Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

AIDE: Agentically Improve Visual Language Model with Domain Experts

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 A vision-language model can improve itself by rewriting its own weak training examples with help from lightweight domain experts.

desk verdict A genuinely new pipeline for self-improving VLMs through expert tools, but the experiments don't isolate the expert contribution and the reported numbers are inconsistent. read the letter →

arxiv 2502.09051 v1 pith:56BDYFOX submitted 2025-02-13 cs.CV cs.AIcs.MA

Ming-Chang Chiu , Fuxiao Liu , Karan Sapra , Andrew Tao , Yaser Jacoob , Xuezhe Ma , Zhiding Yu , Guilin Liu This is my paper
classification cs.CVcs.AIcs.MA
keywords visuallanguagemodelsself-improvementdomainexpertsagenticframeworkdatasynthesisOCRgroundingbenchmarkevaluation
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 claims that a vision-language model can improve itself without a larger teacher or human labels. The proposed loop, AIDE, lets the model's own Selector choose which training examples need work and which lightweight specialist tool to call, then a Synthesizer rewrites those examples into enriched, reasoning-flavored answers; retraining on the rewritten set lifts Eagle-8B by 1.2 percent on the multimodal reasoning benchmark MMMU, 0.77 percent on MMBench, 52 points on MME, and about 1 percent on MathVista and ChartQA. If true, this removes the bottleneck of needing a superior model to distill from and makes continuous self-improvement plausible for state-of-the-art systems.

What carries the argument

The central machinery is a two-agent loop: a Selector that reads each training instance together with descriptions of available expert tools and decides whether to call one, and a Synthesizer that aggregates the original instance, the expert output, and explicit conflict resolution into a new answer. A third device, small-step prompting, breaks the rewriting into a single focused task, such as generating only the reasoning, because the same VLM asked to produce a full enriched answer in one prompt often fails to follow instructions. N-gram filtering is the quality gate before integration.

What would settle it

Score a random sample of rewritten answers against ground-truth labels, either by human judges or a strong held-out evaluator; if the rewritten answers are not more often correct than the original answers, the reported benchmark gains are best explained by distribution shift rather than by improved capability.

Watch

Extended reading notes

Core claim

On the paper's own terms, AIDE is a way for a VLM to improve its own training data by acting as an agent over that data. The Selector inspects each instance, judges whether a domain expert could add useful information, and invokes the expert; the Synthesizer then merges the original question-answer, the expert output, and conflict resolution into a new, more detailed response. After n-gram filtering, retraining on these enriched instances improves Eagle-8B on MMMU by 1.2%, MMBench by 0.77%, MME by 52, MathVista by 1.1%, and ChartQA by 1.1%, without using a larger VLM or human supervision.

Load-bearing premise

The rewritten training answers are assumed to be at least as correct as the originals, even though the same model writes them and only n-gram filtering checks their quality.

Editorial extensions

If this is right

  • A VLM at the frontier no longer needs a superior model to improve; the same model can drive its own training-data enrichment.
  • The loop can be redirected at new skills by adding specialist tools, so improvements can track the tasks a user cares about.
  • Selector choices double as a data-quality signal: the model flagged OCR-heavy subsets such as synthdog, suggesting AIDE can audit dataset difficulty.
  • Because only lightweight off-the-shelf experts and one training iteration are used, the cost stays well below distilling from a frontier teacher.
  • With verifiers added, the same mechanism could become a continuous training loop rather than a one-shot data rewrite.

Reading between the lines

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

  • The authors leave implicit that the benchmark gains have not been shown to come from increased correctness; a control that scores rewritten answers against gold labels would separate capability gains from distribution shift.
  • A natural extension is to use the Selector's per-source selection rates as a difficulty prior for curriculum learning, since it consistently over-selected document-heavy data.
  • Because the Synthesizer is the very model being improved, the loop may plateau or encode its own errors in the absence of any external correctness signal; combining AIDE with a verifier or reward model is the obvious next test.
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. AIDE is a four-stage agentic pipeline in which a VLM (Eagle-8B) acts as both Selector and Synthesizer: the Selector identifies training instances that could benefit from additional expert information and invokes two lightweight domain experts (PaddleOCR and Grounded-SAM); the Synthesizer then rewrites the original instance into a more detailed, reasoning-style response, optionally retaining the original answer; the rewritten data are filtered by simple n-gram heuristics and fine-tuned into the base model. Experiments are reported on MMMU, MME, MMBench, MathVista, ChartQA, SciQA, POPE, and TextVQA, claiming notable gains without larger VLMs or human supervision. The paper also analyzes which data sources the Selector chooses and provides qualitative examples of enriched responses.

