Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

FinLMM-R1: Enhancing Financial Reasoning in LMM through Scalable Data and Reward Design

T0 review · 3 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 An automated report-to-QA pipeline plus a two-stage RL framework with image-selection, length, and adversarial rewards makes a 3B multimodal model more accurate than prior reasoning LMMs on financial and general benchmarks.

desk verdict A useful, well-executed extension of LMM-R1 whose main financial result needs a leakage check before you trust the numbers. read the letter →

arxiv 2506.13066 v1 pith:BHYAW3OZ submitted 2025-06-16 cs.CL

Kai Lan , Jiayong Zhu , Jiangtong Li , Dawei Cheng , Guang Chen , Changjun Jiang This is my paper
classification cs.CL
keywords financialmultimodalreasoningautomateddatapipelineimage-questionalignmentreinforcementlearningGRPOadversarialrewardlengthimageselection
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

FinLMM-R1 is an attempt to close two gaps that keep large multimodal models from reasoning well about financial charts and tables: the lack of high-quality, scalable reasoning data, and reward functions that only check final answers. The paper argues that separating question-answer generation from image-question alignment in a data pipeline solves the first gap, producing 89,378 image-question-answer triplets from 23,397 Chinese financial reports. For the second gap, it extends a two-stage rule-based RL recipe with three extra rewards: image selection, reasoning-length control, and a BERT-based adversarial reward on thinking quality. On the paper's seven benchmarks, the resulting 3B model outperforms LMM-R1 and other reasoning LMMs in both general and financial multimodal tasks, with the largest gain on its own financial test set (total accuracy 59.36% versus 53.01%). A sympathetic reader would care because the recipe is automated and domain-agnostic in principle, so the same data-and-reward loop could lift reasoning in other document-heavy fields.

What carries the argument

The machinery has two coupled parts. The Automated and Scalable Pipeline (ASP) is a separate paradigm for data construction: an LLM generates chart-relevant multiple-choice QA pairs from cleaned report text, and an LMM independently aligns each question to exactly one image, with three-fold LMM inference plus graduate-student review to filter unanswerable or ambiguous triplets. This separation is what turns misaligned report text and figures into reliable image-question-answer triplets at scale. The Thinking with Adversarial Reward in LMM (TAR-LMM) training framework is a two-stage GRPO reinforcement-learning setup: stage 1 uses text-only math problems with format and accuracy rewards; stage 2 uses multi-image contrastive samples with the original rewards plus image-selection, length, and adversarial rewards. The adversarial reward is a BERT-based discriminator, trained periodically on generated thinking traces labeled by answer correctness, that scores whether a chain of thought is genuinely predictive of the right answer; the policy model is rewarded for fooling it. The image-selection reward trains the model to pick the correct image among distractors, and the length reward encodes the paper's finding that reasoning-length and answer accuracy have a non-linear relationship.

What would settle it

Compute the maximum nearest-neighbor overlap between FinData (2016-2022) and the 2023 ASP-derived test set at the level of question templates, chart crops, or answer strings; if a large fraction of test questions are near-duplicates of training samples, the 59.36% result is explained by memorization. Conversely, hold out all reports published after 2023 and rerun the comparison: if the FinData-trained model no longer beats LMM-R1, the claimed transferable reasoning gain fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a purpose-built financial multimodal reasoning dataset plus a multi-reward reinforcement-learning framework produces a stronger reasoning LMM than existing reasoning LMMs in the 3B class. The ASP pipeline generates multiple-choice questions with a text LLM from cleaned report text, then uses a vision-language model to match each question to exactly one chart or table, discarding ambiguous pairs and double-checking doubtful ones with human annotators. The TAR-LMM training framework starts from LMM-R1's two-stage GRPO recipe and adds a multi-image setting with distractors and an image-selection reward, a piecewise length reward that penalizes both too-short and overly long thinking, and an adversarial BERT discriminator that scores whether a thinking trace leads to a correct answer. The paper reports that FinLMM-R1(S2)+FinData reaches 59.36% total accuracy on its 2023 financial test set versus 53.01% for LMM-R1, and 28.44% versus 26.59% on the external FAMMA benchmark, while also improving on the general benchmarks MathVista, MMStar, MathVision, and MathVerse.

