Pith. sign in

REVIEW 2 major objections 5 minor 2 cited by

Wiki-R1 closes the pretraining-to-KB-VQA gap by generating a curriculum of retrieval difficulties and sampling only the samples that still give learning signal.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-15 14:42 UTC pith:2QUBORBG

load-bearing objection Solid, data-efficient curriculum RL for noisy multimodal RAG: the retriever-manipulation ladder is the real novelty, gains are real but modest, and the sampling proxy is the softest joint. the 2 major comments →

arxiv 2603.05256 v2 pith:2QUBORBG submitted 2026-03-05 cs.CV

Wiki-R1: Incentivizing Multimodal Reasoning for Knowledge-based VQA via Data and Sampling Curriculum

Shan Ning , Longtian Qiu , Xuming He This is my paper
classification cs.CV
keywords knowledge-based VQAcurriculum learningreinforcement learningretrieval-augmented generationmultimodal LLMsobservation propagationsparse rewards
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Knowledge-based visual question answering forces a pretrained multimodal model to reason over noisy Wikipedia passages whose style it never saw, so ordinary reinforcement learning mostly produces zero-reward trajectories and stalls. Wiki-R1 solves the problem by deliberately manufacturing a sequence of training distributions: it first hands the model only the gold article, then gradually adds distractors, and finally removes the gold guarantee so the data finally matches real retrieval. A second stage estimates how hard each remaining example is by spreading the sparse observed rewards across a graph of similar knowledge articles, then preferentially samples the ones whose accuracy sits near one-half. On Encyclopedic-VQA and InfoSeek the resulting policy sets new state-of-the-art numbers while training on only forty thousand examples. A reader who cares about teaching large multimodal models to use external knowledge under realistic noise now has a concrete recipe that works.

Core claim

The paper shows that a data-generation-based curriculum—controllably altering the retriever’s candidate set and gold-document inclusion, then sampling with difficulty estimates obtained by propagating sparse rewards—systematically incentivizes multimodal reasoning and lifts accuracy past all prior retrieval-augmented systems on two hard knowledge-VQA benchmarks.

What carries the argument

Controllable curriculum data generation (discrete gap levels that force ground-truth inclusion and vary the number of retrieved candidates) together with observation-propagation sampling (non-parametric label propagation of binary rewards over a TF-IDF graph of Wikipedia articles).

Load-bearing premise

The method assumes that the hardness of an unseen question can be reliably guessed from the textual similarity of its Wikipedia article to articles of questions whose rewards have already been observed.

What would settle it

Replace the article-similarity graph with random edges or disable propagation entirely; if the fraction of zero-advantage trajectories and final accuracy remain unchanged, the claimed difficulty-estimation step is inert.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Reinforcement learning becomes practical for multimodal RAG settings whose retrieval noise previously produced near-total reward sparsity.
  • Entity-balanced training sets of only 40 k examples can outperform systems that fine-tune on millions of samples.
  • The same gap-level schedule transfers to any domain where a retriever can be forced to include or exclude gold documents.
  • Models trained this way show stronger generalization on unseen questions and zero-shot transfer to related knowledge-VQA sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Observation propagation over shared latent structure could cut wasted trajectories in any sparse-reward RL domain (code problems sharing libraries, math problems sharing lemmas).
  • If article-level TF-IDF is only a weak proxy, swapping it for question embeddings or a learned difficulty head would further stabilize the sampling schedule.
  • Controllable generation of intermediate distributions is likely more powerful than pure selection curricula whenever the domain shift itself can be parameterized (noise level, context length, modality drop-out).

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. Wiki-R1 addresses sparse-reward RL for knowledge-based VQA by constructing a curriculum of training distributions. Controllable data generation manipulates the retriever via discrete gap levels g (k retrieved candidates and whether ground-truth is forced) so that difficulty rises from pure ground-truth (g=0) to the noisy inference-time distribution (g=G). A curriculum sampling schedule then selects samples whose estimated accuracy is near 0.5; difficulty is obtained by non-parametric label propagation of sparse binary rewards over a graph whose edges are TF-IDF (or Sentence-Transformer) similarities of ground-truth Wikipedia articles. Built on DAPO with Qwen2.5-VL 3B/7B and only 40k entity-balanced examples, the method reports new SOTA numbers: 37.1% on Encyclopedic-VQA (prior 35.5%) and 44.1% on InfoSeek (prior 40.1%), plus strong zero-shot transfer to ViQuAE and competitive oracle-entity upper bounds.