Significance. If the central claim holds, AIDE would be a meaningful contribution: it offers a way to improve an intermediate-scale VLM by combining self-generated reasoning with outputs of small specialized tools, avoiding the need for a larger teacher. The idea of a Selector/Synthesizer loop grounded in external expert models is novel and plausibly extensible. The paper also ships a concrete experimental setup with a data selection analysis. However, significance is currently conditional because the evidence table has internal inconsistencies, the role of the domain experts is not isolated by any ablation, and the quality of synthesized responses is not verified. These gaps prevent the reader from attributing the measured gains to the paper's stated mechanism.

major comments (4)
  1. [Section 4, Table 2] The primary evidence table is internally inconsistent with the claims in the text. The text states ChartQA gains of +1.1% and MME gains of +52, but Table 2 shows ChartQA changing from 80.24 (baseline) to 79.64 (small-step, a drop of 0.60) or to 80.42 (retention, +0.18). For MME, the sum of perception and cognition scores gives +109.2 for small-step and +52.7 for retention; neither matches the stated "+52" as a single number with a clear referent. The TextVQA baseline 7.469 also appears to be a typo for roughly 74.69. These discrepancies need to be corrected and the text must be reconciled with the table before the experimental claims can be evaluated.
  2. [Section 4.1, Table 2] The central mechanism attributed to the domain experts is never isolated. Table 2 compares Eagle-Baseline with Eagle-Small-step and Eagle-Retention, but both AIDE conditions include the full pipeline with PaddleOCR and Grounded-SAM outputs. There is no ablation in which the Synthesizer receives only the original instance and its own small-step reformatting prompt, without any expert tool output. Because the Synthesizer is the same Eagle-8B and the pipeline changes the response format, the reported gains could in principle come from ordinary self-training on reformatted data or from a favorable distribution shift rather than from expert information. This ablation is load-bearing for the paper's claim of "leveraging lightweight domain experts" and should be included.
  3. [Section 3.1] Quality control for synthesized responses is limited to unspecified n-gram filtering and the small-step prompt format; the paper does not measure whether the rewritten answers are factually correct. Since the Synthesizer is the same Eagle-8B being improved, fluent but incorrect rewrites could be encoded during fine-tuning and still raise benchmark numbers if the benchmark distributions align with the rewritten style. The paper itself notes in Section 5 that verifiers could be added later, which confirms that this is an acknowledged limitation. A sample-level human or automated correctness check on the synthesized instances, plus a comparison of performance on verified versus unverified subsets, would substantially strengthen the claim that the gains reflect genuine capability improvement rather than statistical alignment.
  4. [Section 4, Experimental Setup] The generality of the results is currently under-supported. Eagle-8B serves as both Selector and Synthesizer, the baseline is reproduced from the same group's model, and only one iteration of the loop is evaluated. This leaves open the possibility that the gains are specific to the Eagle design or to the chosen data mixture, rather than a property of the AIDE framework. Additional evidence, such as applying AIDE to a second base VLM or varying the Selector/Synthesizer pairing, would help establish that the mechanism transfers. Releasing the code and the selected/synthesized data would also be essential for reproducibility, but no such release is mentioned.
minor comments (6)
  1. [Table 2] The TextVQA baseline value 7.469 is implausibly low for a trained 8B VLM and is likely a typo for 74.69; please correct it.
  2. [Section 4, Results paragraph] The stated MMBench improvement of 0.77% is not directly visible in Table 2, where MMBench changes from 74.1 to 74.4 (small-step) or 74.8 (retention); please clarify how 0.77 is computed.
  3. [References] The citation for PaddleOCR appears as "(pad)" in the text and references, which is cryptic; please use a standard author-year citation.
  4. [Section 4, Results paragraph] The sentence ending with "(Saikh et al., 2022; Li et al., 2023)" appears to cite ScienceQA and POPE references as support for the performance gains, but these are benchmark references and the connection is unclear; please reposition or remove them.
  5. [Figure 3] The left panel's y-axis label and the exact ordering of datasets are not described; please add axis labels and a legend so the breakdown of the 950K selected instances is readable.
  6. [Abstract and Section 4] The abstract says the method improves "state-of-the-art systems," but experiments are reported only for Eagle-8B, an intermediate-scale model; please temper the claim to match the experiment scale.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AIDE's gains are empirical benchmark results, not derived from their own definitions; the self-referential loop lacks a control but is not definitionally circular.

