Pith. sign in

REVIEW 4 major objections 5 minor 98 references

From Answers to Rationales: Self-Aligning Multimodal Reasoning with Answer-Oriented Chain-of-Thought

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Answer-first chain-of-thought training lets multimodal models learn from answers alone and beat human-annotated reasoning data.

desk verdict Useful self-training recipe for multimodal reasoning, but the missing train/test overlap check makes the headline 'beats human-annotated data' claim unverified. read the letter →

arxiv 2507.02984 v2 pith:TRNAVARX submitted 2025-07-01 cs.CL

Wentao Tan , Qiong Cao , Yibing Zhan , Chao Xue , Changxing Ding This is my paper
classification cs.CL
keywords multimodallargelanguagemodelsreasoninganswer-orientedchain-of-thoughtnegativerationalesdirectpreferenceoptimizationiterativeself-trainingvisualquestionansweringautomaticdatageneration
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 a multimodal large language model can train itself to reason better by generating its own rationales from answer labels, without human annotation or a separate language-model helper. Its device is an answer-oriented chain-of-thought (AoT) prompt: give the model the question, the image, and a pre-specified answer—correct for a positive rationale, incorrect for a deliberately plausible wrong rationale—and ask for concise step-by-step reasoning that ends at that answer. These preference pairs are then used in iterative direct preference optimization (DPO), which the paper calls SMART. In the main experiments this lifts Qwen2-VL-7B from 60.0 to 66.3 percent on MathVista and, with 6K samples, matches or beats human-annotated SQA and M3CoT data; the authors note that the recipe still needs a plausible wrong answer for every question, which is not always available. If the claim holds, the practical upshot is that high-quality reasoning training data can be manufactured automatically from multiple-choice questions alone.

What carries the argument

The load-bearing object is the Answer-oriented Chain-of-Thought (AoT) prompt, a proof-style instruction that feeds the model an image, a question, and a pre-specified answer and asks for a short step-by-step rationale ending with that answer. By using the answer as prior knowledge, the prompt steers the model onto the shortest correct path for positive rationales and, when the answer is wrong and the image is augmented with diffusion noise, flipping, or cropping, forces it to invent plausible but false steps for negative rationales. A conclusion filter and a circularity filter clean the generated pairs, which are optimized with the DPO loss inside a frozen reference model; after each round the updated model generates fresher data, producing the iterative generate-optimize loop that the paper calls SMART.

What would settle it

Run a near-duplicate search (token-level and embedding similarity, plus image matching) between the 6K sampled training questions and the test sets of MathVista, M3CoT, MM-Vet, MMCode, and LLaVA-Bench; if more than a trivial fraction of training items have near-identical counterparts in those tests, the reported gains could reflect memorization rather than reasoning.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that conditioning a reasoning prompt on the answer—rather than on the question alone—turns a single multimodal model into both a teacher and a student. With the correct answer supplied, the model produces concise, logically coherent positive rationales; with an incorrect answer and an augmented image, it produces erroneous but internally persuasive negative rationales that are more discriminative than those from naive prompting. Training with these preference pairs through DPO, and repeating the generate-train cycle, improves four different open multimodal models across math, multi-step reasoning, spatial understanding, code, and open-ended generation. The strongest comparison shows AoT DPO data reaching 64.7 percent on MathVista, above the human-annotated SQA dataset (62.9) and level with M3CoT (64.2), while also beating the automatic DD-CoT, CCoT, SeVa, SENA, and MPO baselines.

Load-bearing premise

The load-bearing premise is that the 6,000 training questions sampled from MathV360K are not in disguise the same questions used to evaluate the models; no overlap analysis is reported.

Editorial extensions

If this is right

  • A manually labeled reasoning dataset is no longer required: AoT data from 6K automatic samples beats SQA and ties M3CoT on MathVista, so annotation cost drops sharply.
  • Negative rationales are doing real work: with the same AoT-positive rationales, replacing naive negatives with AoT negatives moves MathVista from 64.1 to 64.7 and M3CoT from 63.3 to 64.0, so the wrong-reasoning half of the dataset is a measurable source of the gain.
  • Iteration compounds the gain: a second generate-train round lifts Qwen2-VL-7B further on MathVista (66.3), M3CoT (65.9), and MM-Vet (66.6) before performance stabilizes.
  • The recipe transfers across model families: four open multimodal models with different architectures and pretraining all improve, implying the method is not tied to one base model.
  • The framework's only input requirement is multiple-choice questions with known wrong answers, which makes it applicable to any MCQ corpus.

