Pith. sign in

REVIEW 3 major objections 6 minor 4 cited by

M2IO-R1: An Efficient RL-Enhanced Reasoning Framework for Multimodal Retrieval Augmented Multimodal Generation

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a 3B-parameter inserter trained with reinforcement learning can decide which retrieved images belong in a generated text answer and where to place them, matching or exceeding much larger models while cutting latency.

desk verdict Plausible, useful RL-for-image-placement work with consistent gains on the main benchmark, but the inference prompt ambiguity and benchmark circularity need resolving before the central claim is solid. read the letter →

arxiv 2508.06328 v1 pith:INNWAIVV submitted 2025-08-08 cs.IR

Zhiyou Xiao , Qinhan Yu , Binghui Li , Geng Chen , Chong Chen , Wentao Zhang This is my paper
classification cs.IR
keywords multimodalretrieval-augmentedgenerationreinforcementlearningGRPOimageinsertioninterleavedtext-imageefficientLLMreasoningMRAMG
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 tackles multimodal retrieval-augmented multimodal generation (MRAMG): producing an answer that mixes generated text with retrieved images. It claims that the hard part is deciding which images to include and where to insert them, and that this decision can be learned by a small model using reinforcement learning with a purely rule-based reward. A 3B-parameter inserter trained with GRPO matches or beats a 72B untrained inserter on placement quality and outperforms single-shot and bipartite-matching baselines on MRAMG benchmarks, with lower latency and cost. The broader point is that RL-style outcome rewards can extend beyond math and code into general-purpose multimodal production tasks.

What carries the argument

The rule-based reward is the load-bearing mechanism: $R_{\text{total}} = R_{\text{format}} + R_{\text{answer}}$, with $R_{\text{answer}} = \alpha R_{\text{rec}} + (1-\alpha) R_{\text{pos}}$ and $\alpha=0.8$. The inserter is a Qwen2.5-VL-3B model fine-tuned with GRPO, and the reward converts a multi-step sequential decision (which images, at which positions) into a single scalar outcome signal, which is what makes the RL training work.

What would settle it

Take a fixed set of queries and candidate images, generate the text answer with the same model used during training and with a different model, and compare Inserter-R1-3B's image recall and position accuracy on the two text sets. A significant drop on the other-generator text would show the placement skill does not transfer across text distributions.

Watch

Extended reading notes

Core claim

The authors propose M2IO-R1, a four-stage pipeline: retrieve multimodal documents, generate a text answer, insert retrieved images, then merge. The central discovery is that image insertion is a trainable reasoning task: the 3B inserter outputs a <think> reasoning trace and then an <answer> dictionary mapping each image ID to the sentence index after which it belongs. Trained by GRPO with a rule-based reward—format compliance plus α times image recall plus (1−α) times per-position placement accuracy against MRAMG-Bench ground truth—the small inserter reaches or exceeds the placement quality of an untrained 72B inserter and beats single-shot and bipartite-matching baselines across MRAMG-Bench

Load-bearing premise

The inserter is trained with rewards computed against ground-truth text answers and image placements, but at inference it must place images in text written by a different generator; the paper assumes placement skill transfers to that unseen text without evidence.

Editorial extensions

If this is right

  • A 3B inserter trained with outcome rewards can match or exceed the placement quality of a 72B untrained inserter, so model scale is not the main driver of multimodal insertion quality.
  • Decomposing MRAMG into text generation and image insertion outperforms single-shot generation, so the bottleneck is the insertion decision, not the text itself.
  • RL-based insertion beats supervised fine-tuning on nearly all metrics, indicating that outcome-driven training captures sequential placement reasoning that imitation does not.
  • The framework lowers latency to 4.34 seconds per instance and cost to $0.24, making multimodal answers inexpensive enough for deployment.
  • Training on web-style data alone still generalizes to academic and instructional sets, so RL insertion does not overfit to one domain.

