Pith. sign in

REVIEW 5 minor 19 references

Coling-UniA at SciVQA 2025: Few-Shot Example Retrieval and Confidence-Informed Ensembling for Multimodal Large Language Models

T0 review · 0 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A confidence-informed ensemble of two open-weights multimodal large language models, using few-shot example retrieval and a 90% confidence gate, outperforms every single configuration on the SciVQA 2025 scientific visual question…

desk verdict A clean, honest shared-task system paper whose test-set superlative is slightly oversold, but the dev-side ablation and public code make it worth refereeing. read the letter →

arxiv 2507.02357 v1 pith:IQOAZBSD submitted 2025-07-03 cs.CL

classification cs.CL
keywords scientificvisualquestionansweringmultimodallargelanguagemodelsfew-shotexampleretrievalconfidencecalibrationconfidence-informedensemblingquestion-typeroutingSciVQA2025
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

Can off-the-shelf multimodal large language models, with no fine-tuning, answer questions about scientific figures competitively? This paper argues yes, and it identifies where the gains come from: not from a better model, but from a confidence-informed ensemble that routes questions to different model-and-example configurations. The system accepts an answer whenever InternVL3-78B, prompted with one few-shot example retrieved by image similarity, reports a confidence of at least 0.9, which covers roughly half of all instances with about 91% accuracy on development data. All remaining questions are sent to the configuration that performed best for that question type, and the resulting ensemble beats every single configuration on both development and test sets. The paper's claim is that on this task, answer confidence—computed as the mean token log-probability—is calibrated enough to serve as a reliable gate.

What carries the argument

The two-stage ensemble is carried by the confidence gate. The first stage computes answer confidence as the exponentiated mean log-probability of all generated answer tokens, and trusts InternVL3-78B (with one few-shot example retrieved by BLIP-2 embedding similarity, filtered by figure type) whenever that score is at least 0.9. The second stage, for the remaining instances, routes by question type to Pixtral-Large-2411 or InternVL3-78B with one or two few-shot examples retrieved by SBERT question similarity or by CLIP/BLIP-2 image-question similarity. The routing table is chosen by repeated 5-fold cross-validation on development data grouped by figure type and question type.

What would settle it

Compute the calibration curve of InternVL3-78B (1s_q_img_f, BLIP-2) on a held-out split of new scientific figures from a different distribution (e.g., different journals or figure types than the 65% line-chart-dominated training data). If the fraction of correct answers among instances with confidence >= 0.9 drops meaningfully below the 0.9 development value, or if the threshold selects far fewer or far more than half the instances, the ensemble's first-stage gating is broken and the overall score should fall back toward the single-model baseline.

Watch

Extended reading notes

Core claim

The paper's central empirical discovery is that a two-stage ensemble built from two unmodified MLLMs outperforms any of its component configurations on the SciVQA 2025 task. The first stage trusts InternVL3-78B, prompted with one few-shot example retrieved by BLIP-2 image-question similarity and filtered by figure type, whenever its answer confidence—computed as the exponentiated mean log-probability of the generated answer tokens—is at least 0.9. This captures about half of all instances with roughly 91% ROUGE-1 F1 accuracy on the development set. The second stage sends every remaining instance to a configuration chosen per question type: Pixtral-Large-2411 with two few-shot examples for binary questions, Pixtral-Large-2411 with two examples retrieved by question similarity plus image similarity for infinite-answer-set questions, and InternVL3-78B with one question-similarity example for everything else. The ensemble reaches 76.9 ROUGE-1 F1 on development and 78.6 on the test set, beating the single best individual run (75.0 and 77.2) and placing the system third in the official ranking with an average F1 of 85.12 across ROUGE-1, ROUGE-L, and BERTScore.

Load-bearing premise

The confidence scores of InternVL3-78B with BLIP-2-based example retrieval are calibrated well enough on the blind test set that a 0.9 threshold, tuned on development data, still selects about half of the instances with high accuracy.

