Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

TaskGalaxy: Scaling Multi-modal Instruction Fine-tuning with Tens of Thousands Vision Task Types

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Fine-tuning with a dataset of 19,227 hierarchical task types and 413,648 samples improves LLaVA-v1.5 and InternVL-Chat-v1.0 on all 16 benchmarks, with average gains of 3.0 to 4.5 points.

desk verdict A genuinely large task-diversity dataset with an automated pipeline and consistent gains, but the missing train/eval image-overlap check leaves the central diversity claim less secure than the paper's 'proves' language suggests. read the letter →

arxiv 2502.09925 v1 pith:P2YD3RDR submitted 2025-02-14 cs.CV cs.AI

classification cs.CVcs.AI
keywords taskdiversityinstructionfine-tuningmultimodallargelanguagemodelshierarchicaltypesdatasetgenerationpipelineGPT-4oCLIPmatchingvisualquestionanswering
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 sets out to show that the diversity of task types in multimodal instruction fine-tuning data is a major lever on model generalization, distinct from raw sample quantity. It introduces TaskGalaxy, a dataset of 413,648 question-answer pairs organized under 19,227 hierarchical task types, built by an almost fully automated pipeline: GPT-4o expands a small set of hand-written seed tasks into a three-level taxonomy, CLIP matches the resulting labels to open-source images, GPT-4o writes questions and answers, and three open-source models score every sample for relevance. Adding this dataset to the original fine-tuning data of LLaVA-v1.5 and InternVL-Chat-v1.0 improved all 16 evaluated benchmarks, by 3.0 to 4.5 points on average across the 15 non-MME benchmarks, with a 68-point gain for LLaVA-v1.5-13B on MME. An ablation holding sample count fixed at 100,000 shows performance rising as the taxonomy grows from 2,000 to 19,227 task types, which the paper takes as direct evidence that task-type coverage is the active ingredient. If that interpretation holds, the recipe for better multimodal fine-tuning is to spend labeling and generation effort on covering more task types rather than merely adding more samples.

What carries the argument

The load-bearing object is the hierarchical task-type taxonomy and the automated pipeline that produces it. A handful of manually defined seed tasks (OCR, image description, logical reasoning, and others) is expanded by GPT-4o into three levels connected by '~' notation, yielding 19,227 distinct task labels, with prompts that explicitly ask for categories not overlapping existing ones. This taxonomy is the independent variable of the paper's main ablation: holding the training set at 100,000 samples while varying the task-type count from 2,000 to 19,227 raises average benchmark performance steadily. The matching stage uses the CLIP cosine similarity $s_j = I(x_i) \cdot T(t_j)$ between the image embedding and the task-label embedding to select the top $k = 10$ candidate task types per image, followed by GPT-4o pruning, GPT-4o question-answer generation, and a three-model referee vote that keeps samples scoring at least 2 out of 3, which is the quality-control gate that lets the pipeline run without human labeling.

What would settle it

Compute image-level overlap between the TaskGalaxy training pool and each evaluation benchmark, then re-run the fine-tuning with every overlapping image removed. If the reported gains (3.0 to 4.5 points on average, and the monotone improvement from 2,000 to 19,227 task types) persist after deduplication, the task-diversity claim stands; if they shrink or vanish, training-set contamination was a carrier of the effect.

Watch

Extended reading notes

Core claim

The paper's central claim is that the range of task types in supervised fine-tuning data is a bottleneck for vision-language models, and that a dataset spanning roughly 19,000 hierarchical task types removes that bottleneck. TaskGalaxy organizes 413,648 image-question-answer samples under 19,227 task types produced by a five-step pipeline: GPT-4o expands dozens of hand-written seed tasks into a three-level taxonomy (115 first-level, 2,796 second-level, and 14,370 third-level types); CLIP scores each image against each task label by cosine similarity and keeps the top 10; GPT-4o prunes labels that do not fit the image content; GPT-4o writes a question-answer pair for each surviving label; and three open-source multimodal models independently score each image-task-question triple, keeping samples that score at least 2 out of 3. Fine-tuning LLaVA-v1.5 (7B and 13B) and InternVL-Chat-v1.0 (7B and 13B) with TaskGalaxy added to the original fine-tuning data improved every one of the 16 benchmarks, with average gains of 4.5, 3.83, 3.0, and 3.64 points on the 15 non-MME benchmarks and a 68-point gain for LLaVA-v1.5-13B on MME. The paper states that these results prove that expanding the diversity of visual task types, together with the associated high-quality question-answer pairs, significantly enhances the generalization capabilities of multimodal models.