Reading between the lines

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

  • Editorial inference: the same answer-conditioning mechanism should transfer beyond multiple-choice formats; any task with a verifiable final answer (code, algebra, constraint satisfaction) can generate positive rationales from correct outputs and negative rationales from perturbed ones.
  • Editorial inference: wrong answers are currently sampled randomly; selecting distractors that are semantically closest to the correct answer would likely make negative rationales harder and the DPO signal stronger, which is a cheap experiment the paper does not run.
  • Editorial inference: because the paper reports no overlap or deduplication analysis between the MathV360K training sample and the evaluation benchmarks, the immediate check is to measure near-duplicate contamination; this determines how much of the reported gain is reasoning rather than memorization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SMART, a self-training framework for multimodal large language models (MLLMs) that iteratively generates reasoning preference data and optimizes with DPO. Its core component, AoT (Answer-oriented Chain-of-Thought), prompts an MLLM with the image, question, and a supplied correct or incorrect answer, asking the model to produce a step-by-step rationale ending in that answer. Correct answers yield positive rationales; incorrect answers, combined with image augmentations, yield supposedly persuasive negative rationales. The generated preference pairs are used for DPO, and the process repeats across iterations. Experiments on four MLLMs and five benchmarks (MathVista, M3CoT, MM-Vet, MMCode, LLaVA-Bench) report consistent gains and claim that AoT-generated data surpasses human-annotated datasets such as SQA and M3CoT in some settings.

Significance. If the claims are validated, this is a useful contribution: a single-MLLM, single-pass data generation method that produces both positive and negative rationales, needs no manual annotation, and is shown to improve several open MLLMs across architectures and sizes. The paper ships code, provides qualitative examples, and includes ablations that isolate the contribution of AoT versus naive prompts, the value of negative rationales, and the effect of iterative training. These are concrete strengths. However, the central comparative claim—that AoT-trained models outperform models trained on human-annotated data—is not yet established because of a missing train/test overlap analysis, a formal inconsistency in how negative rationales enter the DPO loss, and confounded comparisons in the key table. The paper's usefulness therefore depends on fixing these load-bearing points.

major comments (4)
  1. [Section 3.1 and Section 4.2] The 6K MathV360K training subset is never checked for overlap with the evaluation benchmarks MathVista, M3CoT, MM-Vet, or MMCode. Since MathV360K aggregates public multimodal math QA sources (the paper itself mentions ChartQA and Geometry3K), and the evaluation benchmarks are built from public tasks, the absence of exact-match, fuzzy-match, image-hash, or thematic-overlap analysis creates a concrete leakage risk. This is load-bearing: Table 3's headline comparison (AoT DPO 64.7 vs SQA 62.9, M3CoT 64.2 on MathVista) and Table 1's across-the-board gains would be inflated if even a small fraction of the 6K training items are test-question variants. Please report a deduplication analysis, or evaluate on a held-out set with known disjoint provenance.
  2. [Section 3.1 and Section 3.2, Eq. (1)] The DPO loss conditions negative rationales on the original image I_i, but negative rationales are generated from an augmented image I'_i. Section 3.1 defines R'_i ~ θ_k(I'_i, Q_i, A'_i, PAoT), while Eq. (1) scores π_{θ_k}(R'_i | I_i, Q_i). If the implementation follows this equation, the model is asked to assign likelihood to a rationale generated under a different visual input, making the negative sample artificially unlikely and potentially inflating DPO gains. If the implementation instead conditions R'_i on I'_i, then chosen and rejected examples are responses to different prompts, which is not a valid DPO setup. Please clarify the exact training input for R'_i and justify why the preference pair is coherent under that input.
  3. [Table 3] The central claim that AoT-generated data outperforms manually annotated data is supported mainly by comparisons that simultaneously change the data source and the training method. SQA and M3CoT are trained with SFT, whereas the AoT numbers that exceed both (e.g., 64.7 vs 62.9 and 64.2 on MathVista) come from DPO. Although the AoT SFT row (64.1) provides a same-method comparison, the datasets also differ in size (6,000 vs 6,185 vs 7,861), and no DPO baseline is run on human-annotated preference data. To attribute the gains to data quality rather than to the SFT/DPO difference, please add matched comparisons: SFT on equal-size subsets, and DPO on human-annotated preference pairs extracted from SQA/M3CoT.
  4. [Tables 1-3 and Fig. B7] All reported numbers are single runs without error bars, confidence intervals, or significance tests. Several differences that carry the paper's narrative are small, such as 64.7 vs 64.2 for AoT DPO versus M3CoT on MathVista, and 64.1 vs 63.9 on MM-Vet in Table 2. Without repeated seeds or a statistical test, the claim that AoT 'significantly improves' models and 'surpasses' human-annotated data is not established. Please report at least three seeds with means and standard deviations, or a matched-pairs test, for the key comparisons.
minor comments (5)
  1. [Throughout] The benchmark names appear inconsistently as 'MM-V et' and 'LLaV A-Bench' in tables and captions; please standardize them to 'MM-Vet' and 'LLaVA-Bench'.
  2. [Figure 3] The negative-rationale example ends with 'AB = 5.5 units, which matches choice (D) 5', but 5.5 does not equal 5; if this is intentionally a 'persuasive error', the caption should say so explicitly, otherwise the example is confusing.
  3. [Section 4.4] The sentence 'Next, we using AoT and the Naive prompts...' should be 'Next, we use AoT and the Naive prompts...'.
  4. [Figure 1] The right panel reports '+7.1 Overall score' and '+17.9 Overall score' but never defines 'Overall score' or lists the component benchmarks; please add a definition.
  5. [Section 3.1] The circularity filter is applied only to positive rationales, with n=3 set as a 'stringent criterion'; the choice of n=3 and the threshold 'more than three times' appear arbitrary and should be justified or ablated.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the AoT/SMART derivation is empirically self-contained; the only same-author citation (SENA) is non-load-bearing context for iterative DPO.