Significance. If the gains hold, the paper supplies a practical, data-efficient recipe for post-training MLLMs under noisy retrieval and a clear distributional gap from pretraining. The controllable gap-level construction is a clean, reusable idea that goes beyond ordinary sample re-weighting; the ablations (Table 4), training-dynamics plots (Figure 3), multi-run stability (Appendix B), and zero-shot transfer results give the empirical claims real weight. The work is therefore of clear interest to the multimodal RAG and RL-for-reasoning communities, even if the precise contribution of the observation-propagation component remains only partially isolated.

major comments (2)
  1. Table 4 shows that data curriculum alone already lifts DAPO substantially (EVQA overall 31.4 o34.5, InfoSeek 41.5 o43.0). Adding sampling curriculum without observation propagation hurts, and only the full Obs. Prop. recovers the final numbers. The recovery rests on Appendix A.4 / Algorithms 1–2: sparse binary rewards are propagated over a graph whose edges are defined solely by ground-truth article similarity (top-100 TF-IDF or Sentence-Transformer neighbors). The manuscript never measures whether article-level textual similarity predicts question-level solvability under the actual noisy multi-document retrieval used at training time. Without such a correlation (or an ablation that replaces the graph with random/entity-only edges), the headline gains cannot be cleanly attributed to the claimed sampling curriculum rather than to data curriculum plus extra RL steps.
  2. Section 3.3 and Algorithm 1 define gap levels g via (k,γ) but never report the empirical distribution of realized difficulties (fraction of non-zero advantages, mean reward) actually observed at each g. Because the upgrade rule is driven by a sliding-window accuracy threshold τ, it is possible that later gap levels remain easier than intended or that the schedule simply re-samples easy clusters. A short diagnostic table or plot of realized reward statistics per gap level would make the “controllable curriculum” claim falsifiable and would strengthen the causal link between the generation mechanism and the final accuracy.
minor comments (5)
  1. Figure 1 caption and surrounding text refer to “DAPO on KB-VQA” without stating the exact base model or retrieval configuration used for the preliminary curves; a one-sentence clarification would help readers reproduce the sparse-reward diagnosis.
  2. Table 1 reports “Avg.” across benchmarks but the column is not defined; clarify whether it is a simple mean of the two overall scores or a weighted average.
  3. Appendix A.1: the fusion weight λ is chosen on the training set of each benchmark separately (0.985 / 0.997). A short note on sensitivity of final accuracy to λ would be useful, given that the values sit so close to pure visual retrieval.
  4. Typographical inconsistencies appear in several places (e.g., “V olcengine”, “V .and T.”, “stardenotes”). A light copy-edit pass would improve readability.
  5. Section 4.3 states that the reward is exact-match binary; it would be helpful to confirm whether any soft matching or BEM-style partial credit is used during training, given that evaluation on EVQA uses BEM.

Circularity Check

0 steps flagged

No significant circularity: empirical curriculum-RL method with held-out benchmark evaluation; no derivation reduces to its inputs by construction.

full rationale

Wiki-R1 is an empirical post-training method paper. Its central claims are (i) a controllable retrieval-modification schedule (gap levels g=0..G that force inclusion/exclusion of ground-truth articles and vary candidate count) plus (ii) curriculum sampling that selects near-0.5-accuracy examples whose difficulties are estimated by non-parametric label propagation of sparse binary rewards over a TF-IDF/Sentence-Transformer graph of ground-truth Wikipedia articles, and (iii) the resulting accuracy numbers on the official Encyclopedic-VQA and InfoSeek test splits. These numbers are obtained by running the trained policy on held-out data with the public BEM / official scripts; they are not algebraically forced by any fitted parameter or definitional identity. Curriculum thresholds (tau=0.55, window w=300, G=6) and retrieval fusion weights are ordinary hyper-parameters chosen on training/validation accuracy, not quantities that redefine the reported test metrics. Self-citations (e.g., the authors’ prior NoisyGRPO / DA-DPO works) appear only in related-work discussion of RL curricula and are not invoked as uniqueness theorems or load-bearing premises that forbid alternatives. No equation equates a claimed “prediction” to a fitted input, no ansatz is smuggled via self-citation, and no known empirical pattern is merely renamed. The paper is therefore self-contained against external benchmarks; any weakness of the article-similarity proxy for question difficulty is a validity/correctness concern, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on a small set of hand-chosen curriculum and propagation hyper-parameters, the modeling assumption that retriever manipulation produces a useful difficulty continuum, and the assumption that article-level TF-IDF similarity predicts question difficulty. No new physical or mathematical entities are postulated; the free parameters are ordinary ML knobs.

