Pith. sign in

REVIEW 4 major objections 5 minor 24 references

CTR-Guided Generative Query Suggestion in Conversational Search

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a CTR-guided pipeline—multi-source click modeling, CTR-weighted DPO, and iterative calibration—raises click-through rate, relevance, and diversity of conversational query suggestions at the same time.

desk verdict A coherent industrial pipeline whose headline CTR gains are currently unverifiable because the CTR metric is undefined and may be the authors' own model. read the letter →

arxiv 2507.04072 v1 pith:LIX4NMID submitted 2025-07-05 cs.IR

Erxue Min , Hsiu-Yuan Huang , Xihong Yang , Min Yang , Xin Jia , Yunfang Wu , Hengyi Cai , Junfeng Wang
show 2 more authors
Shuaiqiang Wang Dawei Yin
This is my paper · ORCID
classification cs.IR
keywords conversationalsearchquerysuggestionclick-throughratedirectpreferenceoptimizationgenerativeLLMdiversity-awarealignmentiterativecalibration
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 tries to establish that noisy, biased click logs from a conversational search assistant can be turned into a reliable training signal for generating follow-up query suggestions. It proposes a three-part pipeline: a multi-source click-through-rate (CTR) predictor that scores candidate suggestions, a CTR-weighted Direct Preference Optimization step that also rewards semantic diversity, and an iterative loop that recalibrates the predictor as the generator changes. On two real-world assistant tasks it reports large relative CTR gains over supervised fine-tuning—70.36% on Task 1 and 30.72% on Task 2—while relevance and diversity also rise. The significance, if the results hold, is that offline click data alone can keep improving suggestion quality across repeated rounds without collecting costly new online feedback.

What carries the argument

The load-bearing object is the multi-source contextual CTR predictor: a shared BERT encoder produces representations of the target query and each context source (current query, assistant response, conversation history, user profile, co-occurring queries), single-head cross-attention combines each context with the query, a positional embedding encodes the suggestion's rank, and an MLP outputs a predicted click probability. That predictor supplies the reward, the preference-pair filter, the sample weight, and the reweighted training loss. Around it, GQS wraps Direct Preference Optimization (DPO)—a loss that widens the log-probability gap between chosen and rejected responses—with a CTR-based sample weight and an auxiliary diversity DPO term, plus an iterative calibration loop that clips the likelihood ratio between the initial and current generator to reweight the CTR model's training data.

What would settle it

Serve the final GQS suggestions and the SFT baseline side by side to real users in an online A/B test and compare actual click-through rates on the suggested queries. If the reported relative gains of +70.36% (Task 1) and +30.72% (Task 2) do not appear in the click logs, the central claim fails; a cheaper offline check is to recompute Table 1 using a held-out click model or human click labels never used in training.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that preference alignment driven by predicted click-through rates—rather than raw clicks or direct CTR maximization—is the key to better query suggestions. The authors construct preference pairs from a CTR model's scores: the chosen response is the highest-scoring candidate that passes a diversity threshold, and the rejected response is drawn from candidates below the lower tail of the score distribution. They then optimize a weighted DPO objective plus a diversity-preference objective, and iteratively retrain the CTR model with importance-sampled weights to match the drifting generator. Table 1 shows the complete GQS pipeline outperforming SFT, KTO, SimPO, and DPO in CTR, relevance, and diversity on both tasks; ablation and training-round curves support the contribution of each component.

Load-bearing premise

The central claim collapses if the 'CTR' column in Table 1 is not real user click-through rate but the output of the paper's own CTR predictor, because that same predictor supplies the training rewards; the paper never defines the metric in Section 4.2.

Editorial extensions

If this is right

  • If the reported gains are real, query-suggestion systems can be aligned from sparse offline click logs, and each round of deployed suggestions can be reused for further training without waiting for fresh online labels.
  • The diversity-aware DPO component prevents CTR alignment from collapsing suggestion variety; removing it lowers both diversity and CTR in the ablations.
  • The iterative calibration step matters: a fixed CTR model gives almost no improvement across rounds, while the calibrated model adds 7.45 points in Task 1 CTR.
  • Multi-source cross-attention and position embeddings are not optional; removing either degrades AUC and logloss of the CTR predictor.

