Pith. sign in

REVIEW 4 major objections 4 minor 22 references

Enhancing Financial VQA in Vision Language Models using Intermediate Structured Representations

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Structured tables extracted from chart images, rather than the images themselves, sharply improve vision-language model reasoning on financial bar-chart questions, and fine-tuning DePlot on synthetic bar charts is enough to realize the…

desk verdict Clean internal result on same-distribution synthetic charts, but the financial-VQA generalization claim needs out-of-distribution evidence and released artifacts before it persuades. read the letter →

arxiv 2501.04675 v1 pith:IL4XMSSD submitted 2025-01-08 cs.CL cs.AIcs.CVcs.LG

Archita Srivastava , Abhas Kumar , Rajesh Kumar , Prabhakar Srinivasan This is my paper · ORCID
classification cs.CLcs.AIcs.CVcs.LG
keywords chartunderstandingvisualquestionansweringDePlotfine-tuningbarchartsstructuredrepresentationsfinancialdocumentsvisionlanguagemodels
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 argues that feeding a vision-language model a structured table extracted from a chart image, rather than the raw image, produces far more accurate numerical answers to financial chart questions. To get good tables, the authors fine-tune DePlot, a plot-to-table translation model, on 50,000 synthetic bar charts spanning simple, stacked, and grouped layouts. On a held-out test set the fine-tuned table extractor raises categorical mapping F1 from about 51% to 91%, and for one small vision-language model it cuts mean absolute percentage error on chart-based questions from 32.6% to 3.0%. The intended significance is that a relatively cheap intermediate representation, a linearized table, can make even small models competitive with much larger ones that try to read charts directly.

What carries the argument

The load-bearing object is DePlot, a modality-conversion module that turns a chart image into a linearized table string, using '|' to separate columns and '<0x0A>' to mark row ends. The authors fine-tune DePlot on a custom 50,000-image dataset of financial bar charts, then feed the resulting tables to downstream LLMs alongside the image. The evaluation uses Relative Mapping Similarity (RMS), which measures how well predicted row-and-column-to-value mappings align with ground truth, and Relative Number Set Similarity (RNSS), which checks numerical sets; these metrics make the improvement in table fidelity legible and connect the table-extraction step to the downstream QA gains.

What would settle it

Take a set of real financial charts from filings, earnings reports, or financial news with manually verified ground-truth values, run the fine-tuned table extractor and the same table-plus-image LLM setup on them, and compare MAPE and RMS F1 against the synthetic-test numbers; if the advantage over image-only queries shrinks or disappears, the claim that the gains carry outside the synthetic distribution is refuted.

Watch

Extended reading notes

Core claim

The central claim is that accurate intermediate structured representations are the bottleneck for chart question answering: once a chart is correctly rendered as a table, LLM reasoning over that table is dramatically better than over the image, and fine-tuning the table extractor on domain-specific synthetic charts is enough to make this work. The authors demonstrate that base DePlot produces tables with poor categorical alignment, reaching an RMS F1 of 50.93%, while the fine-tuned model reaches 91.06% on the same 1,000-image test set. Downstream, adding the fine-tuned table lowers Qwen2-VL-7B's MAPE from 32.63% to 2.98% and its RMSE from 90.99% to 8.10%, and it lets smaller models such as Qwen2-VL-7B and Llama-3.2-11B outperform GPT-4o when GPT-4o is given only the raw image. The paper presents this as evidence that the quality of the table, not the scale of the reasoning model, drives numerical question-answering performance.

Load-bearing premise

The whole result rests on the assumption that the synthetic bar charts used for training and testing behave like real-world financial charts; if real charts differ in layout, noise, or encoding, the measured accuracy gains may not transfer.

Editorial extensions

If this is right

  • Fine-tuning a chart-to-table model on domain-specific synthetic charts can produce large gains in table fidelity, with RMS F1 rising by roughly 40 percentage points, without changing the downstream LLM.
  • Providing a high-quality table alongside the image reduces MAPE for all tested LLMs, and the relative reduction is largest for the smaller models.
  • With structured tables, smaller vision-language models can beat a much larger model, GPT-4o, that reads the image directly.
  • Because a correct table is the key input, the quality of the intermediate representation, rather than model scale, dominates numerical QA performance in this setting.
  • The same dataset-generation recipe, with domain-specific labels and values, is claimed to extend the approach to other domains such as healthcare or marketing charts.