full rationale

The paper's core claim—that answer-conditioned generation of positive and negative rationales yields training data that improves MLLM reasoning—is not circular. The AoT prompt conditions generation on the correct or an incorrect answer, but the resulting rationales are then used as SFT/DPO training data and evaluated on external, held-out benchmarks (MathVista, M3CoT, MM-Vet, MMCode, LLaVA-Bench). The benchmark scores are not used to filter or fit the generated data, and the M=6K sample size is selected after a data-size ablation (Fig. 4), not reverse-engineered from test results. The same-author citation is SENA [33], which is cited only as one of several iterative-DPO inspirations ('Drawing inspiration from the iterative DPO strategy [33, 53, 54]'); it does not supply a uniqueness theorem or forbid alternative methods, and the experimental comparisons with SeVa, MPO, DD-CoT, and CCoT are reproduced in-paper. The paper even names a 'Circularity Filter,' but that filter is an n-gram repetition filter for data quality, not an instance of logical circularity. The absence of a MathV360K/benchmark overlap analysis is a potential dataset-leakage risk, but that is a correctness concern, not a circularity of the derivation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method introduces no new theoretical entities, but it relies on several domain assumptions about self-training and benchmark validity. The main free parameters are standard hyperparameters and heuristic data-generation controls; the most concerning is the untested assumption of no overlap between training and evaluation data.

free parameters (7)
  • DPO beta = 0.1
    Standard DPO hyperparameter controlling preference strength; set to default, not tuned per-task.
  • learning rate = 2e-6
    Learning rate for DPO training; chosen from standard practice for fine-tuning MLLMs.
  • sample size M = 6000
    Chosen based on Fig. 4 to balance performance and compute; this is a selection on the benchmark result, a mild free parameter.
  • diffusion noise step = 600
    Augmentation strength for negative image generation; not ablated.
  • flip and erase probabilities = 0.5
    Augmentation probabilities for negative image generation; set by hand.
  • n-gram circularity threshold = 3
    Threshold for the circularity filter on positives; described as stringent but not ablated.
  • number of iterations = 1-3 depending on model
    Iteration count is chosen based on performance curves in Fig. B7 and Table 2, which is fitting on the evaluation data.
assumptions (5)
  • domain assumption MLLMs can produce coherent and useful rationales when prompted with the answer as a prior.
    The entire data generation pipeline assumes that the model, when given the correct or incorrect answer, will generate rationales of sufficient quality to train on. Invoked in Section 3.1.
  • domain assumption DPO training on model-generated preference pairs improves reasoning on held-out benchmarks.
    The approach relies on the standard DPO assumption that increasing the probability of chosen rationales and decreasing rejected rationales leads to better behavior. This is a wide-spread assumption but not proven for this setting.
  • domain assumption The evaluation benchmarks (MathVista, M3CoT, MM-Vet, MMCode, LLaVA-Bench) are valid measurements of reasoning ability.
    The paper uses these benchmarks as ground truth for 'reasoning' without justifying their construct validity.
  • ad hoc to paper The sampled MathV360K subset is free of overlap with the evaluation benchmark questions.
    The paper states that samples do not overlap across iterations, but does not report checking overlap with test sets. This assumption is critical and unverified.
  • ad hoc to paper The conclusion filter and circularity filter improve data quality without introducing bias.
    The filtering rules (final step must contain the answer; n-gram repetition threshold) are ad hoc and not ablated, yet they define the training data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Answers to Rationales: Self-Aligning Multimodal Reasoning with Answer-Oriented Chain-of-Thought." pith.science (2026). https://pith.science/paper/TRNAVARX