Reading between the lines

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

  • The paper does not define the CTR metric in the experimental section, and the same predictor supplies the training rewards; if the Table 1 CTR column is that predictor's output rather than real user clicks, the headline gains are partly self-consistency and need independent verification.
  • The same pipeline could transfer to other list-generation tasks with click feedback—recommendation explanations, related searches, product suggestions—where multi-source cross-attention context encoding and list-level preference scoring would carry over.
  • The diversity judge is trained on a limited annotation set and the relevance/diversity metrics are scored by an automated judge; testing against human raters on unseen topics would clarify whether the reported diversity gains generalize.
  • Iterative rounds saturate around round three to five and can slightly decline; a cheaper variant that stops after a few calibrated rounds may capture most of the benefit.
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 GQS, a generative query suggestion framework for conversational AI assistants. It combines (1) a multi-source BERT/cross-attention CTR predictor, (2) CTR-weighted DPO with a diversity-aware auxiliary DPO loss, and (3) iterative rounds in which the CTR predictor is recalibrated by importance weighting. Experiments on two Baidu assistant tasks report large relative CTR improvements (Task 1 +70.36%, Task 2 +30.72%) with gains in relevance and diversity over SFT/DPO baselines. The paper also claims large-scale A/B online validation in the Introduction, but no online results are reported.

Significance. If the reported gains are genuine measures of user behavior, the framework would be a valuable industrial demonstration of using noisy click signals to align generative query suggestion models, with a useful treatment of diversity collapse. The proposed architecture is plausible, the ablation of the CTR model components is informative, and the sensitivity analyses of λ and sample size are a strength. However, the paper currently does not define the CTR evaluation metric, does not report the claimed online A/B test, and relies on an unvalidated LLM judge for relevance and diversity, so the central empirical claim cannot be assessed as stated. The contribution is therefore conditional on substantially stronger evaluation evidence.

major comments (4)
  1. [Section 4.2, Table 1] The 'CTR' metric in Table 1 is never defined. Training uses the paper's own CTR predictor for response scoring (Eq. (2): r(Y_i)=Σ_j \hat p(q_j)) and for preference-pair construction (§3.2), and §3.3 retrains that same predictor. If the Table 1 CTR values are also produced by this predictor (e.g., mean predicted p-hat on the test set), then the +70.36% and +30.72% improvements would largely measure the policy's agreement with the reward model rather than real user engagement, a form of self-consistency. Please state explicitly how the reported CTR is computed: is it an online click-through rate from a logged deployment, or the predictor's output? If the latter, provide evidence that the predictor is calibrated against held-out click logs (e.g., AUC/logloss on a real-click test set, calibration plots) and ideally report the actual observed CTR from a (even small) online or replay experiment.
  2. [Section 1, contributions bullet] The Introduction states 'We validate our method through large-scale A/B online experiments, showing significant improvements in user engagement and suggestion quality,' but no A/B experiment setup, results, or statistical tests appear anywhere in the paper. Table 1 is offline; Figures 3–5 are offline ablations. This claim is currently unverifiable. Either add the online A/B results (with durations, sample sizes, effect sizes, and confidence intervals) or remove the claim and explicitly state that all experiments are offline.
  3. [Appendix B / Section 4.2] Relevance and diversity are scored by GPT-4o with no human-validation or agreement study. The paper's own Limitations section concedes that the diversity estimation module is trained on 'a limited number of annotated samples,' which makes the diversity scores hard to interpret. Please add a human-baseline/agreement analysis (e.g., Cohen's kappa or correlation on a subsample) or report the judge's accuracy against the existing human annotations. Without this, the diversity gains (e.g., Task 1 Div. 86.04 vs. 85.63) are not distinguishable from judge noise.
  4. [Section 4.2, Table 1 and Figures 3–5] No variance estimates or significance tests are provided for any of the reported metrics. With 10,000 test queries (Section 4.1), it is feasible to report confidence intervals or pairwise significance tests (e.g., bootstrap). This is particularly important for the headline CTR improvements: a relative +70% could be driven by a small subset of queries, and the diversity numbers in Table 1 differ by less than one point from SFT in Task 1 (85.63 vs. 86.04).
minor comments (5)
  1. [Section 3.2, Step 1] The threshold δ is used both as a minimum diversity filter for the preferred response and as the difference threshold for constructing diversity pairs in Step 3; these are different quantities and should be given distinct symbols.
  2. [Section 3, first paragraph] The text contains typos: 'conversational AI assistent search' should be 'assistant', and the stray 'dlg' in Figure 2 should be removed.
  3. [Table 1] In the Task 1 KTOclk row, the entry '79.4586.81' appears to be a formatting error; it should likely be '79.45, 86.81' with a separator.
  4. [Appendix D] The first sentence says 'three query suggestion tasks' but only two tasks (T1 and T2) are described; correct the count.
  5. [Figure 7] The axis labels '101', '102', etc. should be typeset as 10^1, 10^2, etc., to avoid confusion with the numbers 101 and 102.