Load-bearing premise

The paper credits the benchmark gains to task-type diversity, but it never verifies that its training images are absent from the evaluation sets, and several training sources (COCO images used by POPE and other benchmarks, TextVQA training images, and DocVQA images) overlap with evaluation imagery, so part of the reported improvement could come from the model having seen evaluation images during fine-tuning rather than from task diversity.

Editorial extensions

If this is right

  • Task-type coverage becomes a primary design axis for fine-tuning datasets: the paper's ablation shows that, at a fixed 100,000 samples, expanding from 2,000 to 19,227 task types raises average benchmark performance.
  • Manual task labeling is no longer the limiting cost: with a small seed taxonomy and the expand-match-generate-score pipeline, a dataset with hundreds of times more task types than previous human-labeled efforts can be built almost automatically.
  • The diversity benefit transfers across model families and scales: all four tested LLaVA-v1.5 and InternVL-Chat-v1.0 variants improve on all 16 benchmarks, and the appendix reports 14-of-16 benchmark gains on the newer InternVL-Chat-V2.0-8B.
  • Task diversity helps more than reasoning-heavy tasks: the paper reports gains on hallucination and perception benchmarks (POPE, HallusionBench, Q-Bench, MME) as well as on chart, diagram, and math benchmarks.
  • The dataset extends without re-annotation: adding new images and task types can be done through the same pipeline, so coverage can grow with new domains.

Reading between the lines

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

  • If task coverage is the carrier, a data-scaling law in 'task space' is plausible: for a fixed fine-tuning budget, spreading samples across more distinct task labels should beat concentrating them in fewer labels, a hypothesis that could be tested by varying the taxonomy while holding images and answer content fixed.
  • Because the taxonomy's long tail includes refusal and safety-related task types, the same pipeline could encode safety behaviors into models through task labeling; the paper does not analyze this effect directly.
  • A controlling experiment the paper does not run would attach task-type labels at random to a fixed set of question-answer pairs; if the diversity benefit survives random labeling, the active ingredient is the answer-content distribution rather than the taxonomy structure.
  • The comparison against equally sized synthetic datasets (ShareGPT-4V, ALLaVA, Cambrian-1) leaves open whether the gains come from task-type count or from characteristics of GPT-4o-generated answers, so an equal-size human-annotated dataset at the same task-type count would separate the two factors.
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

5 major / 5 minor

Summary. TaskGalaxy is a multimodal instruction fine-tuning dataset containing 19,227 hierarchical task types and 413,648 question-answer samples. The construction pipeline expands a small set of manually defined task types using GPT-4o, matches task types to images with CLIP and GPT-4o, generates QA pairs with GPT-4o, and filters them with three open-source vision-language models acting as referees. The paper evaluates the dataset by fine-tuning LLaVA-v1.5-7B/13B and InternVL-Chat-v1.0-7B/13B on the original SFT mix plus TaskGalaxy, reporting improvements on 16 benchmarks relative to the original-data baselines. Ablations vary the number of task types and the sample budget, and the appendix includes an equal-sample-size comparison with other instruction-tuning datasets and an experiment on InternVL-Chat-V2.0-8B.