Load-bearing premise

The 2023 financial test set and the 2016-2022 training data come from the same automated pipeline, and the paper reports no leakage or near-duplicate analysis at the level of templates, charts, or answer patterns, so part of the reported financial gain could reflect distribution overlap rather than transferable reasoning.

Editorial extensions

If this is right

  • With FinData added, the model improves by roughly four points on the paper's financial test set (59.36% versus 55.02% for stage-2 without FinData), and the improvement grows monotonically as the fraction of FinData increases from 0% to 100% in Table 4.
  • Each of the three extra rewards contributes: removing the adversarial reward hurts financial accuracy most (52.83% versus 55.02%), and removing image-selection or length rewards also lowers scores on both general and financial benchmarks in Table 3.
  • The framework transfers to a larger base model: with Qwen2.5-VL-7B, FinLMM-R1(S2)+FinData reaches 65.59% on the financial test set and 51.14% average on general multimodal tasks, beating the 7B LMM-R1 at 61.72% and 48.79%.
  • Financial data does not hurt general reasoning: FinLMM-R1(S2)+FinData performs essentially at parity with FinLMM-R1(S2) on general multimodal benchmarks (47.45% versus 47.43% average), so adding financial data is safe for general-domain performance.
  • Stage-2 multimodal training improves text-only GPQA from 36.82% for FinLMM-R1(S1) to 41.52% for FinLMM-R1(S2), which the paper interprets as evidence of complementarity between text and multimodal reasoning.

Reading between the lines

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

  • Beyond the paper, the ASP two-model protocol should transfer to other document-heavy domains such as legal, medical, or scientific PDFs with minimal change; a testable extension is to run ASP on legal or medical reports and check whether a 3B model gains on existing domain benchmarks.
  • The adversarial BERT discriminator is the reward most tied to reasoning depth, but the paper's own ablation shows it contributes least on hard visual math sets such as MathVision and MathVerse, which the authors attribute to the discriminator struggling with hard thinking; a stronger discriminator, such as an LMM rather than BERT, is the natural next lever for further gains.
  • Because the length reward encodes a specific optimal window of roughly 450 tokens, the approach implicitly bets that concise, information-dense thinking beats longer chains; one could test this directly by varying the optimal window and watching accuracy on hard financial questions, which the paper does not report.
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 / 6 minor

Summary. The paper proposes FinLMM-R1, a framework that combines an Automated and Scalable Pipeline (ASP) for constructing financial multimodal question-answer data from Chinese listed-company reports with TAR-LMM, a two-stage GRPO-based training scheme that extends LMM-R1 by adding image-selection, reasoning-length, and adversarial BERT-based rewards. The authors build 89,378 IQA triplets from 23,397 reports, train Qwen2.5-VL-3B and 7B models, and report gains over LMM-R1 on one text benchmark, four general multimodal benchmarks, a 2023 ASP-derived financial test set, and the external FAMMA benchmark.

Significance. If the claims hold, the paper makes a useful empirical contribution: a scalable recipe for financial multimodal reasoning data and evidence that multi-reward RL improves small LMM reasoning. The paper includes useful strengths: ablations of each reward component, a 7B variant, human QA checks of the constructed data, significance tests on MMStar and MathVista, and an external financial benchmark (FAMMA) that partially validates the approach. The main caveat is that the headline financial comparison is measured on a test set generated by the same pipeline as the training data, and no leakage or near-duplicate analysis is provided, so the magnitude of the financial-domain claim is uncertain pending such analysis.