full rationale

The paper reports an empirical pipeline (select instances, execute expert tools, synthesize enriched answers, fine-tune) and evaluates on external benchmarks such as MMMU, MMBench, MME, MathVista, and ChartQA. The claimed improvements are measured outcomes, not quantities derived from the framework's definitions; no parameter is fitted to a benchmark and then reported as a prediction. Using Eagle-8B as both Selector and Synthesizer is self-referential in spirit, but not circular in the formal sense: the synthesizer receives external expert outputs (PaddleOCR, Grounded-SAM) plus original instances, and the evaluation is on held-out benchmarks. The lack of an ablation that removes expert outputs is a real experimental confound—gains could partly come from small-step prompting or distribution shift—but a missing control is a correctness/experimental-design concern, not a circular reduction. Citations to the authors' prior Eagle model and to Chiu et al. (2024a,b) are provenance or speculative ('may explain'), not load-bearing justifications for the central claim. No equation, fitted parameter, or citation chain reduces a reported result to its own input; therefore the paper exhibits no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The pipeline has no explicit fitted numerical constants, but it leans on several unverified assumptions: expert outputs are correct and useful, the VLM Selector judges data quality reliably, and the training data does not contaminate the benchmarks. The n-gram filtering threshold and the number of selected instances are unspecified free choices. No new physical entities are introduced; Selector and Synthesizer are software roles within the pipeline.

free parameters (2)
  • n-gram filtering threshold = Not reported
    Integration (Section 3.1) says simple heuristics like n-gram filtering are used; the threshold and n are unspecified, and no ablation shows their effect.
  • Selected data budget = ~950K instances
    The VLM Selector chose about 950K of 5M multimodal instances; the fraction is a design-dependent output, not a controlled variable, and no sensitivity analysis is given.
assumptions (4)
  • domain assumption Expert tool outputs are factually correct and useful for training
    Section 4 (Expert Tool Choice) invokes PaddleOCR and Grounded-SAM as lightweight experts, but no verification of the correctness of their outputs on selected instances is reported.
  • domain assumption The VLM can act as a reliable judge of data quality
    Section 3 (Selector) says the selector judges whether expert information may be beneficial, with no oracle or human validation of the selection decisions.
  • domain assumption Enriched data transfers to benchmarks without dataset contamination
    No contamination analysis is provided between the Cambrian1-7M training data and the reported evaluation benchmarks (MMMU, MMBench, MME, etc.).
  • domain assumption The reproduced Eagle-8B baseline matches the public checkpoint
    Table 2 labels 'Eagle-Baseline (reproduced)' but does not state the exact checkpoint, training recipe, or random seed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AIDE: Agentically Improve Visual Language Model with Domain Experts." pith.science (2026). https://pith.science/paper/56BDYFOX

@misc{pith2026250209051,
  author       = {Pith},
  title        = {Pith review of: AIDE: Agentically Improve Visual Language Model with Domain Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56BDYFOX}},
  note         = {Machine review of arXiv:2502.09051}
}
read the original abstract

The enhancement of Visual Language Models (VLMs) has traditionally relied on knowledge distillation from larger, more capable models. This dependence creates a fundamental bottleneck for improving state-of-the-art systems, particularly when no superior models exist. We introduce AIDE (Agentic Improvement through Domain Experts), a novel framework that enables VLMs to autonomously enhance their capabilities by leveraging specialized domain expert models. AIDE operates through a four-stage process: (1) identifying instances for refinement, (2) engaging domain experts for targeted analysis, (3) synthesizing expert outputs with existing data, and (4) integrating enhanced instances into the training pipeline. Experiments on multiple benchmarks, including MMMU, MME, MMBench, etc., demonstrate AIDE's ability to achieve notable performance gains without relying on larger VLMs nor human supervision. Our framework provides a scalable, resource-efficient approach to continuous VLM improvement, addressing critical limitations in current methodologies, particularly valuable when larger models are unavailable to access.

Figures

Figures reproduced from arXiv: 2502.09051 by the authors.

Figure 1
Figure 1. AIDE Workflow. AIDEconsists of two agents, a Selector and a Synthesizer. The Selector interacts with the data instances and autonomously invoke the expert tools as it deems fit. The Synthesizer collects information from the original data instances along with outputs from the select experts and generate enriched response. Model RefCOCO RefCOCO+ RefCOCOg val testA testB val testA testB val test general-purpose model K… view at source ↗
Figure 2
Figure 2. Small-step Prompting. We observe even when VLM is able to answer the query (middle-column), sometimes the instruction following is not stable. And simplifying the prompt into smaller steps by giving the answer (last column) gives more detailed responses. See Sec. 4.3 for comparisons. 3.1 Integration After generating enhanced samples, the integration incorporates them back into the training pipeline. This involves fi… view at source ↗
Figure 3
Figure 3. Left: Breakdown of selected data instances by VLM-Selector. Synthdog takes the most proportion of the selection. Right: Ratio of data instances selected by the VLM-Selector to the total instances in the original Cambrian-1 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparisons of the original and the new answer produced by [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VL-GenRM: Enhancing Vision-Language Verification via Vision Experts and Iterative Training

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A vision-expert-filtered, chain-of-thought-guided, iteratively fine-tuned reward model boosts a compact 7B model's ability to judge vision-language responses, especially detecting hallucinations.