@misc{pith2026250702984,
  author       = {Pith},
  title        = {Pith review of: From Answers to Rationales: Self-Aligning Multimodal Reasoning with Answer-Oriented Chain-of-Thought},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRNAVARX}},
  note         = {Machine review of arXiv:2507.02984}
}
read the original abstract

Achieving human-like reasoning capabilities in Multimodal Large Language Models (MLLMs) has long been a goal. Current methods primarily focus on synthesizing positive rationales, typically relying on manual annotations or complex systems. Moreover, they often overlook negative reasoning, which limits the model's generalization ability and robustness in multimodal inference. To address this gap, we propose a novel framework: \textbf{S}elf-Aligning \textbf{M}ultimodal Reasoning with \textbf{A}nswer-O\textbf{r}iented Chain-of-\textbf{T}hought (SMART). SMART employs an answer-oriented chain-of-thought (AoT) prompt to automatically construct high-quality data. Drawing inspiration from human proof-based strategies, AoT leverages both correct and incorrect answers to extract key visual information that links questions and answers. When provided with correct answers, the model produces strong positive rationales. Conversely, when correct answers are replaced with incorrect alternatives, the model generates an erroneous yet compelling reasoning path, serving as a form of discriminative negative rationale. Models trained with AoT-generated data outperform those trained on manually annotated datasets, demonstrating superior reasoning capabilities. Consequently, SMART establishes an iterative generation-optimization method that continually enhances the model's reasoning skills. Experiments indicate that the SMART framework significantly improves various MLLMs, regardless of model architecture, parameter size, or pre-training dataset. The code is available at https://github.com/WentaoTan/SMART.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

98 extracted references · 43 canonical work pages

  1. [1]

    arXiv preprint arXiv:2310.09478 (2023)

    Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chan- dra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)

    arXiv 2023

  2. [2]

    arXiv preprint arXiv:2306.15195 (2023)

    Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleash- ing multimodal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)

    arXiv 2023

  3. [3]

    arXiv preprint arXiv:2306.14824 (2023)

    Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos- 2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)

    arXiv 2023

  4. [4]

    arXiv preprint arXiv:2408.16500 (2024) 14

    Hong, W., Wang, W., Ding, M., Yu, W., Lv, Q., Wang, Y., Cheng, Y., Huang, S., Ji, J., Xue, Z., et al.: Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500 (2024) 14

    arXiv 2024

  5. [5]

    In: International Conference on Machine Learning, pp

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: International Conference on Machine Learning, pp. 19730–19742 (2023). PMLR

    2023

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 24185–24198 (2024)

    2024

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Ye, Q., Xu, H., Ye, J., Yan, M., Hu, A., Liu, H., Qian, Q., Zhang, J., Huang, F.: mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13040–13051 (2024)

    2024

  8. [8]

    arXiv:2311.07575 (2023)

    Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., Han, J., Huang, S., Zhang, Y., He, X., Li, H., Qiao, Y.J.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv:2311.07575 (2023)

    arXiv 2023

Show all 98 references
  1. [9]

    arXiv preprint arXiv:2404.16821 (2024)

    Chen, Z., Wang, W., Tian, H., Ye, S., Gao, Z., Cui, E., Tong, W., Hu, K., Luo, J., Ma, Z., et al.: How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821 (2024)

    2024 arXiv

  2. [10]

    arXiv preprint arXiv:2403.05525 (2024)

    Lu, H., Liu, W., Zhang, B., Wang, B., Dong, K., Liu, B., Sun, J., Ren, T., Li, Z., Sun, Y., et al.: Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525 (2024)

    2024 arXiv

  3. [11]

    Young, A., Chen, B., Li, C., Huang, C., Zhang, G., Zhang, G., Li, H., Zhu, J., Chen, J., Chang, J., et al.: Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652 (2024)

    2024 arXiv

  4. [12]

    https://llava-vl.github

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: LLaV A-NeXT: Improved reasoning, OCR, and world knowledge (2024). https://llava-vl.github. io/blog/2024-01-30-llava-next/

    2024

  5. [13]

    arXiv preprint arXiv:2408.01800 (2024)

    Yao, Y., Yu, T., Zhang, A., Wang, C., Cui, J., Zhu, H., Cai, T., Li, H., Zhao, W., He, Z., et al.: Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800 (2024)

    2024 arXiv

  6. [14]

    arXiv preprint arXiv:2409.12191 (2024)

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., Lin, J.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint...