Reading between the lines

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

  • If table quality is truly the bottleneck, then any reliable chart-to-table system, not just DePlot, should reproduce the same downstream gains, making the claim testable in a model-agnostic way.
  • The large improvement likely reflects the LLM's difficulty in reading exact numeric values from images; once the table removes perceptual noise, residual errors should track table-extraction accuracy rather than image understanding.
  • Because training and test charts come from the same randomized generator, the measured gains probably overestimate performance on real-world charts with noise, OCR errors, and complex encodings; testing on real financial report pages would clarify this.
  • The same table-first recipe could be applied to other structured visual inputs, such as line charts, scatter plots, or diagrams, wherever a textual derendering of the visual exists.
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 / 4 minor

Summary. The paper proposes fine-tuning DePlot on 50,000 synthetic financial bar charts (simple, stacked, and grouped) and using the resulting tables as intermediate structured representations for question answering with vision-language models and LLMs. It reports that fine-tuned DePlot improves RNSS from 89.67% to 98.07% and RMS F1 from 50.93% to 91.06% on a held-out set from the same generator, and that providing the fine-tuned table to GPT-4o, Qwen2-VL-2B/7B, and LLaMa-11B reduces MAPE and RMSE relative to image-only prompting. An appendix provides qualitative examples comparing base and fine-tuned tables, along with zero-shot VLM inference on three charts.

Significance. The central idea—domain-adaptive fine-tuning of a plot-to-table module to improve downstream numerical reasoning—is sensible, and the internal experiments are self-consistent, providing a clear proof-of-concept on the synthetic distribution. The paper describes the dataset generation and metrics in enough detail to be reproduced in principle, and it explicitly defines the RNSS and RMS scoring rules. However, the central claim is about financial VQA broadly, and the evidence is entirely from a single synthetic generator with a fixed vocabulary and layout distribution. The paper's own limitations paragraph concedes the absence of evaluation on complex or out-of-distribution charts. Consequently, the headline numbers should be read as distribution-specific results, not as established generalization. The approach merits publication only after external validation on a public chart benchmark or real financial documents.

major comments (4)
  1. [§5.3, §8, Table IV] The training, test, and QA charts are all produced by the same Matplotlib/Seaborn generator with a fixed financial vocabulary (Section 5.1), and the QA pairs are generated from the same ground-truth tables. Because the fine-tuned DePlot is tested on images that share fonts, colors, label sets, and value ranges with training, the large improvements (RMS F1 50.93% to 91.06%; Qwen-7b MAPE 32.63% to 2.98%) may substantially reflect fitting the generator's distribution rather than general chart-reading ability. The limitations paragraph in Section 7 explicitly acknowledges that the model has not been evaluated on more complex visualizations and that domain-specific fine-tuning may limit generalization. To support the paper's general claim, an out-of-distribution evaluation on a public chart benchmark (e.g., ChartQA, ChartOCR, or a collection of real financial charts) is required; without it, the external-validity gap is load-bearing.
  2. [§6.2 vs. Table III] The base DePlot average RNSS is reported as 89.67% in the text of Section 6.2, but Table III lists it as 86.67%. This discrepancy is not explained, and it directly affects the reported RNSS improvement of "approximately 8.4 percentage points" in Section 8.1. The authors should correct the inconsistent number and recompute the gain consistently, or justify why the two values differ.
  3. [Table IV and §9.2] The QA results are point estimates from a single run over 100 charts, with no confidence intervals, error bars, or significance tests. Moreover, the base-table configuration is not consistently better than image-only: for GPT-4o, RMSE increases from 60.51 to 65.89, and for Qwen-2b, from 80.84 to 90.52. The statement that "transitioning from the image-only configuration to the fine-tuned table configuration consistently and significantly reduces both MAPE and RMSE" is supported for the fine-tuned table, but the base-table comparison should be reported with uncertainty and the negative cases should be discussed rather than swept into an average.
  4. [§7.3 and §7.2] The fine-tuning section specifies 10 epochs, AdamW, and an H100 GPU, but omits learning rate, batch size, gradient accumulation, warmup, the exact DePlot checkpoint used for initialization, and the input resolution or patch configuration (beyond the 1,024-patch limit). Without these details—and without code or dataset release—the fine-tuning results are difficult to reproduce independently. The authors should provide the full hyperparameter configuration and, ideally, release the dataset and trained checkpoints.
minor comments (4)
  1. [Section numbering] The conclusion is numbered "7" while Section 7 is "Fine-Tuning DEPLOT on the Custom Bar Chart Dataset"; subsequent sections continue to 9, so the numbering is inconsistent and should be renumbered.
  2. [§8.1 and Table II] The sentence "the fine-tuned model achieves its highest RNSS score of 98.07% and an RMS F1 score of 91.51% at epoch 6 and epoch 2, respectively" matches Table II, but Table III compares the fine-tuned model at epoch 6 for both metrics. Clarify whether the comparison model uses epoch-6 weights for RMS or the epoch-2 checkpoint, and why.
  3. [§6.1 equations] The matching matrix X in the RNSS and RMS equations is not fully specified; please state that it is obtained via a minimal-cost assignment, and give the value of the threshold parameter τ used for the Levenshtein-distance partial credit in RMS.
  4. [References] Reference [14] appears to have incomplete metadata (only a date is given); please provide a full citation. Also, reference [19] is repeated as reference [16] in spirit; check for duplicate or missing entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: fine-tuning gains are measured on a held-out synthetic split; same-generator evaluation limits external validity but does not make the result definitionally circular.