Circularity Check

1 steps flagged · score 7.0 of 10

Reported CTR improvements reduce to the paper's own CTR predictor: the training reward is the predicted CTR and the 'CTR' metric is never defined separately, so the headline gains are self-consistency, not demonstrated real engagement.

  1. fitted input called prediction [Section 3.2 (Eq. 2), Section 3.3 (Eqs. 7-9), Section 4.1/4.2 (CTR metric)]
    "Using the CTR model described in Section 2, we compute the total predicted click likelihood of each response: r(Yi) = NX j=1 ˆp(qj), (2) ... we utilize three metrics: Click-Through Rate (CTR), Relevance (Rel.) and Diversity (Div.)."

    Eq. (2) sets the response reward r(Y_i) to the sum of predicted click probabilities p̂(q_j) from the paper's own CTR model. The DPO objective (Eqs. 3-5) is optimized to prefer responses with larger r(Y), and the iterative procedure (Eqs. 7-9) retrains the same CTR model and repeats. Section 4.1 lists 'Click-Through Rate (CTR)' as an evaluation metric but never defines how it is computed; no other CTR measure is given in the paper. If the CTR column in Table 1 is produced by this predictor, DPO directly maximizes the reported metric, making the +70.36% Task-1 and +30.72% Task-2 gains forced by construction rather than evidence of real user engagement.

full rationale

The paper's central claimed improvement is in CTR, but the only CTR-related quantity defined anywhere is the predicted click probability from the multi-source CTR model (Section 3.1, Eq. 2). This same predicted CTR is used as the reward for constructing DPO preference pairs (Section 3.2) and for iterative retraining (Section 3.3). The evaluation section lists CTR as a metric but gives no formula, no data source, and no procedure, leaving the reader unable to distinguish the reported 'CTR Impr.' from the model's own predictions. Since the training objective is to increase the predicted CTR of generated responses, any evaluation using that same predictor is circular: the improvements are a mathematical consequence of the optimization, not an independent outcome. The relevance and diversity metrics, scored by GPT-4o (Appendix B), are external and not circular. No problematic self-citation was found. The claimed A/B online experiments are absent from the paper, so the real-user-engagement interpretation is unsupported. Overall, the central CTR claim reduces by construction to the fitted CTR model, warranting a high circularity score.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central results rest on several hand-set hyperparameters whose values are not reported, plus domain assumptions about click signal validity, diversity-scorer generalization, and GPT-4o judge quality. No new physical or ontological entities are introduced; the auxiliary models are learned predictors over existing features.

free parameters (7)
  • lambda (diversity loss coefficient) = 0.1
    Chosen by sensitivity analysis in Appendix C; balances CTR alignment and diversity in Eq. 6.
  • beta (DPO temperature) = not reported
    Controls the DPO softness in Eq. 3; a hand-set hyperparameter whose value is omitted.
  • gamma (CTR-gap scaling factor) = not reported
    Scales the sample weight alpha in Eq. 4; value omitted.
  • delta (diversity threshold) = not reported
    Used in Section 3.2 to filter low-diversity responses and to construct diversity preference pairs; value omitted.
  • epsilon (importance-weight clipping bound) = not reported
    Clips the importance weight w(t) in Eq. 7, controlling how much distribution-shift correction is allowed; value omitted.
  • rejection threshold (mu - 2 sigma) = 2 sigma
    Lower-tail cutoff for selecting rejected responses in Section 3.2; a design choice, not derived.
  • number of iterative DPO rounds = approximately 5
    Section 4.4 shows saturation after round 3; number of rounds is a design choice affecting the final result.
assumptions (5)
  • domain assumption DPO remains a valid preference optimization objective when the 'response' is a list of N generated queries rather than a single sequence.
    Section 3.2 applies the standard DPO loss (Eq. 3) to list-structured responses without a derivation that the list likelihood ratio behaves like a sequence likelihood.
  • domain assumption Click behavior is a valid proxy for user preference after CTR-model correction.
    The entire framework in Sections 1 and 3.1 relies on clicks as preference signals, while acknowledging they are noisy and biased; the correction is assumed sufficient.
  • domain assumption The BERT-based diversity scorer, trained on limited human annotations, generalizes to unseen topics and domains.
    Section 3.2 and the Limitations section state the scorer uses a limited number of annotated samples; generalization is assumed without evidence.
  • domain assumption GPT-4o relevance and diversity judgments correlate with human judgments.
    Appendix B uses GPT-4o as an automatic judge for both metrics and reports no human agreement or calibration study.
  • domain assumption The clipped importance-weight ratio in Eq. 7 corrects the distribution shift between the initial and current generation policies.
    Eqs. 7-8 apply importance weighting to retrain the CTR model, but the clipping bound and the use of list-level weights are not validated against an unclipped estimator or a ground-truth shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CTR-Guided Generative Query Suggestion in Conversational Search." pith.science (2026). https://pith.science/paper/LIX4NMID