Reading between the lines

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

  • The paper never tests the order effect of splitting text generation and insertion; an ablation that inserts images in the same text with a single model would isolate the decomposition's contribution.
  • A plausible extension is to apply the same inserter to video or audio segments, where the placement decision is analogous.
  • The think-then-answer trace is decorative in the reward (only the final dictionary is scored); measuring whether removing the trace degrades performance would show whether the reasoning tokens are load-bearing or just a formatting artifact.
  • The paper uses a fixed α=0.8; an adaptive or per-domain α could further improve position accuracy on order-sensitive sets like Manual and Recipe.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper proposes M2IO-R1, a four-stage pipeline for multimodal retrieval-augmented multimodal generation (MRAMG): retrieve, generate a textual answer, insert images with an RL-based inserter, and merge. The inserter is a 3B Qwen2.5-VL model fine-tuned with GRPO using a rule-based reward combining format, image recall, and position accuracy. The authors report that M2IO-R1-3B outperforms Single-Shot and Rule-Based baselines on MRAMG-Bench, and matches or approaches a 72B base inserter, with lower latency. Experiments also include M2RAG and FTII-Bench, an ablation over the reward weight α, and a web-only training variant.

Significance. If the claims hold, the paper is a useful demonstration of RL-based image insertion for MRAMG, with a lightweight model beating heuristics and a training-free 72B inserter on the in-house benchmark. Strengths include a clear decomposition of the task, a simple outcome-based reward, external validation on two benchmarks from other groups (M2RAG, FTII-Bench), and a hyperparameter ablation over α. However, the central evaluation is currently under-specified: the provided inference and training prompts in Appendix A.4 feed the inserter a 'Ground truth answer dict' rather than the generated text described in Section 3.2. If this is what was run, the main tables measure insertion into oracle text, not into GPT-4o/Qwen outputs, so the claimed end-to-end gains are not tested. The interpretation of the numbers also needs error bars or significance tests.

major comments (3)
  1. [§3.2, §3.3, Appendix A.4] The inference prompt for M2IO-R1 in Appendix A.4 is identical to the training prompt and instructs the model to condition on a 'Ground truth answer dict: {ground_truth_dict}'. The paper never states that this placeholder is replaced by the generated answer Atxt at inference. If it is not replaced, Table 1 and Tables 5–11 measure insertion into the MRAMG-Bench ground-truth text rather than the text produced by GPT-4o/Qwen, so the central claim that the RL inserter improves over baselines in the proposed pipeline is unsupported. If it is replaced, the paper must state this explicitly and address the train/inference distribution shift: the reward (Eqs. 6–8) is computed against ground-truth image sets and ground-truth sentence indices, while inference inputs have different sentence structure and content. The same ambiguity affects the M2IO-Base and M2IO-SFT prompts, so the comparison may be
  2. [§4.2, Tables 1, 4] The primary evidence for the abstract's claim that M2IO-R1 'outperforms baselines in both quality and efficiency' comes from MRAMG-Bench, whose training set is sampled from the same benchmark used for reward computation and main evaluation, and which was introduced by an overlapping author group (Yu et al. 2025). On the independent M2RAG benchmark (Table 4), M2IO-R1-3B has Pos 70.5, below Single-Shot's 80.1, and on FTII-Bench it still lags M2IO-Base-72B by a substantial F1 margin (52.7 vs 64.9). The paper should either qualify the global claim or provide analysis of why the in-house gains do not transfer to the external benchmarks.
  3. [§4.2, Table 1] No error bars, variance, or significance tests are reported. Several headline improvements are small; for example, on Web with GPT-4o, M2IO-R1-3B's Ovr (84.4) is identical to M2IO-Base-72B (84.4), and its Rec/F1 gains over Single-Shot are roughly 2–3 points. Without multiple runs or a statistical test, the claimed superiority over baselines and the 'punching above its weight' conclusion are not firmly established.
minor comments (6)
  1. [§4.3, Table 2] The latency comparison reports M2IO at 4.34s while stating that the computational overhead of the locally deployed 3B model is disregarded. Please specify whether the reported number includes the inserter's inference time or only API calls, since a deployment-relevant latency comparison should include the local model.
  2. [§4.3] The 'Web-Style-Only MRAMG-Bench' setting is described as out-of-domain, but it is still evaluated on MRAMG-Bench. Consider renaming it to 'web-only training' or clarifying what domain shift is being tested.
  3. [§4.1, Table 4 caption] For M2RAG, Pos and Rel are obtained via LLM-as-a-Judge. Please specify the judge model and prompt, and state whether the same judge was used for all strategies to ensure a fair comparison.
  4. [Appendix A.4] The 'Training Prompt for M2IO-R1' and the 'Image Insertion Prompt for M2IO-R1' are identical in the appendix. Please clarify whether the same prompt is used at inference and whether the placeholder in the base/SFT prompts is also replaced by generated text.
  5. [§3.2] The framework is first referred to as 'MIMO-R1' before the name M2IO-R1 is used. This inconsistency should be fixed.
  6. [§4.3, Tables 10–11] The claim that α=0.8 is 'optimal' is based on qualitative balance. Some datasets prefer other values (e.g., Web Rec 96.1 at α=1, Arxiv F1 70.4 at α=0.5). Please report a selection criterion or note the sensitivity.

