Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A multimodal LLM reading parsed datasheet images, tables, and a few retrieved examples can generate EDA symbol and footprint files for simple and moderate components—86% symbol accuracy, 80% footprint accuracy—and can populate a 1,000-part

desk verdict Worth reading for the new task and benchmark, but the headline accuracy figures contradict the paper's own metric definitions and tables. read the letter →

arxiv 2607.19767 v1 pith:HJUQOBHO submitted 2026-07-22 cs.AI

Yichen Shi , Yuzhi Liu , Zhuofu Tao , Li Huang , Yuhao Gao , Ting-Jung Lin , Lei Hel This is my paper
classification cs.AI
keywords PCBdesigncomponentlibraryschematicsymbolfootprintmultimodallargelanguagemodelin-contextlearningvisualpromptingchain-of-thought
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

SFgen is an agentic pipeline that turns electronic-component datasheets into the two files PCB designers need: schematic symbols and layout footprints. The paper argues that a multimodal large language model (MLLM), prompted with extracted datasheet images and tables, a few retrieved template examples, numbered pad annotations, chain-of-thought reasoning, and a self-correction step, can generate these files in EDA text format with 86% symbol accuracy and 80% footprint accuracy on a 100-component benchmark. The same pipeline built SFnet, a database of datasheets, symbols, and footprints for 1,000 components. A sympathetic reader cares because manual component-library creation is slow and error-prone; if the claim holds, a large share of the library work for simple and moderate components can be automated and folded into a continuously growing resource for automated PCB design.

What carries the argument

The carrying mechanism is the SFgen pipeline, a datasheet-to-file flow: a PDF parser extracts images and text, an image classifier sorts them into symbol diagrams, footprint diagrams, and pin-function tables, and an MLLM (a model that reads both text and images) converts the resulting constraints into EDA text files. The MLLM's output is steered by in-context learning from a template library, visual prompting that places numeric IDs on pad diagrams so the model can refer to pads by index, chain-of-thought prompting that asks the model to describe pads step by step, and a second-pass modification prompt that re-reads the generated footprint and corrects symmetry and position errors.

What would settle it

Take a datasheet SFgen handles well, manually extract its pin table and pad coordinates, and compare them against both SFgen's parsed constraints and its generated files: if SFgen's extracted constraints already contain a missing or wrong pin/pad, or if the MLLM's output matches the datasheet but not the paper's ground-truth file, then the accuracy numbers measure something other than generation fidelity and the central claim is not settled.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that symbol and footprint generation can be treated as conditional text generation: extract pin configurations, pin functions, pad sizes, pad positions, package type, and pin count from a datasheet; feed them to an MLLM along with a small number of similar example pairs; and have the model emit a symbol or footprint file in a standard EDA text format. The paper reports ACC_N of 96% for basic symbols and 86% for standard symbols, with footprint ACC_N of 96% and 80% for the same groups; for high-density components (41–100 pins) exact pin/pad count accuracy drops to 0%, and the paper uses this to delimit the method's current reach. Correct-coun

Load-bearing premise

The accuracies presuppose that the datasheet-extraction stage recovers every pin, pad, and dimension completely and correctly and that the author-built ground-truth files are themselves correct; if either fails, the reported figures would not describe real device behavior.

Editorial extensions

If this is right

  • For components with up to roughly 40 pins, a first-pass symbol and footprint can be produced in minutes per part instead of manual hours, with designers correcting only the remaining discrepancies.
  • SFnet gives the field a common resource of 1,000 paired datasheets, symbols, and footprints, which can serve as training data or test cases for later PCB-automation work.
  • The 100-component benchmark and its metrics—pin/pad-count accuracy, type accuracy, normalized area and position differences, and mean opinion scores—offer a standard way to compare future methods.
  • Because the outputs are text-format EDA files, generated libraries are editable and portable across tools, so partial automation lowers the barrier to building vendor-specific libraries.
  • The sharp accuracy drop at 41–100 pins shows that high-density parts still need manual or improved treatment; the method's current promise is in automating the bulk of simple and moderate components.