free parameters (5)
  • curriculum upgrade threshold τ = 0.55
    Sliding-window accuracy threshold that triggers gap-level increase; set to 0.55 after sensitivity search.
  • maximum gap level G = 6
    Hardest retrieval setting (no forced ground-truth, k=G-1); chosen as 6.
  • sliding window size w = 300
    Number of recent samples used to compute moving accuracy for curriculum upgrade.
  • label-propagation smoothing α = 0.8
    Balances observed vs. propagated rewards; default neighborhood around 0.8.
  • visual-textual fusion weight λ = 0.985 / 0.997
    Selected on training-set recall separately for each benchmark (0.985 EVQA, 0.997 InfoSeek).
axioms (4)
  • domain assumption Manipulating the number of retrieved candidates k and the forced inclusion of the ground-truth article produces a monotonic difficulty continuum that bridges pretraining to the target KB-VQA distribution.
    Stated in §3.3 and used to define gap levels g=0..G; not independently validated beyond the observed training curves.
  • domain assumption Samples whose expected accuracy is near 0.5 yield the strongest non-zero advantages for policy-gradient RL.
    Cited from prior curriculum-RL work and used to center the Gaussian sampling schedule (§3.4).
  • ad hoc to paper Textual similarity of ground-truth Wikipedia articles (TF-IDF or Sentence-Transformer) is a sufficient proxy for correlating question-level solvability under noisy retrieval.
    Core of the observation-propagation graph (Appendix A.4); ablation in Table 5 shows modest sensitivity but no external validation.
  • domain assumption Binary exact-match reward is an adequate training signal for open-ended encyclopedic answers.
    Reward definition in §4.3; evaluation later uses BEM, creating a train-test metric mismatch that is left unanalyzed.
invented entities (2)
  • gap level g and retrieval modification function φ_g(k,γ) no independent evidence
    purpose: Discrete control knobs that generate the curriculum of training distributions.
    Defined in §3.3; purely algorithmic constructs with no independent physical existence.
  • observation-propagation graph K over VQA samples no independent evidence
    purpose: Propagate sparse binary rewards to unobserved samples for curriculum sampling.
    Constructed via top-100 TF-IDF edges (Appendix A.4); utility is internal to the method.

pith-pipeline@v1.1.0-grok45 · 23036 in / 3263 out tokens · 25944 ms · 2026-07-15T14:42:22.250274+00:00 · methodology

0 comments
read the original abstract

Knowledge-Based Visual Question Answering (KB-VQA) requires models to answer questions about an image by integrating external knowledge, posing significant challenges due to noisy retrieval and the structured, encyclopedic nature of the knowledge base. These characteristics create a distributional gap from pretrained multimodal large language models (MLLMs), making effective reasoning and domain adaptation difficult in the post-training stage. In this work, we propose \textit{Wiki-R1}, a data-generation-based curriculum reinforcement learning framework that systematically incentivizes reasoning in MLLMs for KB-VQA. Wiki-R1 constructs a sequence of training distributions aligned with the model's evolving capability, bridging the gap from pretraining to the KB-VQA target distribution. We introduce \textit{controllable curriculum data generation}, which manipulates the retriever to produce samples at desired difficulty levels, and a \textit{curriculum sampling strategy} that selects informative samples likely to yield non-zero advantages during RL updates. Sample difficulty is estimated using observed rewards and propagated to unobserved samples to guide learning. Experiments on two KB-VQA benchmarks, Encyclopedic VQA and InfoSeek, demonstrate that Wiki-R1 achieves new state-of-the-art results, improving accuracy from 35.5\% to 37.1\% on Encyclopedic VQA and from 40.1\% to 44.1\% on InfoSeek. The project page is available at https://artanic30.github.io/project_pages/WikiR1/.