    2024 arXiv

  7. [15]

    arXiv preprint arXiv:2311.07397 (2023) 15

    Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397 (2023) 15

    2023 arXiv

  8. [16]

    NeurIPS (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. NeurIPS (2023)

    2023

  9. [17]

    arXiv:2306.13394 (2023)

    Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394 (2023)

    2023 arXiv

  10. [18]

    In: Proceedings of the IEEE International Conference on Computer Vision, pp

    Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 2425–2433 (2015)

    2015

  11. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed- bench: Benchmarking multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13299–13308 (2024)

    2024

  12. [20]

    arXiv preprint arXiv:2308.02490 (2023)

    Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)

    2023 arXiv

  13. [21]

    In: International Conference on Learning Representations (ICLR) (2024)

    Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.-W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  14. [22]

    ACL (2024)

    Chen, Q., Qin, L., Zhang, J., Chen, Z., Xu, X., Che, W.: M 3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought. ACL (2024)

    2024

  15. [23]

    NeurIPS (2022)

    Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.-W., Zhu, S.-C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. NeurIPS (2022)

    2022

  16. [24]

    arXiv preprint arXiv:2302.00923 (2023)

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)

    2023 arXiv

  17. [25]

    NeurIPS (2023)

    Zheng, G., Yang, B., Tang, J., Zhou, H.-Y., Yang, S.: Ddcot: Duty-distinct chain- of-thought prompting for multimodal reasoning in language models. NeurIPS (2023)

    2023

  18. [26]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Mondal, D., Modi, S., Panda, S., Singh, R., Rao, G.S.: Kam-cot: Knowledge augmented multimodal chain-of-thoughts reasoning. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 18798–18806 (2024)

    2024

  19. [27]

    ACM MM (2024) 16

    Gao, T., Chen, P., Zhang, M., Fu, C., Shen, Y., Zhang, Y., Zhang, S., Zheng, X., Sun, X., Cao, L., et al.: Cantor: Inspiring multimodal chain-of-thought of mllm. ACM MM (2024) 16

    2024

  20. [28]

    In: CVPR (2024)

    Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. In: CVPR (2024)

    2024

  21. [29]

    arXiv preprint arXiv:2311.09193 (2023)

    Wu, Y., Zhang, P., Xiong, W., Oguz, B., Gee, J.C., Nie, Y.: The role of chain-of-thought in complex vision-language reasoning task. arXiv preprint arXiv:2311.09193 (2023)

    2023 arXiv

  22. [30]

    arXiv preprint arXiv:2405.19716 (2024)

    Deng, Y., Lu, P., Yin, F., Hu, Z., Shen, S., Zou, J., Chang, K.-W., Wang, W.: Enhancing large vision language models with self-training on image comprehen- sion. arXiv preprint arXiv:2405.19716 (2024)

    2024 arXiv

  23. [31]

    arXiv preprint arXiv:2406.11280 (2024)

    Ahn, D., Choi, Y., Kim, S., Yu, Y., Kang, D., Choi, J.: i-srt: Aligning large multi- modal models for videos by iterative self-retrospective judgment. arXiv preprint arXiv:2406.11280 (2024)

    2024 arXiv

  24. [32]

    arXiv preprint arXiv:2405.15973 (2024)

    Wang, X., Chen, J., Wang, Z., Zhou, Y., Zhou, Y., Yao, H., Zhou, T., Gold- stein, T., Bhatia, P., Huang, F., et al.: Enhancing visual-language modality alignment in large vision language models via self-improvement. arXiv preprint arXiv:2405.15973 (2024)

    2024 arXiv

  25. [33]

    arXiv preprint arXiv:2412.15650 (2024)

    Tan, W., Cao, Q., Zhan, Y., Xue, C., Ding, C.: Beyond human data: Align- ing multimodal large language models by iterative self-evolution. arXiv preprint arXiv:2412.15650 (2024)

    2024 arXiv

  26. [34]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Kembhavi, A., Seo, M., Schwenk, D., Choi, J., Farhadi, A., Hajishirzi, H.: Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4999–5007 (2017)

    2017

  27. [35]

    In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pp

    Kembhavi, A., Salvato, M., Kolve, E., Seo, M., Hajishirzi, H., Farhadi, A.: A dia- gram is worth a dozen images. In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pp. 235–251 (2016). Springer