major comments (3)
  1. [Sec. 5.2.2, Sec. B.3, Sec. A.1] The headline financial result in Table 2 (FinLMM-R1(S2)+FinData 59.36 vs LMM-R1 53.01 on the 'Ours' column) is measured on a 2023 test set produced by the same ASP pipeline that generated the FinData training split described in Sec. A.1, with only a temporal separation between 2016-2022 and 2023. The paper reports no chart-level, template-level, or answer-pattern overlap analysis between the FinData training sample and this test set, even though financial reports are highly templated and the same DeepSeek-V3 question-generation prompts and QwenVL alignment pipeline are used for both. This is load-bearing because the 6.35-point gap on the ASP-derived set is much larger than the 0.98-point gap on the external FAMMA benchmark (Table 2, FAMMA Total: 28.44 vs 27.46), and Table 4's monotonically increasing FinData scaling curve may partly reflect distribution overlap rather than transferable reasoning. Please add a quantitative near-duplicate/leakage study (image similarity, template and answer-pattern analysis, document-level split) and either re-report the financial results on a held-out external benchmark or on a test set constructed by a distinct pipeline.
  2. [Sec. H / Sec. 5.2.2] The significance test in Sec. H is run only on MMStar and MathVista; no significance test or confidence interval is provided for the financial benchmarks in Table 2, which are the central domain claim of the paper. The stage-2 vs LMM-R1 differences on the 'Ours' column (e.g., AR 53.52 vs 46.13, Tot. 59.36 vs 53.01) are reported as point estimates without variance, so the stability of the 6.35-point headline gain is unknown. Please run the same seeded multi-run protocol on the ASP-derived financial test set and on FAMMA, and report p-values or CIs for all reported benchmark categories.
  3. [Sec. 5.1 / Tables 1-2] Tables 1 and 2 copy LMM-R1 results from [1] while evaluating the proposed models with the authors' own prompt, scoring, and inference pipeline; this is not a controlled comparison unless the evaluation protocol is identical. The paper already executes LMM-R1 under its own protocol for the significance test in Sec. H, so those re-computed numbers should be used in the main tables, or the paper should explicitly document that the copied results were generated with the same prompts, extraction rules, and seeds.
minor comments (6)
  1. [Sec. 5.3] The heading 'Albation Study' contains a typo; it should be 'Ablation Study'.
  2. [Sec. 5.1] The phrase 'financial multimodal reasinng benchmarks' contains a typo; it should be 'reasoning benchmarks'.
  3. [Figure 1 / Figure 5] Figure 1's caption contains 'illstration', and the sub-figure labels are hard to read; please provide a higher-resolution figure and expand the caption text.
  4. [Appendix C.3 / Figure 5] Figure 5 in the provided manuscript contains garbled text tokens (repeated '/uni000000...' sequences) that appear to be PDF-extraction artifacts; please regenerate the figure with readable axis labels.
  5. [Eq. (3) / Table 6] The five weights in Eq. (3) are not explicitly mapped to the hyperparameters in Table 6; please state explicitly that λ0..λ4 correspond to format, accuracy, BERT/adversarial, image-selection, and length weights with values 0.1, 0.5, 0.15, 0.2, and 0.05, respectively.
  6. [Sec. 2.3 / Appendix B.3] FAMMA is cited as [47] in Sec. 2.3 and as [61] in Appendix B.3; please unify the citation and verify the reference entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity: the central training claims are anchored by external benchmarks, and the ASP-derived financial evaluation is a temporal train/test split rather than a definitional reduction.

full rationale

The paper's main derivation chain is empirical rather than definitional: ASP builds IQA triplets, FinData is a random 2016-2022 subset, and the model is evaluated on a 2023 ASP-derived test set plus external benchmarks (GPQA, MathVision, MathVerse, MMStar, MathVista, FAMMA). No equation reduces a predicted quantity to a fitted input, and no parameter is renamed as a prediction. The TAR-LMM rewards are optimization signals, not fitted constants that later appear as results. LMM-R1 is cited as prior work, but the authors of FinLMM-R1 are not the authors of LMM-R1, so there is no load-bearing self-citation chain. The one legitimate validity concern is that FinData and the 2023 financial test set come from the same ASP pipeline and the paper reports no chart/template/answer-pattern overlap analysis; that is a correctness and leakage-risk issue, not a circularity-by-construction issue. The external FAMMA and general benchmarks provide independent evidence for the training framework, and the financial result, while weaker under external evaluation, is not definitionally forced by the paper's own setup. Accordingly, no circular step can be exhibited with the specificity required by the review rules.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central empirical claims depend on multiple hand-set reward weights and length thresholds, and on the premise that ASP-generated triplets are visually answerable. The adversarial discriminator is an introduced component with no independent validation.