full rationale

The paper's central empirical claims are (i) fine-tuning DePlot improves RMS F1 and RNSS on table extraction (Tables I–III) and (ii) providing the fine-tuned table Ft to LLMs reduces MAPE/RMSE relative to image-only queries (Table IV). These are measured on a held-out test set of 1,000 images plus 100 QA charts that are disjoint from the 50,000-image training set described in Sections 5, 7, and 9. The reported scores compare model outputs to ground-truth tables and QA answers; they are not parameters fitted to the test set and then relabeled as predictions. The fine-tuned DePlot weights are learned on the training split, and Ft is produced at inference on unseen images, so the downstream QA improvement is not forced by construction. The image-only condition never supplies the ground-truth table to the LLM, so the comparison between image-only and table-augmented conditions is meaningful. No load-bearing step reduces to a self-citation: references [1], [3], and [4] are external prior work by other authors, and no uniqueness theorem or ansatz is imported from the present authors' prior publications. The manuscript's own limitations paragraph concedes that the model 'has not been thoroughly evaluated on more complex visualizations' and that 'domain-specific fine-tuning may limit generalization to unfamiliar chart types.' That is a genuine external-validity caveat — training and test charts come from the same randomized Matplotlib/Seaborn generator, so the measured gains partly reflect learning that generator's distribution — but it is a correctness/generalization risk, not a circularity risk under the specified criteria, because the evaluation is held-out and the claimed quantities are not equivalent to the inputs by definition. Accordingly, no circular step is present and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method adds no new theoretical entities. The central empirical quantities are the model weights, which are learned on a held-out split of the synthetic dataset, and the hand-chosen design parameters listed above.

free parameters (3)
  • Chart-type split in fine-tuning set = 50% simple, 30% stacked, 20% grouped
    Hand-chosen distribution in Section 7; it shapes the per-type performance balance and is not varied or justified against other splits.
  • RNSS tolerance threshold = 10%
    The +/-10% tolerance in RNSS (Section 6.1 A) is inherited from ChartQA and strongly affects all reported numeric similarity scores.
  • Image patch budget = 1024 patches
    Configuration choice in Section 7.2 that controls how much visual detail DePlot can encode; no ablation is provided.
assumptions (3)
  • domain assumption Synthetic charts represent real financial bar charts
    Invoked throughout Sections 5 and 8; the practical value of fine-tuning rests on this transfer, which is not validated against real chart data.
  • domain assumption Ground-truth tables exactly describe the rendered images
    True by construction because images are generated from tables (Section 5.3); it is the basis for training labels, RNSS/RMS targets, and QA ground truth.
  • domain assumption DePlot's pretrained weights are a sound starting point
    The fine-tuning initializes from the base DePlot model (Section 7.2) and relies on its general chart-to-table capability without independent verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Financial VQA in Vision Language Models using Intermediate Structured Representations." pith.science (2026). https://pith.science/paper/IL4XMSSD