Reading between the lines

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

  • The paper reports no extraction-quality metric, so the tightest next test is whether the bottleneck is generation or the initial parse: a reader could measure pin/pad recovery from the PDF parser separately and compare it to final accuracy.
  • If template retrieval matters as much as the prompt design suggests, a self-improving loop—feeding corrected outputs back into the template library—could push generation quality toward the high-density range without new model training.
  • Because accuracy is defined on counts, types, and geometry rather than on electrical validity or manufacturability, a generated footprint that scores well on the stated metrics may still need design-rule checking before production; that gap is outside the paper's metrics.
  • The same agentic pattern could generalize to other library artifacts, such as 3D models or simulation models, which are also specification-to-file tasks with well-defined constraints.
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

3 major / 5 minor

Summary. The paper presents SFgen, an agentic MLLM-based pipeline that parses PDF datasheets of electronic components and generates schematic symbols and PCB footprints in EDA text formats. The pipeline combines image classification/extraction, in-context learning from a template library, visual prompting, and chain-of-thought prompting, with an iterative modification step for footprints. The authors report 86% symbol-generation accuracy and 80% footprint-generation accuracy, and use SFgen to construct SFnet, a database of 1,000 components with datasheets, symbols, and footprints. They also introduce a benchmark of 100 components and propose quantitative metrics (ACC_N, ACC_T, Dif_A, Dif_P) plus mean opinion scores (MOS-S, MOS-F) for evaluation.

Significance. If the claimed accuracy figures are reliable, SFgen would be a useful step toward automating component-library creation, which is currently a labor-intensive bottleneck in PCB design. The construction of SFnet as a first large-scale dataset of component symbols/footprints is also a potentially valuable community resource, and the proposed benchmark and metrics could support future comparisons. The paper's strengths include a clear pipeline description, concrete generation flows, and an ablation study for footprint generation. However, the central numerical claims are not currently supported by the paper's own tables, and the evaluation protocol has internal inconsistencies that must be resolved before the significance of the results can be assessed.

major comments (3)
  1. [§3.2, Eqs. (2)-(3); Tables 3 and 4] The definition of ACC_T states that it is computed only when the generated pin/pad count is correct; otherwise the entire component is considered incorrect. Under this definition, the High-Density group, which has ACC_N=0.00% in both Table 3 and Table 4, must have ACC_T=0.00% (or be undefined). Yet the tables report High-Density ACC_T=49.32% and 35.40%, respectively. This contradiction means the reported ACC_T values do not follow from the stated evaluation protocol. The authors must either correct the metric definition, recompute the tables, or explain how the reported ACC_T values were obtained.
  2. [Abstract and §6; Tables 3 and 4] The headline claims of '86% accuracy for symbol generation and 80% accuracy for footprint generation' cannot be derived from Tables 3 and 4. Weighted recomputation from the group sizes in Table 1 gives roughly 78% overall ACC_N for symbols and 71% overall ACC_N for footprints, not 86%/80%. The values 86% and 80% coincide with the Standard-group ACC_N in Tables 3 and 4, suggesting the headline may quote a subgroup. As written, the abstract and conclusion overstate the overall performance. The authors should report an explicit overall aggregation (weighted by group size) or clearly state that the headline figures refer to a specific complexity group.
  3. [§2.2, Fig. 2; §4] The pipeline's accuracy depends on the datasheet extraction stage producing complete and correct constraints: pin functions, pin configurations, pad positions, and pad sizes must all survive PDF parsing and image classification. The paper reports no metric for extraction quality. If a pin or pad is missed or misread, the downstream MLLM cannot recover it, so the reported ACC_N/ACC_T values do not reflect the true end-to-end accuracy. I recommend adding an extraction-stage evaluation (e.g., correct/incorrect extraction rate on the benchmark) or at least a manual audit of extraction failures to support the end-to-end claims.
minor comments (5)
  1. [§3.1, Table 1] The group complexity thresholds are '1-5', '6-20', '21-40', and '41-100' pins. Figure 6 shows the distribution but the text does not justify these thresholds. A brief rationale would help readers interpret the benchmark design.
  2. [§3.2, Table 2] The MOS rubric is defined on a 1-5 scale, but no inter-rater reliability or number of evaluators is reported. Since MOS-S and MOS-F are used in the main results, at least the number of evaluators and whether scores were averaged should be stated.
  3. [§4.2, Fig. 7] The figure caption lists 'L293DDSymbol' / 'L, D, SW' / 'NRF905' but the text does not explain what the reader should inspect in the visual comparison. Adding annotations for the key differences (e.g., pin labels, pin counts) would make the qualitative evaluation more informative.
  4. [§5, Table 5] The component type 'MOS' is ambiguous; it should be spelled out (e.g., MOSFET) for consistency with the other categories. Also, Table 5 lists quantities that sum to 1000, but the text says 'more than 1,000'; please clarify whether the database contains exactly 1000 or a larger number.
  5. [§2.4, Fig. 5] The visual-prompting example shows a footprint with six pads, but the text says 'the MLLM incorrectly identifies four pads' without VP. The figure is helpful, but the caption could be more explicit about how the numbering in the 'With VP' panel maps to the final pad count.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy claims are empirical evaluations against manually built ground truth, not derivations from the input definitions.