Circularity Check

1 steps flagged · score 4.0 of 10

The RL inserter is trained and primarily evaluated on the same author-created MRAMG-Bench, and the only supplied inference prompt still names the 'Ground truth answer dict', so the central end-to-end claim is not cleanly separated from the benchmark's own annotations.

  1. other [Section 3.2 (Eq. 3) vs Appendix A.4, Image Insertion Prompt for M2IO-R1 / M2IO-Base and M2IO-SFT]
    "Section 3.2: 'After generating the textual answer, we split the answer Atxt into sentences... Aimg = M(Pinsert, q, S, Iq)'. Appendix A.4: 'Given a question, a dictionary of ground truth answers (where keys are sentence indices and values are sentence content)... Question: {question} Ground truth answer dict: {ground_truth_dict}'."

    The pipeline definition (Eq. 3) makes the inserter's input the generated text S, but the only inference prompt supplied defines the input as the 'Ground truth answer dict' and is identical to the training prompt. The paper never states that this placeholder is replaced by the generated answer during evaluation. Read literally, the inserter receives MRAMG-Bench ground-truth answer text at inference, while its reward (Eqs. 6-8) and the main evaluation metrics were also computed from MRAMG-Bench ground-truth annotations. The reported gains would then show image insertion into an oracle text, not into GPT-4o/Qwen answers, so the central 'punching above its weight' claim is not derived from the proposed pipeline. The dependence of Table 1 on the T-Generator suggests the placeholder was probably

full rationale

The core RL training loop (GRPO with a rule-based reward on image selection and placement) is not circular by construction: the reward is a legitimate training signal, and the method is also tested on two externally-created benchmarks (M2RAG and FTII-Bench), so the author-created MRAMG-Bench self-citation is not the only evidence. However, the paper's own appendix defines the inference-time inserter prompt as a 'Ground truth answer dict' and never explains how the generated answer replaces it, while the training reward and the primary evaluation metrics both come from the same MRAMG-Bench ground truth. This creates a concrete, quotable ambiguity in which the main results could reduce to an oracle-text insertion task on the training benchmark rather than an end-to-end evaluation of RL-enhanced generation. Because the numerical tables vary with the text generator, the intended behavior was probably to use generated text, but the manuscript does not say so; this missing step prevents the central claim from being fully supported as written. No other load-bearing self-citation or definitional identification was found; the framework itself is a reasonable decomposition, and external benchmarks provide partial independent grounding.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on one fitted hyperparameter (alpha), plus domain assumptions about the benchmark's annotation quality, the sufficiency of the upstream retriever/generator, and the applicability of GRPO to this task. No new theoretical entities are introduced.

free parameters (1)
  • reward balance alpha = 0.8
    Weight between image recall and position accuracy in R_answer (Eq. 8). Selected via ablation on MRAMG-Bench (Appendix A.5), i.e., tuned on the evaluation benchmark.
assumptions (3)
  • domain assumption MRAMG-Bench ground truth image placements are correct and reflect optimal multimodal answers.
    Used as the reward signal during RL training (Section 3.3) and as evaluation ground truth (Section 4.1). If the annotations are biased or noisy, the inserter learns to replicate that bias and the reported scores overstate quality.
  • domain assumption The retriever (BGE-M3) and text generators (GPT-4o, Qwen2.5-VL) provide sufficient context for the inserter to make good placement decisions.
    The pipeline feeds retrieved text/images and generated text into the inserter (Section 3.2). Retrieval misses or generation errors degrade the inserter's inputs without being modeled.
  • domain assumption GRPO with a rule-based reward can optimize the discrete image placement decision in a 3B VLM.
    The paper assumes RL generalizes from math/code reasoning to this multimodal insertion task (Sections 1 and 3.3). This is an empirical claim, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M2IO-R1: An Efficient RL-Enhanced Reasoning Framework for Multimodal Retrieval Augmented Multimodal Generation." pith.science (2026). https://pith.science/paper/INNWAIVV

@misc{pith2026250806328,
  author       = {Pith},
  title        = {Pith review of: M2IO-R1: An Efficient RL-Enhanced Reasoning Framework for Multimodal Retrieval Augmented Multimodal Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INNWAIVV}},
  note         = {Machine review of arXiv:2508.06328}
}
read the original abstract