Editorial extensions

If this is right

  • The confidence-informed ensemble reaches 78.6 ROUGE-1 F1 on the test set, improving on the best individual configuration by 1.4 points and on the shared-task baseline by about 4 points.
  • Providing one answerable and one unanswerable few-shot example improves the model's precision at identifying unanswerable questions, compared with using a single example.
  • Few-shot example retrieval by question similarity and by question-image similarity yield similar overall performance, so the simpler question-only retrieval is usually sufficient.
  • Performance varies strongly by question type, with binary and multiple-choice questions answered far better than infinite-answer-set questions about visual features, which remain the main weakness.
  • InternVL3-78B's confidence estimates are better calibrated than Pixtral-Large-2411's on this task, which is why the high-confidence gate is assigned to InternVL3-78B.

Reading between the lines

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

  • The confidence-gating recipe could transfer to other VQA benchmarks with the same two models; a cheap test is to run the same two-stage ensemble on ChartQA or PlotQA and compare the gate's selected fraction and accuracy against development-calibrated expectations.
  • Because 65% of the development figures are line charts, the per-figure-type routing is effectively tuned for line charts; on a balanced figure distribution the chosen configurations would likely differ, and the ensemble's margin over the single model might shrink.
  • The shared task provides oracle metadata (caption, figure type, subfigure count) as input; a real-world deployment would need to predict these, so a useful extension is to measure the score drop when the metadata is replaced by an automatic figure-type classifier.
  • The paper compares BLIP-2 retrieval against question-similarity retrieval but does not isolate pure image-only retrieval; since BLIP-2 embeddings mostly reflect image content, a direct image-only baseline would test whether the question component adds any signal.
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

0 major / 5 minor

Summary. The paper describes a system submitted to the SciVQA 2025 shared task on scientific visual question answering. The system combines two open-weight multimodal LLMs (InternVL3-78B and Pixtral-Large-2411) in a zero/few-shot setting, with several few-shot example retrieval strategies based on SBERT question similarity, CLIP question-image similarity, and BLIP-2 embeddings. A question/figure-type ensemble selects the best configuration per instance group on the development set, and a confidence-informed ensemble first accepts answers from InternVL3-78B (1s_q_img_f, BLIP-2) when the mean token-log-probability confidence is at least 0.9, then routes remaining instances to per-question-type configurations. The paper reports extensive development-set ablations (Table 2) and blind test results where the system ranks third out of seven with an average ROUGE-1/ROUGE-L/BERTScore F1 of 85.12.

Significance. If the results are taken at face value, the paper makes a modest but useful contribution: a reproducible, compute-constrained recipe for applying MLLMs to scientific VQA without fine-tuning. The strengths are the public code, the transparent ablation over individual configurations on the development set, and the external validation from the official leaderboard. The confidence-informed ensembling idea is simple and effective, showing a consistent gain over its component model on both dev (76.9 vs 75.0 R1-F1) and test (78.6 vs 77.2) for the evaluated baseline. The main limitation is that the test-set comparison is reported for only one individual configuration, so the claim that the ensemble outperforms 'using only one approach' on test is scoped to that baseline rather than to the best individual test configuration; this is a reporting issue that can be fixed by qualification.

minor comments (5)
  1. [Section 4, Table 2] The test-set section reports only three rows (one individual configuration and two ensembles), so the statement that ensembling 'improve[s] the performance compared to using only one approach' should be scoped to the InternVL (1s_q_img_f, BLIP2) baseline actually evaluated. Please either report additional individual configurations on the test set (if available) or replace 'work best' with 'work best among the configurations evaluated on the test set' and clarify whether test runs were limited by submission constraints.
  2. [Section 3.4] The 0.9 confidence threshold is selected on the development set without a sensitivity analysis; a small threshold sweep (e.g., 0.8–0.95) on the development set would help readers judge how robust the choice is and whether the calibration-based justification is stable.
  3. [Section 3.4, Table 2, Table 7, Appendix A.2] There are several typos: 'configruations' in Section 3.4; 'InvernVL' in Table 2's test row header; 'InternvL3-78B' in Table 7's note; and a stray colon in the key 'answer_options:' in the prompt template shown in Appendix A.2.
  4. [Figure 6] The caption describes the plot as a calibration plot but does not specify what is plotted; adding axis labels and a diagonal reference would clarify whether it shows binned empirical R1-F1 versus mean confidence and what the 'calibration gap' represents.
  5. [Section 4] No significance tests are reported, and the test set is a single blind sample; the 1.4-point test gain (78.6 vs 77.2) should be presented as a point estimate rather than a definitive improvement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports empirical system evaluations against an external shared-task benchmark, and its development-set model selection is standard practice rather than a fitted-input prediction.