Significance. If the causal interpretation is accepted, the paper makes a strong practical contribution: TaskGalaxy is two orders of magnitude larger in task-type coverage than prior instruction-tuning datasets, the generation pipeline is largely automated, the dataset is publicly released, and the gains are shown across multiple base model families. The equal-sample-size comparison in Appendix A.7 and the extension to a newer model in Appendix A.8 are useful controls, as is the chain-of-thought ablation. However, the central attribution of the benchmark gains to task diversity is weakened by the absence of a train/eval image-overlap analysis and by the reliance on an uncontrolled primary comparison; both are fixable empirically rather than fundamental flaws in the method.

major comments (5)
  1. [§2.2, Table A-1; §3.1] The TaskGalaxy image pool shares image families with the evaluation benchmarks, and the paper does not check or control for this. Table A-1 explicitly lists 'coco train2017' (118,326 images), 'text vqa train images' (3,485), 'ocr vqa images' (12,614), and 'share textvqa images' (55), while Table 1 lists Visual Genome, ShareGPT4V, and MathV360K, which are known to contain COCO, SA, SBU, and DocVQA images. POPE is built from COCO subsets and TextVQA uses TextVQA images, so training images may overlap with evaluation images in two of the sixteen benchmarks. The paper reports no exact-duplicate or near-duplicate overlap statistics and no evaluation restricted to images disjoint from the training pool. Because the central claim is that task diversity, not image familiarity, drives the improvements, this gap is load-bearing. A concrete fix is to compute overlap between TaskGalaxy images and each evaluation benchmark's image set, and to re-report the affected benchmarks on the disjoint subset; ideally this should be done for all benchmarks.
  2. [§3.2, Table 2; §A.7] The primary comparison in Table 2 adds all 413k TaskGalaxy samples on top of the original SFT mix, so the average gain conflates increased data quantity with increased task diversity. Appendix A.7 is the correct control and TaskGalaxy still wins at matched sample size on most benchmarks, but the main text and abstract continue to treat the uncontrolled comparison as the primary evidence, and the sentence in Section 3.2 that this 'proves' the diversity claim is stronger than the experiment supports. The authors should either move the matched-sample-size comparison into the main paper as the headline result or add an explicit data-quantity-matched ablation; at minimum the wording should be softened to 'consistent with' rather than 'proves.'
  3. [§2.2, Referee Screening; §3.1] Two of the three referee models used to filter TaskGalaxy samples (InternVL-Chat-v1.5 and InternVL2-26B) belong to the InternVL family, and one of the evaluated models is InternVL-Chat-v1.0. Consequently, the selected samples may be preferentially well-matched to what InternVL-family models can score and answer, which could partly explain the InternVL-specific improvements. This concern is mitigated by the fact that the same dataset also helps LLaVA, which is outside the judge family, but the paper should acknowledge the potential selection-family effect and, ideally, run the referee filtering with at least one judge outside the InternVL family to quantify its influence.
  4. [§3.2, Table 2] All results in Table 2 appear to be single runs without variance estimates or significance tests. Several reported gains are small (e.g., LLaVA-v1.5-13B POPE 86.03→86.20; InternVL-Chat-v1.0-13B MME 1525→1534; LLaVA-v1.5-13B LLaVA-Wild 62.6→63.1), and the claim of improvement on all 16 benchmarks would be more convincing with multiple random seeds or per-benchmark significance testing. At minimum, the paper should state whether the differences are stable across seeds and report standard deviations for the main table.
  5. [§A.7, Table A-5] The equal-sample-size comparison contains implausible baseline numbers that undermine its role as a control. For example, fine-tuning LLaVA-v1.5-7B with LLaVA-OneVision data is reported with a SEED score of 30.06 and with ALLaVA-4V data as 40.52, far below the Baseline's 58.62; the InternVL row shows a similar pattern (40.87 and 48.36 vs. Baseline 59.06). These values are not consistent with expected performance levels and suggest a possible evaluation or sampling error. The authors should verify the evaluation code and data-sampling procedure for these rows before the matched-sample-size comparison can be used as evidence.
minor comments (5)
  1. [Throughout] There are several typos and inconsistencies: 'comparision' in Table A-4, 'tunning' in Introduction and Related Work, 'properities' in Section 2.2, 'leval' in Appendix A.2, and 'the third part ... screen the ten images' in Appendix A.4 should presumably read 'ten task types'. The abstract says 413,648 samples while Table A-4 says 431K; the numbers should be reconciled.
  2. [§A.6] The benchmark name is written inconsistently as 'MM-VeT' and 'MMVet' across the paper and appendix; please unify the notation.
  3. [Table A-6] The table title says 'InternVL-Chat-v1.0-8B' while the surrounding text and Section A.8 refer to 'InternVL-Chat-V2.0-8B'; the table header should match the model actually used.
  4. [Figure 6 and §3.3] The horizontal-axis labels in Figure 6 are not fully self-explanatory: the left panel reports the number of task types at a constant 100k sample budget, and the right panel reports the per-task-type sample cap (5 to 55) resulting in different total sample sizes. Please make these axes explicit and describe in the caption how samples are selected when the number of task types is reduced.
  5. [§6, Ethics Statement] The ethics statement is generic boilerplate about informed consent and human participants, which is not directly applicable to a dataset generated from public image collections and synthetic QA pairs. The licensing discussion in Appendix A.5 is more relevant; consider replacing the generic paragraph with a dataset-specific statement about image provenance, licensing, and content risks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TaskGalaxy is an empirically evaluated dataset, with no fitted parameter or self-citation chain that makes its benchmark conclusions equivalent to its construction inputs.