free parameters (5)
  • Stage 2 reward weights = format=0.1, accuracy=0.5, length=0.05, adversarial=0.15, image_selection=0.2
    Hand chosen in Table 6; no sensitivity analysis is reported.
  • Length reward thresholds = Lmin=300, Lopt=450, Lmax=600 tokens; rmin, rpen, gamma unspecified
    The piecewise length reward in Section 4.3 depends on these hand-set values.
  • BERT discriminator retraining interval = N=10 steps
    The discriminator update frequency is chosen without ablations.
  • Data filtering thresholds = PDFs with image coverage below 70% discarded; reports with 2 to 10 valid images retained
    Sections 3.2 and 3.4 use hand-set thresholds that shape the dataset composition.
  • Stage 1 reward weight = lambda=0.5 for accuracy and format
    Section 4.2 sets the balance without a reported sweep.
assumptions (5)
  • standard math GRPO with group-normalized rewards provides a valid policy optimization signal for LMMs.
    Section 4.1 adopts GRPO from DeepSeekMath; this is a standard RL result with no new proof.
  • domain assumption The BERT discriminator's confidence score is a meaningful proxy for reasoning quality.
    Section 4.3 relies on this; if the discriminator exploits superficial cues, the adversarial reward could reward style over substance.
  • domain assumption The ASP-generated questions are answerable from the paired image alone.
    Sections 3.3 and 3.4 rely on QwenVL alignment plus human filtering to guarantee visual answerability.
  • domain assumption The 2023 ASP-generated test set is a fair measure of financial multimodal reasoning.
    Section 5.1 uses this set for the main financial evaluation; no cross-split leakage analysis is provided.
  • domain assumption Text extraction and chart reference marks such as 'Chart x' and 'Figure x' reliably identify the correct image.
    Sections 3.1 and 3.2 assume PDF layout parsing preserves image-text correspondence.
invented entities (1)
  • Adversarial BERT discriminator
    purpose: Provides a reward for thinking-process quality by predicting answer correctness from the content inside think tags.
    It is a trained reward model without a released checkpoint or standalone evaluation; its reliability is only indirectly measured through downstream accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FinLMM-R1: Enhancing Financial Reasoning in LMM through Scalable Data and Reward Design." pith.science (2026). https://pith.science/paper/BHYAW3OZ

@misc{pith2026250613066,
  author       = {Pith},
  title        = {Pith review of: FinLMM-R1: Enhancing Financial Reasoning in LMM through Scalable Data and Reward Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHYAW3OZ}},
  note         = {Machine review of arXiv:2506.13066}
}
read the original abstract

Large Multimodal Models (LMMs) demonstrate significant cross-modal reasoning capabilities. However, financial applications face challenges due to the lack of high-quality multimodal reasoning datasets and the inefficiency of existing training paradigms for reasoning enhancement. To address these issues, we propose an integrated framework, FinLMM-R1, combining an automated and scalable pipeline for data construction with enhanced training strategies to improve the multimodal reasoning of LMM. The Automated and Scalable Pipeline (ASP) resolves textual-visual misalignment in financial reports through a separate paradigm of question-answer generation and image-question alignment, ensuring data integrity and extraction efficiency. Through ASP, we collect 89,378 aligned image-question pairs from 23,397 financial reports, covering tasks such as arithmetic reasoning, statistics reasoning, financial explanation, and financial knowledge. Moreover, we introduce the Thinking with Adversarial Reward in LMM (TAR-LMM), extending the prior two-stage training framework [1] with additional reward mechanisms. In the first stage, we focus on text-only tasks with format and accuracy rewards to guide the model in generating well-structured thinking contents. In the second stage, we construct multi-image contrastive samples with additional reward components including image selection, thinking content length, and adversarial reward to jointly optimize the LMM across visual perception, reasoning efficiency, and logical coherence. Extensive experiments on 7 benchmarks show ASP-derived dataset and training framework significantly improve answer accuracy and reasoning depth over existing reasoning LMMs in both general and financial multimodal contexts.

Figures

Figures reproduced from arXiv: 2506.13066 by the authors.