full rationale

The paper contains no derivation chain whose conclusion is equivalent to an input by construction. Its central claim is that the confidence-informed ensemble outperforms individual configurations on the development and test sets of an external benchmark (SciVQA 2025). The ensemble design and the confidence threshold are chosen using development-set performance, which is standard model selection rather than circular reasoning: the reported test-set numbers are new data points on blind data. The test-set comparison in Table 2 reports only one individual configuration alongside the ensembles, so the claim that the ensemble beats 'the single best individual configuration' on test is not fully demonstrated by the displayed rows; however, this is a limitation of evidence presentation, not a circularity. There are no load-bearing self-citations: the cited datasets, models, and metrics are external. No fitted parameter is renamed as a prediction, and no result is defined in terms of another result it is supposed to predict. The honest finding is that the paper is self-contained as an empirical system description with no significant circularity.

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

The central claim rests on the reliability of pre-trained models and the provided metadata, plus the assumption that dev-set calibration carries over to the test set.

free parameters (2)
  • confidence_threshold = 0.9
    Chosen based on calibration plot on the development set (Section 3.4); directly controls which answers are accepted from the calibrated model, affecting roughly half the instances.
  • few_shot_k = 1 or 2
    Number of few-shot examples used; 2 for Pixtral, 1 for InternVL3-78B due to context window (Section 3.2); chosen based on dev set performance.
assumptions (4)
  • domain assumption Pretrained models (InternVL3-78B, Pixtral-Large, SBERT, CLIP, BLIP-2) provide meaningful features for retrieval and generation.
    The method relies on these off-the-shelf models without fine-tuning (Section 2).
  • domain assumption Oracle-style metadata (image caption, figure type, compound flag, answer options) is available and accurate for test instances.
    The shared task provides this metadata; the prompt uses it (Section 2, Appendix A.2).
  • domain assumption Development set (train+val combined) is representative of the test distribution.
    Model selection and confidence threshold are based on dev set; test results confirm, but this is an assumption (Sections 3-4).
  • domain assumption Unanswerable questions in train and test follow similar patterns.
    The system learns to identify unanswerable questions from few-shot examples; the authors note a limitation that real-world unanswerable questions may differ (Discussion, Limitations).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coling-UniA at SciVQA 2025: Few-Shot Example Retrieval and Confidence-Informed Ensembling for Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/IQOAZBSD

@misc{pith2026250702357,
  author       = {Pith},
  title        = {Pith review of: Coling-UniA at SciVQA 2025: Few-Shot Example Retrieval and Confidence-Informed Ensembling for Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IQOAZBSD}},
  note         = {Machine review of arXiv:2507.02357}
}
read the original abstract

This paper describes our system for the SciVQA 2025 Shared Task on Scientific Visual Question Answering. Our system employs an ensemble of two Multimodal Large Language Models and various few-shot example retrieval strategies. The model and few-shot setting are selected based on the figure and question type. We also select answers based on the models' confidence levels. On the blind test data, our system ranks third out of seven with an average F1 score of 85.12 across ROUGE-1, ROUGE-L, and BERTS. Our code is publicly available.

Figures

Figures reproduced from arXiv: 2507.02357 by the authors.