full rationale

The paper's derivation chain is empirical rather than definitional: TaskGalaxy is constructed by expanding manually seeded task types with GPT-4o, matching them to open-source images with CLIP, and filtering generated Q&A with three external open-source judges. The central claim, that incorporating TaskGalaxy improves LLaVA-v1.5 and InternVL-Chat-v1.0 across 16 benchmarks, is then tested on held-out standardized benchmarks. No equation in the paper reduces a predicted quantity to a fitted parameter or to the dataset's own definition. The only same-family element is that InternVL-Chat-v1.5 and InternVL2-26B are used as referee filters while InternVL-Chat-v1.0 is the evaluated model; this is not load-bearing for the benchmark claim and the paper itself acknowledges judge fallibility, e.g., 'The Internvl model, however, provided a score, which may reflect a lack of mathematical reasoning capabilities, leading to an incorrect evaluation.' The train/evaluation image-overlap issue (COCO and TextVQA images appearing in Visual Genome, ShareGPT4V, and MathV360K) is a real contamination risk for causal attribution, but it is a validity concern, not a circularity: it does not make any benchmark result equivalent to the dataset's construction by definition. There is no load-bearing self-citation chain, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. Accordingly, the honest circularity finding is no significant circularity.

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

The central claim rests on four domain assumptions: the reliability of GPT-4o's task expansion, the usefulness of CLIP matching, the reliability of the open-source judges, and the absence of benchmark contamination. The last assumption is the most fragile and is not verified in the paper.

free parameters (4)
  • CLIP top-k = 10
    Number of task types retrieved per image in the match stage; chosen by hand without a sensitivity study.
  • Referee score threshold = 2 (out of 3)
    Samples are kept only if at least two of three open-source VLM judges give a positive score. Chosen by hand.
  • Max samples per task type = 55
    For balance, 1-55 samples are randomly selected per task type. The upper limit is a design choice; the ablation varies it from 5 to 55.
  • Number of hierarchy levels = 3
    The task taxonomy is expanded to three levels; deeper hierarchies are not explored.
assumptions (4)
  • domain assumption GPT-4o can generate non-overlapping, comprehensive task types from seed prompts covering all multimodal content-understanding scenarios.
    Invoked in Section 2.2 Hierarchical Task Type Generation; if outputs are overlapping or biased, the 19,227 task types would be inflated and noisy.
  • domain assumption CLIP image-text cosine similarity is a useful first-pass signal for matching task types to images.
    Used in Section 2.2 Match and Filter (Eq. 1). The paper itself notes CLIP can produce illusory matches, which is why a GPT-4o filter is added.
  • domain assumption The three open-source VLM judges provide reliable relevance scores, and a majority threshold of 2 correctly selects high-quality samples.
    Used in Section 2.2 Referee Screening. The paper gives failure examples where even these judges mis-score.
  • domain assumption The image sources do not overlap with evaluation benchmark images, or any overlap does not materially affect benchmark scores.
    Section 2.2 Image Collection (Table 1) lists COCO-based sources (Visual Genome, ShareGPT4V) and DocVQA (via MathV360K), which are related to TextVQA and POPE evaluation images. No decontamination check is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TaskGalaxy: Scaling Multi-modal Instruction Fine-tuning with Tens of Thousands Vision Task Types." pith.science (2026). https://pith.science/paper/P2YD3RDR