Figure 1
Figure 1. The illstration of (a) Automated and Scalable Pipeline for multimodal financial data [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The illustration shows the ASP for multimodal financial data collection. We first extract [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The illustration shows the TAR-LMM training framework. In the first stage, we focus on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative example showcasing how our FinLMM-R1 framework achieves successful [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The Training Curve of FinLMM-R1 in Stage-2. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: The statistics of FinData in (a) question type, (b) image quantity distribution by document, [PITH_FULL_IMAGE:figures/full_fig_p018_6.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. Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A new benchmark shows LLMs' financial calculation accuracy collapses without explicit formulas and degrades further when they must generate multi-metric tables.

Reference graph

Works this paper leans on

90 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    Zhang, M

    Peng, Y ., G. Zhang, M. Zhang, et al. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl.arXiv preprint arXiv:2503.07536, 2025

    arXiv 2025

  2. [2]

    Bhatia, G., E. M. B. Nagoudi, H. Cavusoglu, et al. Fintral: A family of gpt-4 level multimodal financial large language models. InACL Findings. 2024

    2024

  3. [3]

    Wang, Z., Y . Li, J. Wu, et al. Finvis-gpt: A multimodal large language model for financial chart analysis. InFinLLM@IJCAI. 2023

    2023

  4. [4]

    Radford, A., J. W. Kim, C. Hallacy, et al. Learning transferable visual models from natural language supervision. InICML, pages 8748–8763. 2021

    2021

  5. [5]

    Li, J., D. Li, S. Savarese, et al. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InICML, pages 19730–19742. 2023

    2023

  6. [6]

    Liu, H., C. Li, Q. Wu, et al. Visual instruction tuning. InNeurIPS, pages 34892–34916. 2023

    2023

  7. [7]

    Brown, T., B. Mann, N. Ryder, et al. Language models are few-shot learners. InNeurIPS, vol. 33, pages 1877–1901. 2020

    1901

  8. [8]

    Wei, J., X. Wang, D. Schuurmans, et al. Chain-of-thought prompting elicits reasoning in large language models. InNeurIPS, pages 24824–24837. 2022

    2022

Show all 90 references
  1. [9]

    Donahue, P

    Alayrac, J.-B., J. Donahue, P. Luc, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 23716–23736. 2022

    2022

  2. [10]

    Li, W., H. Fan, Y . Wong, et al. Improving context understanding in multimodal large language models via multimodal composition learning. InICML. 2024

    2024

  3. [11]

    Mitra, L

    Huang, B., C. Mitra, L. Karlinsky, et al. Multimodal task vectors enable many-shot multimodal in-context learning. InNeurIPS, pages 22124–22153. 2024

    2024

  4. [12]

    Shao, H., S. Qian, H. Xiao, et al. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. InNeurIPS, pages 8612–8642. 2024

    2024

  5. [13]

    Chen, Z., Q. Zhou, Y . Shen, et al. Visual chain-of-thought prompting for knowledge-based visual reasoning. InAAAI, pages 1254–1262. 2024

    2024

  6. [14]

    Zhao, H., Z. Cai, S. Si, et al. Mmicl: Empowering vision-language model with multi-modal in-context learning ability. InICLR. 2024

    2024

  7. [15]

    Zheng, G., B. Yang, J. Tang, et al. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. InNeurIPS, pages 5168–5191. 2023

    2023

  8. [16]

    Lu, P., B. Peng, H. Cheng, et al. Chameleon: Plug-and-play compositional reasoning with large language models. InNeurIPS, pages 43447–43478. 2023

    2023

  9. [17]

    Yao, S., J. Zhao, D. Yu, et al. React: Synergizing reasoning and acting in language models. InICLR. 2024

    2024

  10. [18]

    Tang, X., A. Zou, Z. Zhang, et al. Medagents: Large language models as collaborators for zero-shot medical reasoning. InACL Findings. 2024

    2024

  11. [19]

    Yang, Z., L. Li, J. Wang, et al. Mm-react: Prompting chatgpt for multimodal reasoning and action.arXiv preprint arXiv:2303.11381, 2023

    2023 arXiv

  12. [20]

    Liang, Y

    Qin, Y ., S. Liang, Y . Ye, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis.arXiv preprint arXiv:2307.16789, 2023. 10