Current research on Multimodal Retrieval-Augmented Generation (MRAG) enables diverse multimodal inputs but remains limited to single-modality outputs, restricting expressive capacity and practical utility. In contrast, real-world applications often demand both multimodal inputs and multimodal outputs for effective communication and grounded reasoning. Motivated by the recent success of Reinforcement Learning (RL) in complex reasoning tasks for Large Language Models (LLMs), we adopt RL as a principled and effective paradigm to address the multi-step, outcome-driven challenges inherent in multimodal output generation. Here, we introduce M2IO-R1, a novel framework for Multimodal Retrieval-Augmented Multimodal Generation (MRAMG) that supports both multimodal inputs and outputs. Central to our framework is an RL-based inserter, Inserter-R1-3B, trained with Group Relative Policy Optimization to guide image selection and placement in a controllable and semantically aligned manner. Empirical results show that our lightweight 3B inserter achieves strong reasoning capabilities with significantly reduced latency, outperforming baselines in both quality and efficiency.

Figures

Figures reproduced from arXiv: 2508.06328 by the authors.

Figure 1
Figure 1. Illustration of two practical scenarios where users [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Fig1: Step1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. A case comparing the plain text output of M2IO-SFT-3B, which omits the expected illustrations, with the comprehensive [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: A case contrasting the multimodal output of M2IO-Base-3B, where images are placed inappropriately, with the [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: A case contrasting the multimodal output of M2IO-SFT-3B, which exhibits incorrect image selection, with the expected [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: A case comparing the multimodal output of M2IO-Base-3B, which exhibits incorrect image ordering, with the expected [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Gen-Searcher: Reinforcing Agentic Search for Image Generation

    cs.CV 2026-03 unverdicted novelty 7.0 of 10

    Gen-Searcher is the first trained search-augmented image generation agent using SFT followed by GRPO reinforcement learning with dual text-image rewards, delivering 15-16 point gains on knowledge-intensive benchmarks.

  2. VIG-RL: Learning to Search and Insert for Verified Image Grounding

    cs.IR 2026-07 conditional novelty 6.0 of 10

    An RL-trained ReAct agent learns joint search–selection–insertion of retrieved authentic images, setting SOTA on MRAMG-Bench Verified Image Grounding.

  3. Gen-Searcher: Reinforcing Agentic Search for Image Generation

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    Gen-Searcher is the first search-augmented image generation agent trained with SFT followed by agentic RL using dual text and image rewards on custom datasets and the KnowGen benchmark.

  4. Scaling Beyond Context: A Survey of Multimodal Retrieval-Augmented Generation for Document Understanding

    cs.CL 2025-10 accept novelty 4.0 of 10

    A systematic survey of Multimodal RAG for document understanding proposing a taxonomy based on domain, retrieval modality, and granularity while reviewing graph structures, agentic frameworks, datasets, benchmarks, ap...

Reference graph

Works this paper leans on

46 extracted references · 12 canonical work pages · cited by 3 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Anthropic. 2024. Claude 3.5 Sonnet. https://www.anthropic.com/news/claude-3-5-sonnet

    work page 2024

  4. [4]

    Asai, A.; Wu, Z.; Wang, Y.; Sil, A.; and Hajishirzi, H. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection

    2024

  5. [5]

    Bao, F.; Li, C.; Zhu, J.; and Zhang, B. 2022. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503

    arXiv 2022

  6. [6]

    T.; Filimonau, V.; and Sezerel, H

    Bui, H. T.; Filimonau, V.; and Sezerel, H. 2024. AI-thenticity: Exploring the effect of perceived authenticity of AI-generated visual content on tourist patronage intentions. Journal of Destination Marketing & Management, 34: 100956

    work page 2024

  7. [7]

    Chen, D.; Chen, R.; Pu, S.; Liu, Z.; Wu, Y.; Chen, C.; Liu, B.; Huang, Y.; Wan, Y.; Zhou, P.; et al. 2024 a . Interleaved scene graphs for interleaved text-and-image generation assessment. arXiv preprint arXiv:2411.17188

    arXiv 2024

  8. [8]

    Chen, J.; Xiao, S.; Zhang, P.; Luo, K.; Lian, D.; and Liu, Z. 2024 b . BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. arXiv:2402.03216

    arXiv 2024

Show all 46 references
  1. [9]