@misc{pith2026250209925,
  author       = {Pith},
  title        = {Pith review of: TaskGalaxy: Scaling Multi-modal Instruction Fine-tuning with Tens of Thousands Vision Task Types},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P2YD3RDR}},
  note         = {Machine review of arXiv:2502.09925}
}
read the original abstract

Multimodal visual language models are gaining prominence in open-world applications, driven by advancements in model architectures, training techniques, and high-quality data. However, their performance is often limited by insufficient task-specific data, leading to poor generalization and biased outputs. Existing efforts to increase task diversity in fine-tuning datasets are hindered by the labor-intensive process of manual task labeling, which typically produces only a few hundred task types. To address this, we propose TaskGalaxy, a large-scale multimodal instruction fine-tuning dataset comprising 19,227 hierarchical task types and 413,648 samples. TaskGalaxy utilizes GPT-4o to enrich task diversity by expanding from a small set of manually defined tasks, with CLIP and GPT-4o filtering those that best match open-source images, and generating relevant question-answer pairs. Multiple models are employed to ensure sample quality. This automated process enhances both task diversity and data quality, reducing manual intervention. Incorporating TaskGalaxy into LLaVA-v1.5 and InternVL-Chat-v1.0 models shows substantial performance improvements across 16 benchmarks, demonstrating the critical importance of task diversity. TaskGalaxy is publicly released at https://github.com/Kwai-YuanQi/TaskGalaxy.

Figures

Figures reproduced from arXiv: 2502.09925 by the authors.

Figure 1
Figure 1. An illustration of the benefits of high task type coverage in TaskGalaxy for the SFT stage. We presented the performance of LLaVA-v1.5-13B and InternVL-Chat-v1.0-7B models, both before and after integrating TaskGalaxy into the fine-tuning dataset. supervised fine-tuning phase to enhance instruction adherence and logical reasoning. Despite these efforts, instructional datasets still face limitations in task diversity… view at source ↗
Figure 2
Figure 2. An overview of the task type and high-quality question-answer pairs generation pipeline for TaskGalaxy. We initially define the first level of visual task types, along with a small number of second and third level task types. Subsequently, we instruct GPT-4o to extend these to a broader range of task types. We then collect image modalities from existing publicly available datasets for matching task types with images… view at source ↗
Figure 3
Figure 3. The prompt template used in GPT-4o API for first-level task type generation. Hierarchical Task Type Generation. To address the challenge of expanding the range of visual task types in the multimodal domain, manual organization is impractical due to its time-intensive nature. Instead, we leverage a robust multimodal model to automate the construction of diverse task types with minimal human intervention. Rather than … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Sample images, task types, and Q&A in TaskGalaxy. The Task Type refers to the visual task related to the image. Question and Answers are generated by GPT-4o and subsequently filtered by three refereeing models. Match and Filter. Given the complexity and labor-intensive…
Figure 5
Figure 5. Figure 5: Distribution of the number of images across the 19,227 task types in TaskGalaxy. The ranges 1-10, 21-40 and etc. indicate the number of samples associated with different task types in TaskGalaxy. The corresponding ratios represent the proportion of task types that fall…
Figure 6
Figure 6. Figure 6: Plot showing the change in baseline performance with variations in the number of tasks and the total number of samples. Left: The effect of varying the number of task types (ranging from 2k to 19,227) on model performance, while maintaining a constant total sample size…
Figure 7
Figure 7. Figure 7: Plot illustrating the varia￾tion in MME Perception performance scores in relation to the number of task types and the total number of samples. The left bar represents per￾formance across different numbers of task types, while the right bar represents performance across…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Kwai Keye-VL 1.5 Technical Report

    cs.CV 2025-09 conditional novelty 6.0 of 10

    Keye-VL-1.5 combines similarity-based Slow-Fast video token allocation with progressive context extension and iterative RL, reporting leading video-understanding results among 8B-scale multimodal models.