@misc{pith2026250704072,
  author       = {Pith},
  title        = {Pith review of: CTR-Guided Generative Query Suggestion in Conversational Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIX4NMID}},
  note         = {Machine review of arXiv:2507.04072}
}
read the original abstract

Generating effective query suggestions in conversational search requires aligning model outputs with user preferences, which is challenging due to sparse and noisy click signals. We propose GQS, a generative framework that integrates click modeling and preference optimization to enhance real-world user engagement. GQS consists of three key components: (1) a Multi-Source CTR Modeling module that captures diverse contextual signals to estimate fine-grained click-through rates; (2) a Diversity-Aware Preference Alignment strategy using CTR-weighted Direct Preference Optimization (DPO), which balances relevance and semantic diversity; and (3) a CTR-Calibrated Iterative Optimization process that jointly refines the CTR and generation models across training rounds. Experiments on two real-world tasks demonstrate that GQS outperforms strong baselines in CTR, relevance, and diversity.

Figures

Figures reproduced from arXiv: 2507.04072 by the authors.

Figure 1
Figure 1. Examples of AI assistant interfaces where [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of our proposed method. (a) The overall pipeline of the framework. (b) Multi-Source [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. CTR progression over training rounds in Task [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: CTR model performance (AUC and Logloss) under different architectural components. Aligned scenario, GQS improves CTR, relevance, and diversity by 5.21, 2.28, and 22.66 points, re￾spectively. 4.3 Ablation Studies In this subsection, we present experiments to eval￾uate t…
Figure 6
Figure 6. Figure 6: COO query information construction and refilling as side information. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Sensitivity analysis results. (a) Effect of diver [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Illustration of the three query suggestion tasks [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 8 canonical work pages

  1. [1]

    Andrea Bacciu, Enrico Palumbo, Andreas Damianou, Nicola Tonellotto, and Fabrizio Silvestri. 2024. https://arxiv.org/abs/2405.19749 Generating query recommendations via llms . Preprint, arXiv:2405.19749

    arXiv 2024

  2. [2]

    Jiaju Chen, Wang Wenjie, Chongming Gao, Peng Wu, Jianxiong Wei, and Qingsong Hua. 2024. https://doi.org/10.1145/3626772.3657736 Treatment effect estimation for user interest exploration on recommender systems . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, page 1861–1871, New...

    arXiv 2024

  3. [3]

    Xiaocong Chen, Siyu Wang, Julian McAuley, Dietmar Jannach, and Lina Yao. 2023. https://arxiv.org/abs/2308.11336 On the opportunities and challenges of offline reinforcement learning for recommender systems . Preprint, arXiv:2308.11336

    work page Pith review arXiv 2023

  4. [4]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. https://arxiv.org/abs/2402.01306 Kto: Model alignment as prospect theoretic optimization . Preprint, arXiv:2402.01306

    arXiv 2024

  5. [5]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian Mcauley. 2023. https://doi.org/10.1145/3583780.3614949 Large language models as zero-shot conversational recommenders . In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM '23, ...

    arXiv 2023

  6. [6]

    Md Aminul Islam, Kathryn Vasilaky, and Elena Zheleva. 2025. https://arxiv.org/abs/2506.06989 Correcting for position bias in learning to rank: A control function approach . Preprint, arXiv:2506.06989

    work page arXiv 2025

  7. [7]

    Jinqiu Jin, Sihao Ding, Wenjie Wang, and Fuli Feng. 2024. https://doi.org/10.1145/3589335.3651576 Understanding and counteracting feature-level bias in click-through rate prediction . In Companion Proceedings of the ACM Web Conference 2024, WWW '24, page 838–841, New York, NY, USA. Association for Computing Machinery

    arXiv 2024

  8. [8]