@misc{pith2026250104675,
  author       = {Pith},
  title        = {Pith review of: Enhancing Financial VQA in Vision Language Models using Intermediate Structured Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IL4XMSSD}},
  note         = {Machine review of arXiv:2501.04675}
}
read the original abstract

Chart interpretation is crucial for visual data analysis, but accurately extracting information from charts poses significant challenges for automated models. This study investigates the fine-tuning of DEPLOT, a modality conversion module that translates the image of a plot or chart to a linearized table, on a custom dataset of 50,000 bar charts. The dataset comprises simple, stacked, and grouped bar charts, targeting the unique structural features of these visualizations. The finetuned DEPLOT model is evaluated against its base version using a test set of 1,000 images and two metrics: Relative Mapping Similarity (RMS), which measures categorical mapping accuracy, and Relative Number Set Similarity (RNSS), which evaluates numerical interpretation accuracy. To further explore the reasoning capabilities of large language models (LLMs), we curate an additional set of 100 bar chart images paired with question answer sets. Our findings demonstrate that providing a structured intermediate table alongside the image significantly enhances LLM reasoning performance compared to direct image queries.

Figures

Figures reproduced from arXiv: 2501.04675 by the authors.

Figure 2
Figure 2. Stacked Horizontal Bar Chart Example: “TITLE |Global Competitive Advantage <0x0A> Market Trend Analysis |Aomori Bank |Bank of Yokohama |State Street <0x0A> Costs |948 |983 |837 <0x0A> Invoices |680 |547 |532 <0x0A> Capital |709 |937 |830” [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 1
Figure 1. Simple Vertical Bar Chart 2. Stacked Horizontal Bar Chart: An example chart, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Grouped Vertical Bar Chart Each description ensures that all chart types—whether simple, stacked, or grouped—are consistently formatted, with labels and categories adjusted according to the layout. The data is structured to support effective training of the DEPLOT model by providing a rich, diverse dataset that reflects various financial scenarios and chart configurations. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: (Gt): “TITLE |Sustainable Development Goals <0x0A> Inter￾est Expense($) |Product Customization Metric <0x0A> Cash |-16 <0x0A> Audit |-8 <0x0A> Operating |30 <0x0A> Balances |-3” corresponding ground truth table (Gt) provides a structured syntax representation of the un…
Figure 5
Figure 5. Figure 5: (Gt): “TITLE |Corporate Ethics <0x0A> Product Reliability Met￾ric |Shimane Bank |Handelsbanken <0x0A> Leverage |285 |328 <0x0A> Investments |191 |418 <0x0A> Accounting |173 |127 <0x0A> Payroll |209 |294” [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (Gt): “TITLE |Market Segmentation <0x0A> Asset Turnover Ratio |Bank of Montreal |WisdomTree Japan |Royal Bank of Scotland <0x0A> Reserves |759 |870 |691 <0x0A> Equity |812 |508 |805 <0x0A> Securities |692 |627 |959” • Image Preprocessing: Each bar chart image is proces…
Figure 7
Figure 7. Figure 7: (Ft): TITLE |Strategic Employee Development <0x0A> Employee Department |Asset Turnover Ratio <0x0A> Margins |833 |Futures |328 |Statements |262 (Bt): TITLE |<0x0A> Strategic Employee Development |Struct E<0x0A> Margins |83 |Futures<0x0A> Employee Development |327 |Stat…
Figure 9
Figure 9. Figure 9: Ground Truth (Gt) :“TITLE |Venture Capital Funding <0x0A> EBT Dollars($) |Hokkoku Bank |Sumitomo Mitsui Financial Group |Tokyo Marine Asset Management |Nomura Securities |Monex Group <0x0A> Vari￾able |985 |978 |605 |840 |541 <0x0A> Projections |870 |922 |994 |767 |640 …
Figure 10
Figure 10. Figure 10: Ground Truth (Gt) :“TITLE |Strategic Marketing Campaigns <0x0A> Skill Level |Rakuten Securities |Citibank |Mizuho Bank |WisdomTree Japan |China Merchants Bank <0x0A> Income |788 |647 |627 |629 |706 <0x0A> Royalties |898 |871 |533 |707 |565 <0x0A> Capital |748 |989 |80…
Figure 8
Figure 8. Figure 8: Ground Truth (Gt) :“TITLE |Digital Marketing Trends <0x0A> Product Reliability Metric |Dividends Per Share($) <0x0A> Reserves |-35 |Margins |61 |Cash |-34 |Payables |-59” [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 9 canonical work pages