Reference graph

Works this paper leans on

27 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://github.com/PaddlePaddle/PaddleOCR

    PaddleOCR . https://github.com/PaddlePaddle/PaddleOCR. Accessed: 2024-06-30

    work page 2024

  2. [2]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski,...

    arXiv 2022

  3. [3]

    Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. 2023. ShareGPT4V : Improving large multi-modal models with better captions. arXiv:2311.12793

    arXiv 2023

  4. [4]

    Ming-Chang Chiu, Yingfei Wang, Derrick Eui Gyu Kim, Pin-Yu Chen, and Xuezhe Ma. 2024 a . https://arxiv.org/abs/2212.08650 Colorsense: A study on color vision in machine visual recognition . Preprint, arXiv:2212.08650

    work page Pith review arXiv 2024

  5. [5]

    Ming-Chang Chiu, Shicheng Wen, Pin-Yu Chen, and Xuezhe Ma. 2024 b . https://arxiv.org/abs/2412.03927 Megacoin: Enhancing medium-grained color perception for vision-language models . Preprint, arXiv:2412.03927

    arXiv 2024

  6. [6]

    Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jan Kautz, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. 2024. https://www.arxiv.org/abs/2407.17453 Vila ^2 : Vila augmented vila . arXiv preprint arXiv:2407.17453

    arXiv 2024

  7. [7]

    Hao Fei, Yuan Yao, Zhuosheng Zhang, Fuxiao Liu, Ao Zhang, and Tat-Seng Chua. 2024. From multimodal llm to human-level ai: Modality, instruction, reasoning, efficiency and beyond. In LREC-Coling Tutorials

    work page 2024

  8. [9]

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. 2024. https://arxiv.org/abs/2306.13394 Mme: A comprehensive evaluation benchmark for multimodal large language models . Preprint, arXiv:2306.13394

    arXiv 2024

Show all 27 references
  1. [10]

    Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. 2023. https://www.arxiv.org/abs/2302...

    2023 arXiv

  2. [11]

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. arXiv:2305.10355

    2023 arXiv

  3. [12]

    Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023 a . https://www.arxiv.org/abs/2306.14565 Mitigating hallucination in large multi-modal models via robust instruction tuning . arXiv preprint arXiv:2306.14565

    2023 arXiv

  4. [13]

    Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. 2023 b . MMC : Advancing multimodal chart understanding with large-scale instruction tuning. arXiv:2311.10774

    2023 arXiv

  5. [14]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023 c . https://www.arxiv.org/abs/2304.08485 Visual instruction tuning . arXiv preprint arXiv:2304.08485

    2023 arXiv

  6. [15]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. 2023 d . https://www.arxiv.org/abs/2303.05499 Grounding dino: Marrying dino with grounded pre-training for open-set object detection . arXiv prepri...

    2023 arXiv

  7. [16]

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. 2023 e . MMBench : Is your multi-modal model an all-around player? arXiv:2307.06281

    2023 arXiv

  8. [17]

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2024. MathVista : Evaluating mathematical reasoning of foundation models in visual contexts. In ICLR

    2024

  9. [18]

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 2022. ChartQA : A benchmark for question answering about charts with visual and logical reasoning. arXiv:2203.10244

    2022 arXiv

  10. [19]

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. 2024. https://www.arxiv.org/abs/2401.14159 Grounded sam: Assembling open-wo...

    2024 arXiv

  11. [20]

    Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. 2022. ScienceQA : A novel resource for question answering on scholarly articles. International Journal on Digital Libraries

    2022

  12. [21]

    Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, Bryan Catanzaro, Andrew Tao, Jan Kautz, Zhiding Yu, and Guilin Liu. 2024. Eagle: Exploring the design space for multimodal llms with mixtu...

    2024 arXiv

  13. [22]

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Ziteng Wang, Rob Fergus, Yann LeCun, and Saining Xie. 2024. https://www.arxiv.org/abs/2406.16860v2 Cambrian-1: A fully open, vision-...

    2024 arXiv

  14. [23]

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560

    2022 arXiv

  15. [24]

    Yue Xiang, Ni Yuansheng, Zhang Kai, Zheng Tianyu, Liu Ruoqi, Zhang Ge, Stevens Samuel, Jiang Dongfu, Ren Weiming, Sun Yuxuan, et al. 2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conferenc...

    2024

  16. [25]

    Liu Yuan, Duan Haodong, Zhang Yuanhan, Li Bo, Zhang Songyang, Zhao Wangbo, Yuan Yike, Wang Jiaqi, He Conghui, Liu Ziwei, et al. 2024. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216--233. Springer

    2024

  17. [26]

    Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pe...

    2021 arXiv

  18. [27]

    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...

  19. [28]

    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.