  2. Kwai Keye-VL Technical Report

    cs.CV 2025-07 conditional novelty 6.0 of 10

    Kwai Keye-VL shows that a five-mode chain-of-thought cold-start plus mix-mode reinforcement learning can push an 8B multimodal model to strong short-video and general vision-language performance.

Reference graph

Works this paper leans on

42 extracted references · 11 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

  2. [4]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,

  3. [5]

    Additionally, the average performance across 15 benchmarks, excluding MME, increased by approx- imately 1.3 points with CoT

    in benchmarks such as MME, LLaV A-in-the-wild, and Q-Bench. Additionally, the average performance across 15 benchmarks, excluding MME, increased by approx- imately 1.3 points with CoT. These findings underscore the value of incorporating CoT prompting into multimodal models. Table A-7: Performance comparison of CoT validity verification. +max 5 refers to ...

  4. [7]

    Mme: A comprehensive evaluation benchmark for multimodal large language models

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394,

  5. [8]

    Chatglm: A family of large language models from glm-130b to glm-4 all tools

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793,

  6. [9]

    Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models

    11 Published as a conference paper at ICLR 2025 Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Pro- ceedings of the IEEE/CVF Confere...

  7. [10]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276,

  8. [14]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pp. 19730–19742. PMLR, 2023c. Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al...

Show all 42 references
  1. [15]

    Clevr-math: A dataset for compositional language, visual and mathematical reasoning

    Adam Dahlgren Lindstr ¨om and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358,

  2. [16]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 26296–26306, 2024a. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Vi...

  3. [17]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–25...

  4. [19]

    Ocr-vqa: Visual question answering by reading text in images

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pp. 947–952. IEEE,

  5. [20]

    Vicente Ordonez, Girish Kulkarni, and Tamara Berg

    Available at https://openai.com/index/ gpt-4v-system-card (accessed on February 14, 2025). Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24,

  6. [22]

    Solving geometry problems: Combining text and diagram interpretation

    Minjoon Seo, Hannaneh Hajishirzi, Ali Farhadi, Oren Etzioni, and Clint Malcolm. Solving geometry problems: Combining text and diagram interpretation. In Proceedings of the 2015 conference on empirical methods in natural language processing, pp. 1466–1476,

  7. [23]

    Math-llava: Bootstrapping mathematical reasoning for multimodal large language models

    Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294,

  8. [24]

    Aligning large multimodal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525,

  9. [25]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860,

  10. [26]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Baptiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023a. Hugo Touvron...

  11. [27]

    Q-bench: A benchmark for general-purpose foundation models on low-level vision

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181,

  12. [28]

    Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks

    Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394,

  13. [29]

    Vision-flan: Scaling human-labeled tasks in visual instruction tuning

    Zhiyang Xu, Chao Feng, Rulin Shao, Trevor Ashby, Ying Shen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. Vision-flan: Scaling human-labeled tasks in visual instruction tuning. arXiv preprint arXiv:2402.11690,

  14. [30]

    Baichuan 2: Open large-scale language models

    14 Published as a conference paper at ICLR 2025 Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305,

  15. [31]

    mplug-owl: Modularization empowers large language models with multimodality

    Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178,

  16. [32]

    Mm-vet: Evaluating large multimodal models for integrated capabilities.arXiv preprint arXiv:2308.02490,

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities.arXiv preprint arXiv:2308.02490,

  17. [33]

    Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs

    Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582, 2024a. Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, ...