  1. [1]

    Deplot: One-shot visual language reasoning by plot-to-table translation,

    F. Liu, J. M. Eisenschlos, F. Piccinno, S. Krichene, C. Pang, K. Lee, M. Joshi, W. Chen, N. Collier, and Y . Altun, “Deplot: One-shot visual language reasoning by plot-to-table translation,” 2023. [Online]. Available: https://arxiv.org/abs/2212.10505

    arXiv 2023

  2. [2]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amo...

  3. [3]

    ChartQA: A benchmark for question answering about charts with visual and logical reasoning,

    A. Masry, X. L. Do, J. Q. Tan, S. Joty, and E. Hoque, “ChartQA: A benchmark for question answering about charts with visual and logical reasoning,” in Findings of the Association for Computational Linguistics: ACL 2022 , S. Muresan, P. Nakov, and A. Villavicencio, Eds. Dublin, Ireland: Association for Computational Linguistics, May 2022, pp. 2263–2279. [O...

    work page 2022

  4. [4]

    Chartocr: Data extraction from charts images via a deep hybrid framework,

    J. Luo, Z. Li, J. Wang, and C.-Y . Lin, “Chartocr: Data extraction from charts images via a deep hybrid framework,” in 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). The Computer Vision Foundation, January 2021. [Online]. Available: https://www.microsoft.com/en-us/research/publication/ chartocr-data-extraction-from-charts-images-vi...

    work page 2021

  5. [5]

    Figureseer: Parsing result-figures in research papers,

    N. Siegel, Z. Horvitz, R. Levin, S. Divvala, and A. Farhadi, “Figureseer: Parsing result-figures in research papers,” in Computer Vision – ECCV 2016, B. Leibe, J. Matas, N. Sebe, and M. Welling, Eds. Cham: Springer International Publishing, 2016, pp. 664–680

    work page 2016

  6. [6]

    Matcha: Enhancing visual language pretraining with math reasoning and chart derendering,

    F. Liu, F. Piccinno, S. Krichene, C. Pang, K. Lee, M. Joshi, Y . Altun, N. Collier, and J. M. Eisenschlos, “Matcha: Enhancing visual language pretraining with math reasoning and chart derendering,” 2023. [Online]. Available: https://arxiv.org/abs/2212.09662

    arXiv 2023

  7. [7]

    Pix2struct: Screenshot parsing as pretraining for visual language understanding,

    K. Lee, M. Joshi, I. Turc, H. Hu, F. Liu, J. Eisenschlos, U. Khandelwal, P. Shaw, M.-W. Chang, and K. Toutanova, “Pix2struct: Screenshot parsing as pretraining for visual language understanding,” 2023. [Online]. Available: https://arxiv.org/abs/2210.03347

    arXiv 2023

  8. [8]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin, “Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,” 2024. [Online]. Available: https://arxiv.org/abs/2409.12191

    arXiv 2024

Show all 22 references
  1. [9]

    From data quality to model quality: an exploratory study on deep learning,

    T. He, S. Yu, Z. Wang, J. Li, and Z. Chen, “From data quality to model quality: an exploratory study on deep learning,” 2019. [Online]. Available: https://arxiv.org/abs/1906.11882

    2019 arXiv

  2. [10]

    The effects of data quality on machine learning performance,

    L. Budach, M. Feuerpfeil, N. Ihde, A. Nathansen, N. Noack, H. Patzlaff, F. Naumann, and H. Harmouch, “The effects of data quality on machine learning performance,” 2022. [Online]. Available: https://arxiv.org/abs/2207.14529