    2016

  28. [36]

    Barra, S., Bisogni, C., De Marsico, M., Ricciardi, S.: Visual question answer- ing: Which investigated applications? Pattern Recognition Letters 151, 325–331 (2021)

    2021

  29. [37]

    arXiv preprint arXiv:2103.03874 (2021)

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., Steinhardt, J.: Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874 (2021)

    2021 arXiv

  30. [38]

    In: European Conference on Computer Vision, pp

    Hessel, J., Hwang, J.D., Park, J.S., Zellers, R., Bhagavatula, C., Rohrbach, A., Saenko, K., Choi, Y.: The abduction of sherlock holmes: A dataset for visual abductive reasoning. In: European Conference on Computer Vision, pp. 558–575 (2022). Springer 17

    2022

  31. [39]

    arXiv preprint arXiv:2401.02582 (2024)

    Zhang, D., Yang, J., Lyu, H., Jin, Z., Yao, Y., Chen, M., Luo, J.: Cocot: Con- trastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582 (2024)

    2024 arXiv

  32. [40]

    arXiv preprint arXiv:2403.16999 (2024)

    Shao, H., Qian, S., Xiao, H., Song, G., Zong, Z., Wang, L., Liu, Y., Li, H.: Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999 (2024)

    2024 arXiv

  33. [41]

    Journal of machine learning research 21(140), 1–67 (2020)

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21(140), 1–67 (2020)

    2020

  34. [42]

    : Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. : Training language models to follow instructions with human feedback. Advances in neural information processing systems 35, 27730–27744 (2022)

    2022

  35. [43]

    arXiv preprint arXiv:2407.07895 (2024)

    Li, F., Zhang, R., Zhang, H., Zhang, Y., Li, B., Li, W., Ma, Z., Li, C.: Llava- next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895 (2024)

    2024 arXiv

  36. [44]

    In: Proceedings of the 37th International Conference on Neural Information Processing Systems (2024)

    Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: towards general-purpose vision-language models with instruc- tion tuning. In: Proceedings of the 37th International Conference on Neural Information Processing Systems (2024)

    2024

  37. [45]

    https://openai.com/research/ gpt-4v-system-card

    OpenAI.: GPT-4V(ision) System Card (2023). https://openai.com/research/ gpt-4v-system-card

    2023

  38. [46]

    arXiv preprint arXiv:2402.04792 (2024)

    Guo, S., Zhang, B., Liu, T., Liu, T., Khalman, M., Llinares, F., Rame, A., Mes- nard, T., Zhao, Y., Piot, B., et al.: Direct language model alignment from online ai feedback. arXiv preprint arXiv:2402.04792 (2024)

    2024 arXiv

  39. [47]

    arXiv preprint arXiv:2401.10020 (2024)

    Yuan, W., Pang, R.Y., Cho, K., Sukhbaatar, S., Xu, J., Weston, J.: Self-rewarding language models. arXiv preprint arXiv:2401.10020 (2024)

    2024 arXiv

  40. [48]

    arXiv preprint arXiv:2403.08635 (2024)

    Calandriello, D., Guo, D., Munos, R., Rowland, M., Tang, Y., Pires, B.A., Richemond, P.H., Lan, C.L., Valko, M., Liu, T., et al.: Human alignment of large language models through online preference optimisation. arXiv preprint arXiv:2403.08635 (2024)

    2024 arXiv

  41. [49]

    arXiv preprint arXiv:2405.07863 (2024)

    Dong, H., Xiong, W., Pang, B., Wang, H., Zhao, H., Zhou, Y., Jiang, N., Sahoo, D., Xiong, C., Zhang, T.: Rlhf workflow: From reward modeling to online rlhf. arXiv preprint arXiv:2405.07863 (2024)

    2024 arXiv

  42. [50]

    arXiv preprint arXiv:2401.01335 18 (2024)

    Chen, Z., Deng, Y., Yuan, H., Ji, K., Gu, Q.: Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335 18 (2024)

    2024 arXiv

  43. [51]

    Advances in Neural Information Processing Systems 35, 15476–15488 (2022)

    Zelikman, E., Wu, Y., Mu, J., Goodman, N.: Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems 35, 15476–15488 (2022)

    2022

  44. [52]

    arXiv preprint arXiv:2404.19733 (2024)

    Pang, R.Y., Yuan, W., Cho, K., He, H., Sukhbaatar, S., Weston, J.: Iterative reasoning preference optimization. arXiv preprint arXiv:2404.19733 (2024)