full rationale

I walked the claimed derivation chain. SFgen's central outputs are generated by an MLLM pipeline from datasheets and then compared with manually designed ground-truth symbol and footprint files (§3.1, Tables 3–4). No equation defines the predicted output in terms of the evaluation target: Eq. (1) is prompt construction, and Eqs. (2)–(5) are metrics applied after generation. The benchmark, ground-truth labels, and MOS rubric are created in-house, which is a self-validation/independence concern, but it is not definitional circularity: the MOS rubric (Table 2) is not fed back into the prompt, and the accuracy metrics are not fitted parameters. There is no load-bearing self-citation; the references are external tools and models (GPT-4, MinerU, etc.), and no uniqueness theorem or prior result by the same authors is used to force the method. I therefore find no circular step. Separately, and as a correctness concern rather than a circularity one, the Abstract's '86% symbol-generation accuracy and 80% footprint-generation accuracy' cannot be reproduced from Tables 3–4 under the paper's own ACC_T rule, since the High-Density group has ACC_N = 0.00% yet non-zero ACC_T (49.32% and 35.40%). This internal inconsistency should be fixed or explained by the authors, but it does not make the derivation circular.

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

The central accuracy claims rest on four unverified domain assumptions: extraction completeness, ground-truth correctness, file-format sufficiency, and ICL generalization. They also depend on hand-chosen procedural parameters (k, modification-loop iterations, group thresholds, LLM decoding settings). No physical invented entities are introduced.

free parameters (4)
  • ICL example count k and template-retrieval match
    SFgen selects 'best-matching' templates from a template library, but k and the matching function are not specified; these choices affect every generated file.
  • Number of footprint modification iterations
    The modified-prompt loop corrects symmetry/position errors, but the stopping rule or iteration count is not given; more iterations could change results.
  • Benchmark complexity thresholds (pins per group) = 1-5 / 6-20 / 21-40 / 41-100
    Hand-set group boundaries determine the aggregate accuracies in Tables 3-4; different boundaries would change the headline 86%/80% numbers.
  • LLM sampling parameters (model version, temperature, max tokens)
    Only 'GPT-4 through API calls' is reported; sampling variability and decoding parameters are uncontrolled and not disclosed.
assumptions (4)
  • domain assumption The MinerU PDF parser and the CNN/MLP image classifier correctly extract symbol diagrams, footprint diagrams, pin-function tables, and metadata from heterogeneous vendor datasheets.
    Invoked in §2.2 and Fig. 2; no extraction accuracy is reported, so errors propagate to every generated file.
  • domain assumption The ground-truth symbols and footprints in the §3.1 benchmark are authoritative and correct representations of the components.
    All ACC_N, ACC_T, DifA, and DifP scores compare against these files; if the GT is wrong or inconsistent, the metrics are invalid.
  • domain assumption A returned text file in the target EDA format that visually matches the GT is a sufficient criterion for a correct symbol or footprint.
    Metrics count pins/types and compare geometry, but do not check electrical connectivity, manufacturing rules, or design-rule violations.
  • domain assumption GPT-4's in-context learning from k examples generalizes to unseen components of the same package and pin-count class.
    The method is prompt-based and untrained; there is no evidence of the minimal template diversity needed for robustness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation." pith.science (2026). https://pith.science/paper/HJUQOBHO

@misc{pith2026260719767,
  author       = {Pith},
  title        = {Pith review of: Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJUQOBHO}},
  note         = {Machine review of arXiv:2607.19767}
}
read the original abstract

A rich and recognizable component library is the cornerstone of printed circuit board (PCB) design and generation. Traditionally, engineers manually create symbols and footprints and design PCB schematics, which is time-consuming and error-prone. Leveraging multimodal large language models (MLLMs), we develop SFgen, an agentic recognition and generation flow of symbol and footprint for electronic components. SFgen achieves 86% accuracy for symbol generation and 80% accuracy for footprint generation. We use the SFgen method to create SFnet, a database of symbols and footprints. It now has 1000 components and is expanding constantly, which lays the foundation for automatic generation of PCB designs.

Figures

Figures reproduced from arXiv: 2607.19767 by the authors.