    2023 arXiv

  13. [21]

    Liu, X., R. Li, W. Ji, et al. Towards robust multi-modal reasoning via model selection. InICLR. 2024

    2024

  14. [22]

    Guo, D., D. Yang, H. Zhang, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  15. [23]

    Zhang, X., J. Wang, Z. Cheng, et al. Srpo: A cross-domain implementation of large-scale reinforcement learning on llm.arXiv preprint arXiv:2504.14286, 2025

    2025 arXiv

  16. [24]

    Wei, Y ., Y . Peng, X. Wang, et al. Skywork r1v2: Multimodal hybrid reinforcement learning for reasoning. arXiv preprint arXiv:2504.16656, 2025

    2025 arXiv

  17. [25]

    Wang, W., Z. Gao, L. Chen, et al. Visualprm: An effective process reward model for multimodal reasoning. arXiv preprint arXiv:2503.10291, 2025

    2025 arXiv

  18. [26]

    Hu, C.-W

    Xing, Z., X. Hu, C.-W. Fu, et al. Echoink-r1: Exploring audio-visual reasoning in multimodal llms via reinforcement learning.arXiv preprint arXiv:2505.04623, 2025

    2025 arXiv

  19. [27]

    Yang, Y ., X. He, H. Pan, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  20. [28]

    Bai, J., S. Bai, Y . Chu, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  21. [29]

    Dubey, A

    Grattafiori, A., A. Dubey, A. Jauhri, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  22. [30]

    Team, G., R. Anil, S. Borgeaud, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  23. [31]

    Irsoy, S

    Wu, S., O. Irsoy, S. Lu, et al. Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564, 2023

    2023 arXiv

  24. [32]

    Wang, N., H. Yang, C. D. Wang. Fingpt: Instruction tuning benchmark for open-source large language models in financial datasets.arXiv preprint arXiv:2310.04793, 2023

    2023 arXiv

  25. [33]

    Zhang, X., Q. Yang. Xuanyuan 2.0: A large chinese financial chat model with hundreds of billions parameters. InCIKM, pages 4435–4439. 2023

    2023

  26. [34]

    Li, J., Y . Bian, G. Wang, et al. Cfgpt: Chinese financial assistant with large language model.arXiv preprint arXiv:2309.10654, 2023

    2023 arXiv

  27. [35]

    Chen, W., Q. Wang, Z. Long, et al. Disc-finllm: A chinese financial large language model based on multiple experts fine-tuning.arXiv preprint arXiv:2310.15205, 2023

    2023 arXiv

  28. [36]

    Xie, Q., W. Han, X. Zhang, et al. Pixiu: A large language model, instruction data and evaluation benchmark for finance. InNeurIPS. 2023

    2023

  29. [37]

    Liu, Z., X. Guo, F. Lou, et al. Fin-r1: A large language model for financial reasoning through reinforcement learning.arXiv preprint arXiv:2503.16252, 2025

    2025

  30. [38]

    Adler, S

    Achiam, J., S. Adler, S. Agarwal, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  31. [39]

    Shah, R. S., K. Chawla, D. Eidnani, et al. When flue meets flang: Benchmarks and large pre-trained language model for financial domain. InEMNLP. 2022

    2022

  32. [40]

    Lu, D., H. Wu, J. Liang, et al. Bbt-fin: Comprehensive construction of chinese financial domain pre-trained language model, corpus and benchmark.arXiv preprint arXiv:2302.09432, 2023

    2023 arXiv

  33. [41]

    Zhang, L., W. Cai, Z. Liu, et al. Fineval: A chinese financial domain knowledge evaluation benchmark for large language models.arXiv preprint arXiv:2308.09975, 2023

    2023 arXiv

  34. [42]

    Li, H., Y . Cao, Y . Yu, et al. Investorbench: A benchmark for financial decision-making tasks with llm-based agent.arXiv preprint arXiv:2412.18174, 2024

    2024 arXiv

  35. [43]

    Chen, Z., W. Chen, C. Smiley, et al. Finqa: A dataset of numerical reasoning over financial data. In EMNLP. 2021