    Yachen Kang, Diyuan Shi, Jinxin Liu, Li He, and Donglin Wang. 2023. https://arxiv.org/abs/2305.16217 Beyond reward: Offline preference-guided policy optimization . Preprint, arXiv:2305.16217

    arXiv 2023

Show all 24 references
  1. [9]

    Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis, Jelena Luketina, Eric Hambro, Edward Grefenstette, and Roberta Raileanu. 2024. https://arxiv.org/abs/2310.06452 Understanding the effects of rlhf on llm generalisation and diversity . Preprint, arXiv:2310.06452

    2024 arXiv

  2. [10]

    Wenhan Liu, Ziliang Zhao, Yutao Zhu, and Zhicheng Dou. 2024. https://doi.org/10.1145/3589334.3645424 Mining exploratory queries for conversational search . In Proceedings of the ACM Web Conference 2024, WWW '24, page 1386–1394, New York, NY, USA. Association for Computing Machinery

    2024

  3. [11]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. https://arxiv.org/abs/2405.14734 Simpo: Simple preference optimization with a reference-free reward . Preprint, arXiv:2405.14734

    2024 arXiv

  4. [12]

    Kenny Peng, Manish Raghavan, Emma Pierson, Jon Kleinberg, and Nikhil Garg. 2023. https://arxiv.org/abs/2307.15142 Reconciling the accuracy-diversity trade-off in recommendations . Preprint, arXiv:2307.15142

    2023 arXiv

  5. [13]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/a85b405ed65c6477a4fe8302b5e06ce7-Paper-Conference.pdf Direct preference optimization: Your language model is ...

    2023

  6. [14]

    Lei Sang, Honghao Li, Yiwen Zhang, Yi Zhang, and Yun Yang. 2024. https://doi.org/10.1145/3681785 Adagin: Adaptive graph interaction network for click-through rate prediction . ACM Trans. Inf. Syst., 43(1)

    2024 doi

  7. [15]

    Xiaobin Shen, Daniel Lee, Sumit Ranjan, Sai Sree Harsha, Pawan Sevak, and Yunyao Li. 2024. Enhancing discoverability in enterprise conversational systems with proactive question suggestions. arXiv preprint arXiv:2412.10933

    2024 arXiv

  8. [16]

    Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Ernie 2.0: A continual pre-training framework for language understanding. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8968--8975

    2020

  9. [17]

    Shuting Wang, Xin Yu, Mang Wang, Weipeng Chen, Yutao Zhu, and Zhicheng Dou. 2024 a . https://arxiv.org/abs/2406.12566 Richrag: Crafting rich responses for multi-faceted queries in retrieval-augmented generation . Preprint, arXiv:2406.12566

    2024 arXiv

  10. [18]

    Zhenduo Wang, Yuancheng Tu, Corby Rosset, Nick Craswell, Ming Wu, and Qingyao Ai. 2023. https://doi.org/10.1145/3543507.3583420 Zero-shot clarifying question generation for conversational search . In Proceedings of the ACM Web Conference 2023, WWW '23, page 3288–3298, New York...

    2023

  11. [19]

    Zheng Wang, Bingzheng Gan, and Wei Shi. 2024 b . https://doi.org/10.1145/3589334.3645365 Multimodal query suggestion with multi-agent reinforcement learning from human feedback . In Proceedings of the ACM Web Conference 2024, WWW '24, page 1374–1385, New York, NY, USA. Associa...

    2024

  12. [20]

    Yao Xiao, Hai Ye, Linyao Chen, Hwee Tou Ng, Lidong Bing, Xiaoli Li, and Roy Ka-wei Lee. 2025. Finding the sweet spot: Preference data construction for scaling preference optimization. arXiv preprint arXiv:2502.16825

    2025 arXiv

  13. [21]

    Rosie Zhao, Alexandru Meterez, Sham Kakade, Cengiz Pehlevan, Samy Jelassi, and Eran Malach. 2025. https://arxiv.org/abs/2504.07912 Echo chamber: Rl post-training amplifies behaviors learned in pretraining . Preprint, arXiv:2504.07912

    2025 arXiv

  14. [22]

    Ziliang Zhao and Zhicheng Dou. 2024. https://doi.org/10.1145/3589334.3645712 Generating multi-turn clarification for web information seeking . In Proceedings of the ACM Web Conference 2024, WWW '24, page 1539–1548, New York, NY, USA. Association for Computing Machinery

    2024

  15. [23]

    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...

  16. [24]

    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 6, 2026 · model on record in the stance chip above.