Figure 1
Figure 1. System overview. Abbreviations for few-shot [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Percentage of selected one-shot example matching the question type of the input instance. bin = binary question, MC-4 = four answer options, inf = infinite answer set, unansw. = unanswerable, (v) = vi￾sual, (nv) = non-visual, filtered = filter for same figure type, not filtered = search in entire train set, q = question similarity, img = image similarity. bin (nv) bin(v) MC-4 (nv) MC-4 (v) inf (nv) inf (v) unansw. 5… view at source ↗
Figure 3
Figure 3. ROUGE-1 F1 scores per question type. Box [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Number of instances having received high con [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Calibration plot for InternVL3-78B (1s_q_img_f, BLIP) showing that instances with confi￾dence score ≥ 0.9 have high expected accuracy. to depend on whether the question incorporates vi￾sual or non-visual features. Since the vast majority of the remaining instances are …
Figure 7
Figure 7. Figure 7: The zero-shot prompt is formatted based on the annotated metadata via conditional statements. The [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Figure type distribution on development set. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 3 canonical work pages

  1. [1]

    Ekaterina Borisova, Nikolas Rauscher, and Georg Rehm. 2025. SciVQA 2025: Overview of the first scientific visual question answering shared task. In Proceedings of the 5th Workshop on Scholarly Document Processing (SDP), Vienna, Austria

  2. [2]

    Yash Goyal, Tejas Khot, Douglas Summers - Stay, Dhruv Batra, and Devi Parikh. 2017. https://doi.org/10.1109/CVPR.2017.670 Making the V in VQA matter: Elevating the role of image understanding in visual question answering . In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages 6325--6334....

  3. [3]

    Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. 2023. https://doi.org/10.48550/ARXIV.2311.16483 Chartllama: A multimodal LLM for chart understanding and generation . CoRR, abs/2311.16483

  4. [4]

    Lee Giles

    Zeba Karishma, Shaurya Rohatgi, Kavya Shrinivas Puranik, Jian Wu, and C. Lee Giles. 2023. https://ceur-ws.org/Vol-3656/paper2.pdf Acl-fig: A dataset for scientific figure classification . In Proceedings of the Workshop on Scientific Document Understanding co-located with 37th AAAI Conference on Artificial Inteligence (AAAI 2023), Remote, February 14, 2023...

  5. [5]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. 2023. https://proceedings.mlr.press/v202/li23q.html BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models . In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of Proceedings of Machine L...

  6. [6]

    Shengzhi Li and Nima Tajbakhsh. 2023. https://doi.org/10.48550/ARXIV.2308.03349 Scigraphqa: A large-scale synthetic multi-turn question-answering dataset for scientific graphs . CoRR, abs/2308.03349

  7. [7]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

  8. [8]

    Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Eisenschlos. 2023. https://doi.org/10.18653/v1/2023.acl-long.714 M at C ha: Enhancing visual language pretraining with math reasoning and chart derendering . In Proceedings of the 61st Annual Meeting of the Association for Compu...

Show all 19 references
  1. [9]

    Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 2022. https://doi.org/10.18653/v1/2022.findings-acl.177 C hart QA : A benchmark for question answering about charts with visual and logical reasoning . In Findings of the Association for Computational Ling...

  2. [10]

    Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. 2021. https://doi.org/10.1109/WACV48630.2021.00225 Docvqa: A dataset for VQA on document images . In IEEE Winter Conference on Applications of Computer Vision, WACV 2021, Waikoloa, HI, USA, January 3-8, 2021 , pages 2199-...

  3. [11]

    Khapra, and Pratyush Kumar

    Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. 2020. https://doi.org/10.1109/WACV45572.2020.9093523 Plotqa: Reasoning over scientific plots . In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2...

  4. [12]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. http://proceedings.mlr.press/v139/radford21a.html Learning transferable visual models...

  5. [13]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

  6. [14]

    Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, and Yu Qiao. 2024. https://doi.org/10.48550/ARXIV.2402.12185 Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning . CoR...

  7. [15]

    Liang Zhang, Anwen Hu, Haiyang Xu, Ming Yan, Yichen Xu, Qin Jin, Ji Zhang, and Fei Huang. 2024. https://aclanthology.org/2024.emnlp-main.112 Tinychart: Efficient chart understanding with program-of-thoughts learning and visual token merging . In Proceedings of the 2024 Confere...

  8. [16]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with BERT . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April...

  9. [17]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, and 1 others. 2025. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479

  10. [18]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  11. [19]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.