    2024 arXiv

  45. [53]

    arXiv preprint arXiv:2405.17220 (2024)

    Yu, T., Zhang, H., Yao, Y., Dang, Y., Chen, D., Lu, X., Cui, G., He, T., Liu, Z., Chua, T.-S., et al.: Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220 (2024)

    2024

  46. [54]

    arXiv preprint arXiv:2405.14622 (2024)

    Zhou, Y., Fan, Z., Cheng, D., Yang, S., Chen, Z., Cui, C., Wang, X., Li, Y., Zhang, L., Yao, H.: Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622 (2024)

    2024 arXiv

  47. [55]

    ACM MM (2024)

    Zhu, K., Zhao, L., Ge, Z., Zhang, X.: Self-supervised visual preference alignment. ACM MM (2024)

    2024

  48. [56]

    arXiv preprint arXiv:2406.17294 (2024)

    Shi, W., Hu, Z., Bin, Y., Liu, J., Yang, Y., Ng, S.-K., Bing, L., Lee, R.K.-W.: Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294 (2024)

    2024 arXiv

  49. [57]

    In: Findings of the Association for Computational Linguistics: ACL 2022, pp

    Masry, A., Do, X.L., Tan, J.Q., Joty, S., Hoque, E.: Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In: Findings of the Association for Computational Linguistics: ACL 2022, pp. 2263–2279 (2022)

    2022

  50. [58]

    Lu, P., Gong, R., Jiang, S., Qiu, L., Huang, S., Liang, X., Zhu, S.-c.: Inter- gps: Interpretable geometry problem solving with formal language and symbolic reasoning. In: Proceedings of the 59th Annual Meeting of the Association for Com- putational Linguistics and the 11th In...

    2021

  51. [59]

    NeurIPS (2022)

    Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. NeurIPS (2022)

    2022

  52. [60]

    Advances in neural information processing systems 33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020)

    2020

  53. [61]

    In: Findings of the Association for Computational Linguistics: EMNLP 2024, pp

    Li, K., Tian, Y., Hu, Q., Luo, Z., Huang, Z., Ma, J.: Mmcode: Benchmarking multimodal large language models for code generation with visually rich program- ming problems. In: Findings of the Association for Computational Linguistics: EMNLP 2024, pp. 736–783 (2024)

    2024

  54. [62]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern 19 Recognition, pp

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern 19 Recognition, pp. 26296–26306 (2024)

    2024

  55. [63]

    CoRR abs/2308.12966 (2023)

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. CoRR abs/2308.12966 (2023)

    2023 arXiv

  56. [64]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    2023 arXiv

  57. [65]

    arXiv preprint arXiv:2403.19443 (2024)

    Gou, Q., Nguyen, C.-T.: Mixed preference optimization: Reinforcement learning with data selection and better reference model. arXiv preprint arXiv:2403.19443 (2024)

    2024 arXiv

  58. [66]

    arXiv:2303.08774 (2023)

    OpenAI: Gpt-4 technical report. arXiv:2303.08774 (2023)

    2023 arXiv

  59. [67]

    generate-train

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024) Appendix A Prompt Comparison In the ablation study section, we compare the AoT promp...

    2024 arXiv

  60. [68]

    Sub-answers:

    <sub-question 2> ... Sub-answers:

  61. [69]

    <sub-answer 1> or 'Uncertain'

  62. [70]

    Uncertain

    <sub-answer 2> or 'Uncertain' ... Answer: <One of the options> or 'Uncertain' For a question, assume that you do not have any information about the picture, but try to answer the sub-questions and prioritize whether your general knowledge can answer it, and then consider wheth...

  63. [71]

    The side length of the square is given as 32 units

  64. [72]

    The diameter of the circle is equal to the side length of the square

  65. [73]

    The formula for the circumference C of a circle is given by: C = 2πr, where r is the radius of the circle

    Therefore, the radius of the circle is half of the diameter, which is 32/2 = 16 units. The formula for the circumference C of a circle is given by: C = 2πr, where r is the radius of the circle. Substituting the radius value into the formula: C = 2 π * 16 = 32 π So, the exact c...