    Chen, W.; Hu, H.; Chen, X.; Verga, P.; and Cohen, W. W. 2022. Murag: Multimodal retrieval-augmented generator for open question answering over images and text. arXiv preprint arXiv:2210.02928

    2022 arXiv

  2. [10]

    Chen, Z.; Liu, K.; Wang, Q.; Zhang, W.; Liu, J.; Lin, D.; Chen, K.; and Zhao, F. 2024 c . Agent-flan: Designing data and methods of effective agent tuning for large language models. arXiv preprint arXiv:2403.12881

    2024 arXiv

  3. [11]

    Fan, A.; Jernite, Y.; Perez, E.; Grangier, D.; Weston, J.; and Auli, M. 2019. ELI5: Long form question answering. arXiv preprint arXiv:1907.09190

    2019 arXiv

  4. [12]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  5. [13]

    Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; Chen, Q.; Peng, W.; Feng, X.; Qin, B.; et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232

    2023 arXiv

  6. [14]

    Huang, W.; Jia, B.; Zhai, Z.; Cao, S.; Ye, Z.; Zhao, F.; Xu, Z.; Hu, Y.; and Lin, S. 2025. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749

    2025 arXiv

  7. [15]

    Jaech, A.; Kalai, A.; Lerer, A.; Richardson, A.; El-Kishky, A.; Low, A.; Helyar, A.; Madry, A.; Beutel, A.; Carney, A.; et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  8. [16]

    Jin, B.; Zeng, H.; Yue, Z.; Yoon, J.; Arik, S.; Wang, D.; Zamani, H.; and Han, J. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516

    2025 arXiv

  9. [17]

    u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \

    Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; K \"u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \"a schel, T.; et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33: 9459--9474

    2020

  10. [18]

    Li, T.; Tian, Y.; Li, H.; Deng, M.; and He, K. 2024. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 56424--56445

    2024

  11. [19]

    Lin, C.-Y. 2004. ROUGE : A Package for Automatic Evaluation of Summaries. In Text Summarization Branches Out, 74--81. Barcelona, Spain: Association for Computational Linguistics

    2004

  12. [20]

    J.; and Li, C

    Liu, H.; Son, K.; Yang, J.; Liu, C.; Gao, J.; Lee, Y. J.; and Li, C. 2023. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15148--15158

    2023

  13. [21]

    Liu, Z.; Zang, Y.; Zou, Y.; Liang, Z.; Dong, X.; Cao, Y.; Duan, H.; Lin, D.; and Wang, J. 2025. Visual Agentic Reinforcement Fine-Tuning. arXiv preprint arXiv:2505.14246

    2025 arXiv

  14. [22]

    Ma, Z.-A.; Lan, T.; Tu, R.-C.; Hu, Y.; Huang, H.; and Mao, X.-L. 2024. Multi-modal Retrieval Augmented Multi-modal Generation: A Benchmark, Evaluate Metrics and Strong Baselines. arXiv preprint arXiv:2411.16365

    2024 arXiv

  15. [23]

    Mei, L.; Mo, S.; Yang, Z.; and Chen, C. 2025. A survey of multimodal retrieval-augmented generation. arXiv preprint arXiv:2504.08748

    2025 arXiv

  16. [24]

    Meng, F.; Du, L.; Liu, Z.; Zhou, Z.; Lu, Q.; Fu, D.; Shi, B.; Wang, W.; He, J.; Zhang, K.; et al. 2025. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning. CoRR

    2025

  17. [25]

    OpenAI. 2024. Hello GPT-4o. OpenAI Blog

    2024

  18. [26]

    Peebles, W.; and Xie, S. 2023. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 4195--4205

    2023

  19. [27]

    Peng, Y.; Zhang, G.; Zhang, M.; You, Z.; Liu, J.; Zhu, Q.; Yang, K.; Xu, X.; Geng, X.; and Yang, X. 2025. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536

    2025 arXiv

  20. [28]

    Ruan, J.; Yang, Y.; Lin, Z.; Feng, Y.; Xiong, F.; Tang, Z.; and Li, Z. 2024. Ftii-bench: A comprehensive multimodal benchmark for flow text with image insertion. arXiv preprint arXiv:2410.12564

    2024 arXiv

  21. [29]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  22. [30]

    Sheng, G.; Zhang, C.; Ye, Z.; Wu, X.; Zhang, W.; Zhang, R.; Peng, Y.; Lin, H.; and Wu, C. 2024. HybridFlow: A Flexible and Efficient RLHF Framework. arXiv preprint arXiv: 2409.19256