    2021

  36. [44]

    Xie, Q., W. Han, Z. Chen, et al. Finben: A holistic financial benchmark for large language models. In NeurIPS, pages 95716–95743. 2024

    2024

  37. [45]

    Xu, L., L. Zhu, Y . Wu, et al. Superclue-fin: Graded fine-grained analysis of chinese llms on diverse financial tasks and applications.arXiv preprint arXiv:2404.19063, 2024

    2024 arXiv

  38. [46]

    Openfindata: The open-source financial evaluation dataset for large language models, 2023

    OpenCompass Project. Openfindata: The open-source financial evaluation dataset for large language models, 2023. Available athttps://github.com/open-compass/OpenFinData

    2023

  39. [47]

    Xue, S., T. Chen, F. Zhou, et al. FAMMA: A benchmark for financial multilingual multimodal question answering, 2025

    2025

  40. [48]

    Gan, Z., Y . Lu, D. Zhang, et al. Mme-finance: A multimodal finance benchmark for expert-level understanding and reasoning.arXiv preprint arXiv:2411.03314, 2024. 11

    2024 arXiv

  41. [49]

    Yizhao-findataset, 2024

    CMB AILab. Yizhao-findataset, 2024. Available at https://www.modelscope.cn/datasets/CMB_ AILab/YiZhao-FinDataSet

    2024

  42. [50]

    Liu, A., B. Feng, B. Xue, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  43. [51]

    Bai, S., K. Chen, X. Liu, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  44. [52]

    Shao, Z., P. Wang, Q. Zhu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  45. [53]

    Yang, W., S. Ma, Y . Lin, et al. Towards thinking-optimal scaling of test-time compute for llm reasoning. arXiv preprint arXiv:2502.18080, 2025

    2025

  46. [54]

    Goodfellow, I. J., J. Pouget-Abadie, M. Mirza, et al. Generative adversarial nets. InNeurIPS. 2014

    2014

  47. [55]

    Luo, M., S. Tan, J. Wong, et al. Deepscaler: Surpassing o1-preview with a 1.5 b model by scaling rl. Notion Blog, 2025

    2025

  48. [56]

    Wang, K., J. Pan, W. Shi, et al. Measuring multimodal mathematical reasoning with math-vision dataset. InNeurIPS, pages 95095–95169. 2024

    2024

  49. [57]

    Jiang, Y

    Zhang, R., D. Jiang, Y . Zhang, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? InECCV, pages 169–186. 2024

    2024

  50. [58]

    Chen, L., J. Li, X. Dong, et al. Are we on the right way for evaluating large vision-language models? In NeurIPS. 2024

    2024

  51. [59]

    Bansal, T

    Lu, P., H. Bansal, T. Xia, et al. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. InICLR. 2024

    2024

  52. [60]

    Rein, D., B. L. Hou, A. C. Stickland, et al. Gpqa: A graduate-level google-proof q&a benchmark. In COLM. 2024

    2024

  53. [61]

    Xue, S., X. Li, F. Zhou, et al. Famma: A benchmark for financial domain multilingual multimodal question answering.arXiv preprint arXiv:2410.04526, 2025

    2025 arXiv

  54. [62]

    Chen, Z., W. Wang, Y . Cao, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024

    2024 arXiv

  55. [63]

    Yao, Y ., T. Yu, A. Zhang, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    2024 arXiv

  56. [64]

    Lu, P., L. Qiu, J. Chen, et al. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning.arXiv preprint arXiv:2110.13214, 2021

    2021 arXiv

  57. [65]

    Kembhavi, A., M. Seo, D. Schwenk, et al. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. InCVPR, pages 4999–5007. 2017

    2017

  58. [66]

    Zhang, X., C. Wu, Z. Zhao, et al. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023

    2023 arXiv

  59. [67]

    Bagal, R

    Mathew, M., V . Bagal, R. Tito, et al. Infographicvqa. InWACV, pages 1697–1706. 2022

    2022

  60. [68]

    Qiu, K.-W

    Lu, P., L. Qiu, K.-W. Chang, et al. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning.arXiv preprint arXiv:2209.14610, 2022

    2022 arXiv

  61. [69]