Figure 1
Figure 1. Comparison of the traditional manual design process and SFgen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Based on the open-source PDF parsing tool [5], the pipeline extracts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. Information extraction from component datasheets. As summarized in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (10 more)
Figure 3
Figure 3. Figure 3: Symbol-generation flow: (1) select the best-matching ICL cases from the tem￾plate library, (2) construct a prompt with multimodal information from the target datasheet, (3) generate a symbol file with the MLLM, and (4) visualize and compare the result against ground tr…
Figure 4
Figure 4. Figure 4: Footprint-generation flow: (1) attach visual prompts to datasheet images, (2) select the best-matching ICL cases from the template library, (3) construct the full prompt, (4) generate a footprint with the MLLM, (5) incrementally improve the foot￾print through prompt mo…
Figure 5
Figure 5. Figure 5: Illustration of visual prompting (VP): without VP, the MLLM incorrectly iden￾tifies four pads; with VP, it correctly identifies six pads. 3 Evaluation Protocol 3.1 Benchmark Construction Electronic components vary substantially in functionality and package configu￾rati…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualization results of generated symbols as footprint complexity increases, with a sharp drop in the High-Density group because of the MLLM’s limited ability to recognize complex images [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visualization results of generated footprints 6 Conclusions and Future Work This paper presents SFgen, an MLLM-driven pipeline that extracts compo￾nent information from datasheets and generates symbol and footprint files. SF￾gen combines in-context learning (ICL), visu…
Figure 9
Figure 9. Figure 9: Ablation study of footprint generation [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Design prompt for symbol generation [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Modification prompt for footprint generation [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Design prompt for footprint generation [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 1 canonical work pages

  1. [1]

    Microprocess

    Alain Amroun , title =. Microprocess. Microsystems , volume =. 1994 , url =. doi:10.1016/0141-9331(94)90099-X , timestamp =

    work page doi:10.1016/0141-9331(94)90099-x 1994

  2. [2]

    1992 , publisher=

    Michel, Petra and Lauther, Ulrich and Duzy, Peter , volume=. 1992 , publisher=

    1992

  3. [3]

    Rutenbar , editor =

    Rob A. Rutenbar , editor =. Proceedings of the 2015 Symposium on International Symposium on Physical Design , pages =. 2015 , url =. doi:10.1145/2717764.2717780 , timestamp =

    arXiv 2015

  4. [4]

    2024 , organization=

    Lu, Yao and Liu, Shang and Zhang, Qijun and Xie, Zhiyao , booktitle=. 2024 , organization=

    2024

  5. [5]

    Sedra, Adel S and Smith, Kenneth C , year=

  6. [6]

    Marcus and Beatrice Santorini and Mary Ann Marcinkiewicz , title =

    Mitchell P. Marcus and Beatrice Santorini and Mary Ann Marcinkiewicz , title =. Comput. Linguistics , volume =. 1993 , timestamp =

    1993

  7. [7]

    2009 , organization=

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , booktitle=. 2009 , organization=

    2009

  8. [8]

    Razavi, Behzad , year=

Show all 49 references
  1. [9]

    Agarwal, Anant and Lang, Jeffrey , year=

  2. [10]

    Fernandez-Canque, Hernando Lautaro , year=

  3. [11]

    Zumbahlen, Hank , year=

  4. [12]

    Gray, Paul R and Hurst, Paul J and Lewis, Stephen H and Meyer, Robert G , year=

  5. [13]

    1989 , publisher=

    Horowitz, Paul and Hill, Winfield and Robinson, Ian , volume=. 1989 , publisher=

    1989

  6. [14]

    Mancini, Ron , year=

  7. [15]

    National Semiconductor Corporation , year=

  8. [16]

    2007 , publisher=

    Neamen, Donald A , volume=. 2007 , publisher=

    2007

  9. [17]

    2023 , organization=

    Fu, Yonggan and Zhang, Yongan and Yu, Zhongzhi and Li, Sixu and Ye, Zhifan and Li, Chaojian and Wan, Cheng and Lin, Yingyan Celine , booktitle=. 2023 , organization=

    2023

  10. [18]

    2024 , publisher=

    Wu, Haoyuan and He, Zhuolun and Zhang, Xinyun and Yao, Xufeng and Zheng, Su and Zheng, Haisheng and Yu, Bei , journal=. 2024 , publisher=

    2024

  11. [19]

    2023 , organization=

    Blocklove, Jason and Garg, Siddharth and Karri, Ramesh and Pearce, Hammond , booktitle=. 2023 , organization=

    2023

  12. [20]

    Liu, Shang and Fang, Wenji and Lu, Yao and Zhang, Qijun and Zhang, Hongce and Xie, Zhiyao , journal=

  13. [21]

    Chen, Lei and Chen, Yiqi and Chu, Zhufei and Fang, Wenji and Ho, Tsung-Yi and Huang, Yu and Khan, Sadaf and Li, Min and Li, Xingquan and Liang, Yun and others , journal=