  66. [74]

    - Each side of the square is equal to the height of the rectangle, which is given as 32 units

    Identify the Side Length of the Square: - The square is formed by the sides of the rectangle that is equal to the length of the rectangle's sides. - Each side of the square is equal to the height of the rectangle, which is given as 32 units

  67. [75]

    - The diagonal of a square with side length s is given by s\sqrt{2}

    Find the Radius of the Circle: - The diagonal of the square is the diameter of the circle. - The diagonal of a square with side length s is given by s\sqrt{2} . - In this case, the side length s is 32, so the diameter (and hence the radius) is 32\sqrt{2}

  68. [76]

    objects": [ {

    Calculate the Circumference: - The circumference C of a circle is given by the formula C = 2\pi r , where r is the radius. - Using the radius r = 32\sqrt{2} : C = 2\pi \times 32\sqrt{2} = 64\sqrt{2}\pi - Simplifying this further, we have: C = 64 \times 1.414 \pi = 89.472 \pi #...

  69. [77]

    Identify the total number of parts: The image is divided into 12 equal parts

  70. [78]

    objects": [ {

    Count the number of orange parts: In the image, there are 8 orange parts. Now, we can calculate the fraction of the image that is orange: \text{Fraction of orange} = \frac{\text{Number of orange parts}}{\text{Total number of parts}} = \frac{8}{12} Simplifying the fraction: \fr...

  71. [79]

    The image shows a circle divided into 12 equal segments

  72. [80]

    Out of these 12 segments, 5 are shaded orange. Now, we can calculate the fraction of the image that is orange: \text{Fraction of orange} = \frac{\text{Number of orange segments}}{\text{Total number of segments}} = \frac{5}{12} So, the fraction of the image that is orange is 5/...

  73. [81]

    **H**: This stage shows a single cell, which is not a blastocyst

  74. [82]

    **E**: This stage shows several cells grouped together, which is not a blastocyst

  75. [83]

    **K**: This stage shows multiple cells with a central cavity, which is characteristic of a blastocyst

  76. [84]

    **X**: This stage shows several cells arranged in a circular pattern, which is not a blastocyst

  77. [85]

    **Y**: This stage shows many cells with a central cavity, which is characteristic of a blastocyst

  78. [86]

    objects": [{

    **J**: This stage shows a large central cavity with cells around it, which is characteristic of a blastocyst. Based on the analysis, stages **K** and **J** are identified as the blastocyst. However, since the question asks for a single stage, the correct answer is: **K** # DD-...

  79. [87]

    **E** - Early blastocyst 3

    **H** - Single cell 2. **E** - Early blastocyst 3. **K** - Late blastocyst

  80. [88]

    (D) No

    **X** - Early blastocyst 5. **Y** - Midway through blastocyst 6. **J** - Late blastocyst From the options provided: - (A) H - Single cell - (B) J - Late blastocyst - (C) E - Early blastocyst - (D) Y - Midway through blastocyst The blastocyst stage is characterized by the forma...

  81. [89]

    The inferior abdominal aorta is located at the lower body

  82. [90]

    The upper systemic circulation is the part of the circulation system that includes the head, superior vena cava, pulmonaryartery, and pulmonary vein

  83. [91]

    The lower systemic circulation includes the abdominal aorta, kidneys, liver, and intestine. Since the inferior abdominal aorta is part of the lower systemic circulation and does not connect to the upper systemic circulation directly, a blockage in the inferior abdominal aorta ...

  84. [92]

    The inferior abdominal aorta is located below the heart and is part of the lower systemic circulation

  85. [93]

    The upper systemic circulation is located above the heart and includes the abdominal aorta

  86. [94]

    objects": [ {

    The abdominal aorta is the continuation of the inferior abdominal aorta. Since the abdominal aorta is a continuation of the inferior abdominal aorta and is part of the lower systemic circulation, a blockage in the inferior abdominal aorta would indeed affect the lower systemic...

  87. [95]

    The inferior abdominal aorta is part of the lower systemic circulation

  88. [96]

    The upper systemic circulation includes the superior vena cava and the aorta. Since the inferior abdominal aorta is part of the lower systemic circulation and not directly connected to the upper systemic circulation, a blockage in the inferior abdominal aorta would not affect ...

  89. [97]

    At x = 6 : \frac{dy}{dx} = \frac{1}{6 ln(2)}

  90. [98]

    The correct answer is: (C) smaller than Qwen-VL-7B + SMART ： The derivative of the function y = log_2(x) is \frac{1}{x ln(2)}

    At x = 8 : \frac{dy}{dx} = \frac{1}{8 ln(2)} Comparing the two derivatives: \frac{1}{6 ln(2)} \text{ (at } x = 6 and \frac{1}{8 ln(2)} \text{ (at } x = 8 Since 6 ln(2) > 8 ln(2) , it follows that: \frac{1}{6 ln(2)} < \frac{1}{8 ln(2)} Therefore, the derivative at x = 6 is smal...

Pith tools

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