    Natarajan, M

    Singh, A., V . Natarajan, M. Shah, et al. Towards vqa models that can read. InCVPR, pages 8317–8326. 2019

    2019

  62. [70]

    Salvato, E

    Kembhavi, A., M. Salvato, E. Kolve, et al. A diagram is worth a dozen images. InECCV, pages 235–251. 2016

    2016

  63. [71]

    Goyal, Y ., T. Khot, D. Summers-Stay, et al. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. InCVPR, pages 6904–6913. 2017

    2017

  64. [72]

    Kahou, S. E., V . Michalski, A. Atkinson, et al. Figureqa: An annotated figure dataset for visual reasoning. InICLR workshop. 2018

    2018

  65. [73]

    Masry, A., D. X. Long, J. Q. Tan, et al. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. InACL. 2022

    2022

  66. [74]

    Mishra, T

    Lu, P., S. Mishra, T. Xia, et al. Learn to explain: Multimodal reasoning via thought chains for science question answering. InNeurIPS, pages 2507–2521. 2022

    2022

  67. [75]

    Ganguly, M

    Methani, N., P. Ganguly, M. M. Khapra, et al. Plotqa: Reasoning over scientific plots. InWACV, pages 1527–1536. 2020

    2020

  68. [76]

    Cao, J., J. Xiao. An augmented benchmark dataset for geometric question answering through dual parallel text encoding. InCOLING, pages 1511–1520. 2022

    2022

  69. [77]

    Li, Z., X. Wang, E. Stengel-Eskin, et al. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. InCVPR, pages 14963–14973. 2023. 12

    2023

  70. [78]

    Price, S

    Kafle, K., B. Price, S. Cohen, et al. Dvqa: Understanding data visualizations via question answering. In CVPR, pages 5648–5656. 2018

    2018

  71. [79]

    Agrawal, J

    Antol, S., A. Agrawal, J. Lu, et al. Vqa: Visual question answering. InICCV, pages 2425–2433. 2015

    2015

  72. [80]

    Lu, P., R. Gong, S. Jiang, et al. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. InACL. 2021

    2021

  73. [81]

    Palzer, J

    Chang, S., D. Palzer, J. Li, et al. Mapqa: A dataset for question answering on choropleth maps. InNeurIPS. 2022

    2022

  74. [82]

    Chen, J., T. Li, J. Qin, et al. Unigeo: Unifying geometry logical reasoning via reformulating mathematical expression. InEMNLP. 2022

    2022

  75. [83]

    Gurari, D., Q. Li, A. J. Stangl, et al. Vizwiz grand challenge: Answering visual questions from blind people. InCVPR, pages 3608–3617. 2018

    2018

  76. [84]

    Hajishirzi, A

    Seo, M., H. Hajishirzi, A. Farhadi, et al. Solving geometry problems: Combining text and diagram interpretation. InEMNLP, pages 1466–1476. 2015

    2015

  77. [85]

    Lindström, A. D., S. S. Abraham. Clevr-math: A dataset for compositional language, visual and mathemat- ical reasoning.arXiv preprint arXiv:2208.05358, 2022

    2022 arXiv

  78. [86]

    Lau, J. J., S. Gayen, A. Ben Abacha, et al. A dataset of clinically generated visual questions and answers about radiology images.Scientific data, 5(1):1–10, 2018

    2018

  79. [87]

    Junting, W

    Yaowei, Z., L. Junting, W. Shenzhi, et al. Easyr1: An efficient, scalable, multi-modality rl training framework.https://github.com/hiyouga/EasyR1, 2025

    2025

  80. [88]

    Zhang, Z

    Sheng, G., C. Zhang, Z. Ye, et al. Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv: 2409.19256, 2024

    2024 arXiv

  81. [89]

    Gurevych

    Reimers, N., I. Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. InEMNLP- IJCNLP, pages 3982–3992. 2019

    2019

  82. [90]

    A. content

    Devlin, J., M.-W. Chang, K. Lee, et al. Bert: Pre-training of deep bidirectional transformers for language understanding. InNAACL, pages 4171–4186. 2019. 13 The appendix includes additional materials to supplement the main submission. Section A details the training data and im...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.