    2022 arXiv

  3. [11]

    Matplotlib: A 2d graphics environment,

    J. D. Hunter, “Matplotlib: A 2d graphics environment,” Computing in Science & Engineering , vol. 9, no. 3, pp. 90–95, 2007

    2007

  4. [12]

    seaborn: statistical data visualization,

    M. Waskom, “seaborn: statistical data visualization,” The Journal of Open Source Software , vol. 6, no. 60, p. 3021, Apr. 2021

    2021

  5. [13]

    Icdar 2019 competition on scene text visual question answering,

    A. Furkan Biten, R. Tito, A. Mafla, L. Gomez, M. Rusi ˜nol, M. Mathew, C. Jawahar, E. Valveny, and D. Karatzas, “Icdar 2019 competition on scene text visual question answering,” in 2019 International Conference on Document Analysis and Recognition (ICDAR) , 2019, pp. 1563–1570

    2019

  6. [14]

    Enhancing large vision language models with self-training on image comprehension,

    Y . Deng, P. Lu, F. Yin, Z. Hu, S. Shen, J. Zou, K.-W. Chang, and W. Wang, “Enhancing large vision language models with self-training on image comprehension,” 05 2024

    2024

  7. [15]

    Fine-tuning smaller language models for question answering over financial documents,

    K. S. Phogat, S. A. Puranam, S. Dasaratha, C. Harsha, and S. Ramakrishna, “Fine-tuning smaller language models for question answering over financial documents,” 2024. [Online]. Available: https://arxiv.org/abs/2408.12337

    2024 arXiv

  8. [16]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2...

    2023 arXiv

  9. [17]

    Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models,

    M. Deitke, C. Clark, S. Lee, R. Tripathi, Y . Yang, J. S. Park, M. Salehi, N. Muennighoff, K. Lo, L. Soldaini, J. Lu, T. Anderson, E. Bransom, K. Ehsani, H. Ngo, Y . Chen, A. Patel, M. Yatskar, C. Callison-Burch, A. Head, R. Hendrix, F. Bastani, E. VanderBilt, N. Lambert, Y . ...

    2024 arXiv

  10. [18]

    Pixtral 12b,

    P. Agrawal, S. Antoniak, E. B. Hanna, B. Bout, D. Chaplot, J. Chudnovsky, D. Costa, B. D. Monicault, S. Garg, T. Gervet, S. Ghosh, A. H ´eliou, P. Jacob, A. Q. Jiang, K. Khandelwal, T. Lacroix, G. Lample, D. L. Casas, T. Lavril, T. L. Scao, A. Lo, W. Marshall, L. Martin, A. Me...

    2024 arXiv

  11. [19]

    Phi-3 technical report: A highly capable language model locally on your phone,

    M. Abdin, J. Aneja, H. Awadalla, A. Awadallah, A. A. Awan, N. Bach, A. Bahree, A. Bakhtiari, J. Bao, H. Behl et al., “Phi-3 technical report: A highly capable language model locally on your phone,” arXiv preprint arXiv:2404.14219, 2024. 12 A. COMPARATIVE INFERENCE STUDY A.1 Ba...

    2024 arXiv

  12. [150]

    You are a helpful assistant. Help me with my math homework!

    Citibank’s Capital Value is 950, while Mizuho Bank’s Capital Value is 800. Pixtral-12B-2409 [18] The value of Reserves is approximately -30, and the value of Cash is approxi- mately -35. So, the difference in value between Reserves and Cash is 5. Hokkoku Bank, Projections: 100...

    2000

  13. [800]

    The difference in value between Reserves and Cash is: -30 - (-20) = -10 Therefore, the difference in Value between Reserves and Cash is -10

    To find the difference, I’ll subtract Mizuho Bank’s Capital from Citibank’s Capital.1000 (Citibank’s Capital) - 800 (Mizuho Bank’s Capital) = 200 Molmo-7B-D-0924 [17] The Value of Reserves is -30 and the Value of Cash is -20. The difference in value between Reserves and Cash i...

  14. [2020]

    Available: https://arxiv.org/abs/2005.14165

    [Online]. Available: https://arxiv.org/abs/2005.14165

    2005 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.