  18. [34]

    Retrieval-augmented mixture of lora experts for uploadable machine learning.arXiv preprint arXiv:2406.16989,

    Ziyu Zhao, Leilei Gan, Guoyin Wang, Yuwei Hu, Tao Shen, Hongxia Yang, Kun Kuang, and Fei Wu. Retrieval-augmented mixture of lora experts for uploadable machine learning.arXiv preprint arXiv:2406.16989,

  19. [35]

    Minigpt-4: En- hancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En- hancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592,

  20. [36]

    Model tailor: Mitigating catastrophic forgetting in multi-modal large language models

    Didi Zhu, Zhongyi Sun, Zexi Li, Tao Shen, Ke Yan, Shouhong Ding, Kun Kuang, and Chao Wu. Model tailor: Mitigating catastrophic forgetting in multi-modal large language models. ICML 2024,

  21. [37]

    The approximate data sources and their corre- sponding sample sizes are presented in Table 1 of the main text

    15 Published as a conference paper at ICLR 2025 A A PPENDIX A.1 O VERALL REVIEW OF IMAGE SOURCES Considering the accessibility of data sources and the task-related nature of the image data we aim to mine, we have opted for open-source image data. The approximate data sources a...

  22. [38]

    data fusion,

    Please only output your final score without any other characters. p CoT You are a multimodal content understanding expert and you are very good at solving visual question answering tasks, I will give you an image and a question related to this image, the question is: {question...

  23. [39]

    flooding inundates the marina and affects nearby buildings and facilities

    Thus, the answer is incorrect. The task requires logical reasoning and complex algebraic calculations to solve the problem, but the answer does not meet the expected level of complexity required for this task type. Additionally, the image content appears to be more relevant to...

  24. [41]

    The best results are highlighted in bold. Model Method BenchmarksMME MMB MMBCN POPE LLaV AW MMVet TQA SQA MathVista InternVL-Chat-V2.0-8B Baseline 1536 68.52 66.46 86.30 63.20 46.17 66.24 90.58 50.10TaskGalaxy1565 73.88 70.79 86.90 62.85 48.86 70.49 92.71 52.31ChartQA AI2D Q-B...

  25. [2000]

    The best results are highlighted in bold. Model Method BenchmarksMME MMB MMBCN POPE LLaV AW MMVet TQA SQA MathVista LLaV A-v1.5-7B Baseline 1476 63.29 56.45 86.30 47.70 24.70 57.59 68.77 28.20ShareGPT-4V1501 65.97 59.10 86.29 49.20 29.00 57.56 70.60 28.20LLaV A-OneVision1251 5...

  26. [2014]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision–ECCV 2016: 14th Euro- pean Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 , pp. 235–251....

  27. [2017]

    Mimic-it: Multi-modal in-context instruction tuning

    Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023a. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, ...

  28. [2018]

    Referitgame: Referring to objects in photographs of natural scenes

    Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 787–798,

  29. [2019]

    Chartqa: A bench- mark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A bench- mark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244,

  30. [2020]

    Mapqa: A dataset for question answering on choropleth maps

    Shuaichen Chang, David Palzer, Jialin Li, Eric Fosler-Lussier, and Ningchuan Xiao. Mapqa: A dataset for question answering on choropleth maps. arXiv preprint arXiv:2211.08545,

  31. [2021]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    13 Published as a conference paper at ICLR 2025 Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXi...

  32. [2022]

    Allava: Harnessing gpt4v-synthesized data for a lite vision-language model

    Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhi- hong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024a. Keqin Chen, Zhao ...

  33. [2023]

    Direct preference optimization for suppressing hallucinated prior exams in radiology report gen- eration

    Oishi Banerjee, Hong-Yu Zhou, Subathra Adithan, Stephen Kwak, Kay Wu, and Pranav Rajpurkar. Direct preference optimization for suppressing hallucinated prior exams in radiology report gen- eration. arXiv preprint arXiv:2406.06496,

  34. [2024]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou

    Available at https://www.anthropic.com/news/ claude-3-5-sonnet (accessed on February 14, 2025). Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abi...

Pith tools

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