  14. [22]

    2016 14th IEEE International New Circuits and Systems Conference (NEWCAS) , pages=

    Stas, Fran. 2016 14th IEEE International New Circuits and Systems Conference (NEWCAS) , pages=. 2016 , organization=

    2016

  15. [23]

    2021 , publisher=

    Budak, Ahmet Faruk and Gandara, Miguel and Shi, Wei and Pan, David Z and Sun, Nan and Liu, Bo , journal=. 2021 , publisher=

    2021

  16. [24]

    Yao, Wei and Shi, Yiyu and He, Lei and Pamarti, Sudhakar , booktitle=

  17. [25]

    Xing, Wei and Fan, Weijian and Liu, Zhuohua and Yao, Yuan and Hu, Yuanqi , booktitle=

  18. [26]

    arXiv preprint arXiv:1905.13319 , year=

    Mathqa: Towards interpretable math word problem solving with operation-based formalisms , author=. arXiv preprint arXiv:1905.13319 , year=

    1905 arXiv

  19. [27]

    The 36th Conference on Neural Information Processing Systems (NeurIPS) , year=

    Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering , author=. The 36th Conference on Neural Information Processing Systems (NeurIPS) , year=

  20. [28]

    arXiv preprint arXiv:2308.00692 , year=

    LISA: Reasoning Segmentation via Large Language Model , author=. arXiv preprint arXiv:2308.00692 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2311.00176 , year=

    Chipnemo: Domain-adapted llms for chip design , author=. arXiv preprint arXiv:2311.00176 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2305.14019 , year=

    Chipgpt: How far are we from natural language hardware design , author=. arXiv preprint arXiv:2305.14019 , year=

  23. [31]

    2023 , eprint=

    CodeGeeX: A Pre-Trained Model for Code Generation with Multilingual Evaluations on HumanEval-X , author=. 2023 , eprint=

    2023

  24. [32]

    arXiv preprint arXiv:2403.07506 , year=

    Robustness, security, privacy, explainability, efficiency, and usability of large language models for code , author=. arXiv preprint arXiv:2403.07506 , year=

    arXiv

  25. [33]

    Liu, Mingjie and Pinckney, Nathaniel and Khailany, Brucek and Ren, Haoxing , booktitle=

  26. [34]

    2023 Design, Automation & Test in Europe Conference & Exhibition (DATE) , pages=

    Benchmarking large language models for automated verilog rtl code generation , author=. 2023 Design, Automation & Test in Europe Conference & Exhibition (DATE) , pages=. 2023 , organization=

    2023

  27. [35]

    IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , year=

    Chateda: A large language model powered autonomous agent for eda , author=. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , year=

  28. [36]

    2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD) , pages=

    Gpt4aigchip: Towards next-generation ai accelerator design automation via large language models , author=. 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD) , pages=. 2023 , organization=

    2023

  29. [37]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    arXiv

  30. [38]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    arXiv

  31. [39]

    arXiv preprint arXiv:2302.13971 , year=

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    arXiv

  32. [40]

    arXiv preprint arXiv:2308.12966 , year=

    Qwen-vl: A frontier large vision-language model with versatile abilities , author=. arXiv preprint arXiv:2308.12966 , year=

    arXiv

  33. [41]

    Advances in Neural Information Processing Systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in Neural Information Processing Systems , volume=

  34. [42]

    arXiv e-prints , pages=

    The Role of Chain-of-Thought in Complex Vision-Language Reasoning Task , author=. arXiv e-prints , pages=

  35. [43]

    arXiv preprint arXiv:2310.11441 , year=

    Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v , author=. arXiv preprint arXiv:2310.11441 , year=

    arXiv

  36. [44]

    arXiv preprint arXiv:2203.08568 , year=

    In-context learning for few-shot dialogue state tracking , author=. arXiv preprint arXiv:2203.08568 , year=

    arXiv

  37. [45]

    arXiv preprint arXiv:2101.06804 , year=

    What Makes Good In-Context Examples for GPT- 3 ? , author=. arXiv preprint arXiv:2101.06804 , year=

    arXiv

  38. [46]

    Kapsalis, Timo , journal=

  39. [47]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    arXiv

  40. [48]

    Advances in Neural Information Processing Systems , volume=

    Layoutprompter: Awaken the design ability of large language models , author=. Advances in Neural Information Processing Systems , volume=

  41. [49]

    2024 , eprint=

    MinerU: An Open-Source Solution for Precise Document Content Extraction , author=. 2024 , eprint=

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.