domain assumption N=512 stimulus pairs provide sufficient representational richness, and the PCA direction from these pairs is the dominant factual shift.
    Section 3.1 and Appendix C.1; N was chosen by ablation on the same benchmark family.
  • domain assumption Editing via additive perturbation of hidden states with a shared alpha does not degrade model quality outside the edit scope.
    Section 3.2; relies on locality loss and gate, no independent verification.
  • standard math Principal Component Analysis via SVD is differentiable and thus trainable.
    Appendix C.2; a standard mathematical fact used to justify joint optimization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing." pith.science (2026). https://pith.science/paper/WNAEQBS4

@misc{pith2026250518933,
  author       = {Pith},
  title        = {Pith review of: REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WNAEQBS4}},
  note         = {Machine review of arXiv:2505.18933}
}
read the original abstract

Large language model editing methods frequently suffer from overfitting, wherein factual updates can propagate beyond their intended scope, overemphasizing the edited target even when it's contextually inappropriate. To address this challenge, we introduce REACT (Representation Extraction And Controllable Tuning), a unified two-phase framework designed for precise and controllable knowledge editing. In the initial phase, we utilize tailored stimuli to extract latent factual representations and apply Principal Component Analysis with a simple learnbale linear transformation to compute a directional "belief shift" vector for each instance. In the second phase, we apply controllable perturbations to hidden states using the obtained vector with a magnitude scalar, gated by a pre-trained classifier that permits edits only when contextually necessary. Relevant experiments on EVOKE benchmarks demonstrate that REACT significantly reduces overfitting across nearly all evaluation metrics, and experiments on COUNTERFACT and MQuAKE shows that our method preserves balanced basic editing performance (reliability, locality, and generality) under diverse editing scenarios.

Figures

Figures reproduced from arXiv: 2505.18933 by the authors.

Figure 1
Figure 1. Illustration of overfitting in LLM editing. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of our REACT pipeline for controllable knowledge editing. We First construct stimuli prompts and feed them into the LLM to extract layer-wise representations, which are then processed via PCA and an MLP to isolate the key “belief shift” vector. Thereafter, we apply a controllable perturbation (using learned scalar factors) to the model’s hidden states. The pre-trained classifier manages when the edits sh… view at source ↗
Figure 3
Figure 3. Editing results on COUNTERFACT and MQuAKE-CF-v2 in radar chart. Detailed results could be found [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Editing results on EVOKE in radar chart. Values prefixed with “100-” denote the difference between the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 5 canonical work pages

  1. [1]

    Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. 2023. Can we edit multimodal large language models? arXiv preprint arXiv:2310.08475

    arXiv 2023

  2. [2]

    Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2024. Evaluating the ripple effects of knowledge editing in language models. Transactions of the Association for Computational Linguistics, 12:283--298

    2024

  3. [3]

    Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.522 Editing factual knowledge in language models . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491--6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics

    doi:10.18653/v1/2021.emnlp-main.522 2021

  4. [4]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2022. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2203.14465

    arXiv 2022

  5. [5]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and et al. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

    arXiv 2024

  6. [6]

    Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with grace: Lifelong model editing with discrete key-value adaptors. In Advances in Neural Information Processing Systems

    2023

  7. [7]

    Evan Hernandez, Belinda Z Li, and Jacob Andreas. 2023. Inspecting and editing knowledge representations in language models. arXiv preprint arXiv:2306.04542

    arXiv 2023

  8. [8]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022 a . Locating and editing factual associations in GPT . Advances in Neural Information Processing Systems, 36

    work page 2022

Show all 26 references
  1. [9]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023. Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR)

    2023

  2. [10]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022 b . Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229

    2022 arXiv

  3. [11]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2021. Fast model editing at scale. arXiv preprint arXiv:2110.11309

    2021 arXiv

  4. [12]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022 a . https://openreview.net/pdf?id=0DcZxeWfOPt Fast model editing at scale . In International Conference on Learning Representations

    2022

  5. [13]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022 b . https://arxiv.org/pdf/2206.06520.pdf Memory-based model editing at scale . In International Conference on Machine Learning

    2022 arXiv

  6. [14]

    Alex Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248

    2023 arXiv

  7. [15]

    Peng Wang, Ningyu Zhang, Xin Xie, Yunzhi Yao, Bozhong Tian, Mengru Wang, Zekun Xi, Siyuan Cheng, Kangwei Liu, Guozhou Zheng, et al. 2023. Easyedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269

    2023 arXiv

  8. [16]

    Ziyang Xu, Haitian Zhong, Bingrui He, Xueying Wang, and Tianchi Lu. 2024. Ptransips: Identification of phosphorylation sites enhanced by protein plm embeddings. IEEE Journal of Biomedical and Health Informatics

    2024

  9. [17]

    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, Le...

    2025 arXiv

  10. [18]

    Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172

    2023 arXiv

  11. [19]

    Mengqi Zhang, Xiaotian Ye, Qiang Liu, Pengjie Ren, Shu Wu, and Zhumin Chen. 2024 a . https://arxiv.org/abs/2410.07819 Uncovering overfitting in large language model editing . Preprint, arXiv:2410.07819

    2024 arXiv

  12. [20]

    Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, et al. 2024 b . A comprehensive study of knowledge editing for large language models. arXiv preprint arXiv:2401.01286

    2024 arXiv

  13. [21]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong W...

    2023 arXiv

  14. [22]

    Zexuan Zhong, Zhengxuan Wu, Christopher Manning, Christopher Potts, and Danqi Chen. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.971 MQ u AKE : Assessing knowledge editing in language models via multi-hop questions . In Proceedings of the 2023 Conference on Empirical Meth...

    2023 doi

  15. [23]

    Chengrun Zhu, Hieu Pham, Zihang Dai, Chris Cundy, Sean Welleck, and Kyunghyun Cho. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363

    2020 arXiv

  16. [24]

    Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, Zico Kolter, and Dan Hendrycks

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, Zico Ko...

    2023 arXiv

  17. [25]

    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...

  18. [26]

    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.