    2024 arXiv

  23. [31]

    Sun, Q.; Cui, Y.; Zhang, X.; Zhang, F.; Yu, Q.; Wang, Y.; Rao, Y.; Liu, J.; Huang, T.; and Wang, X. 2024. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14398--14409

    2024

  24. [32]

    M.; Hauth, A.; Millican, K.; et al

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

    2023 arXiv

  25. [33]

    I.; Burnell, R.; Bai, L.; Gulati, A.; Tanzer, G.; Vincent, D.; Pan, Z.; Wang, S.; et al

    Team, G.; Georgiev, P.; Lei, V. I.; Burnell, R.; Bai, L.; Gulati, A.; Tanzer, G.; Vincent, D.; Pan, Z.; Wang, S.; et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530

    2024 arXiv

  26. [34]

    Tian, C.; Zhu, X.; Xiong, Y.; Wang, W.; Chen, Z.; Wang, W.; Chen, Y.; Lu, L.; Lu, T.; Zhou, J.; et al. 2024 a . Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208

    2024 arXiv

  27. [35]

    Tian, K.; Jiang, Y.; Yuan, Z.; Peng, B.; and Wang, L. 2024 b . Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37: 84839--84865

    2024

  28. [36]

    Wang, Q.; Ding, R.; Zeng, Y.; Chen, Z.; Chen, L.; Wang, S.; Xie, P.; Huang, F.; and Zhao, F. 2025. VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning. arXiv preprint arXiv:2505.22019

    2025 arXiv

  29. [37]

    Yin, S.; Fu, C.; Zhao, S.; Li, K.; Sun, X.; Xu, T.; and Chen, E. 2024. A survey on multimodal large language models. National Science Review, 11(12): nwae403

    2024

  30. [38]

    Yu, Q.; Xiao, Z.; Li, B.; Wang, Z.; Chen, C.; and Zhang, W. 2025. MRAMG-Bench: A Comprehensive Benchmark for Advancing Multimodal Retrieval-Augmented Multimodal Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information...

    2025

  31. [39]

    Yu, S.; Tang, C.; Xu, B.; Cui, J.; Ran, J.; Yan, Y.; Liu, Z.; Wang, S.; Han, X.; Liu, Z.; et al. 2024. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594

    2024 arXiv

  32. [40]

    Q.; and Artzi, Y

    Zhang, T.; Kishore, V.; Wu, F.; Weinberger, K. Q.; and Artzi, Y. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  33. [41]

    Zhao, P.; Zhang, H.; Yu, Q.; Wang, Z.; Geng, Y.; Fu, F.; Yang, L.; Zhang, W.; and Cui, B. 2024. Retrieval-augmented generation for ai-generated content: A survey. arXiv preprint arXiv:2402.19473

    2024 arXiv

  34. [42]

    Zheng, Y.; Lu, J.; Wang, S.; Feng, Z.; Kuang, D.; and Xiong, Y. 2025. Easyr1: An efficient, scalable, multi-modality rl training framework

    2025

  35. [43]

    Zheng, Y.; Zhang, R.; Zhang, J.; Ye, Y.; Luo, Z.; Feng, Z.; and Ma, Y. 2024. LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). Bangk...

    2024

  36. [44]

    Zhou, G.; Qiu, P.; Chen, C.; Wang, J.; Yang, Z.; Xu, J.; and Qiu, M. 2025. Reinforced mllm: A survey on rl-based reasoning in multimodal large language models. arXiv preprint arXiv:2504.21277

    2025 arXiv

  37. [45]

    J.; Lian, D.; and Xiong, Y

    Zhou, J.; Liu, Z.; Liu, Z.; Xiao, S.; Wang, Y.; Zhao, B.; Zhang, C. J.; Lian, D.; and Xiong, Y. 2024. MegaPairs: Massive Data Synthesis For Universal Multimodal Retrieval. arXiv preprint arXiv:2412.14475

    2024 arXiv

  38. [46]

    S.; Feujio, L.; Maharaj, A.; and Li, Y

    Zhu, Z.; Lee, D.; Zhang, H.; Harsha, S. S.; Feujio, L.; Maharaj, A.; and Li, Y. 2024. Murar: A simple and effective multimodal retrieval and answer refinement framework for multimodal question answering. arXiv preprint arXiv:2408.08521

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.