Figures

Figures reproduced from arXiv: 2603.05256 by Longtian Qiu, Shan Ning, Xuming He.

Figure 1
Figure 1. Figure 1: ( 1a) and ( 1b): Training dynamics of DAPO on KB-VQA. RL optimization suffers from a high proportion of zero-advantage samples and low training accuracy, highlighting the distribution gap between pretraining and the KB-VQA target domain. ( 1c): Motivation of Wiki-R1. To mitigate this gap, Wiki-R1 generates a sequence of training distributions with progressively reduced discrepancies and employs a curriculu… view at source ↗
Figure 2
Figure 2. Figure 2: Left: Controllable curriculum data generation. We manipulate the retriever to generate training samples with gradually increasing difficulty, adaptively aligned with the model’s evolving capability, bridging the gap from pretraining to the KB-VQA target distribution. Right: Curricu￾lum sampling with observation propagation. We adaptively select informative samples likely to produce non-zero advantage durin… view at source ↗
Figure 3
Figure 3. Figure 3: Left: Number of ignored trajectories. Trajectories are ignored when they provide zero advantage and no training signal; a larger number indicates lower training efficiency. Right: Accuracy over training iterations. Performance is reported on the EVQA test set and the InfoSeek validation set. The star denotes an increase in curriculum difficulty during Wiki-R1 training. ting. These results demonstrate that … view at source ↗
Figure 4
Figure 4. Figure 4: Comparison across gap thresholds and smoothing factors. We report EVQA test and InfoSeek validation performance across training iterations under different hyperparameter settings. For the left two figures, the star denotes an increase in curriculum difficulty during Wiki-R1 training. The chosen hyperparameter τ is 0.55 and the α is 0.8 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance over training iterations for three independent runs on EVQA and InfoSeek. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. WikiCLIP: An Efficient Contrastive Baseline for Open-domain Visual Entity Recognition

    cs.CV 2026-03 unverdicted novelty 7.0

    WikiCLIP delivers an efficient contrastive baseline for open-domain visual entity recognition that improves accuracy by 16% on OVEN unseen entities and runs nearly 100 times faster than leading generative models.

  2. WikiCLIP: An Efficient Contrastive Baseline for Open-domain Visual Entity Recognition

    cs.CV 2026-03 conditional novelty 6.0

    WikiCLIP reaches 28.5% OVEN-unseen accuracy (vs 24.5% AutoVER) at 14.5 ms latency by vision-guided LLM embeddings plus hard-negative text swaps.

Reference graph

Works this paper leans on

41 extracted references · 23 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Qwen2.5-vl technical report.ArXiv, abs/2502.13923,

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report.Ar...

    Pith/arXiv arXiv

  2. [2]

    Can pre-trained vision and language models answer visual information-seeking questions? ArXiv, abs/2302.11713,

    Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions? ArXiv, abs/2302.11713,

    Pith/arXiv arXiv

  3. [3]

    Federico Cocchi, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Aug- menting multimodal llms with self-reflective tokens for knowledge-based visual question answer- ing.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9199– 9209,

    2025

  4. [4]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning.ArXiv, abs/2305.06500,

    Pith/arXiv arXiv

  5. [5]

    Bellemare, Jacob Menick, R´emi Munos, and Koray Kavukcuoglu

    Alex Graves, Marc G. Bellemare, Jacob Menick, R´emi Munos, and Koray Kavukcuoglu. Automated curriculum learning for neural networks.ArXiv, abs/1704.03003,

    Pith/arXiv arXiv

  6. [6]

    Hauptmann, Yonatan Bisk, and Jianfeng Gao

    Liangke Gui, Borui Wang, Qiuyuan Huang, Alexander G. Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language.ArXiv, abs/2112.08614,

    Pith/arXiv arXiv

  7. [7]

    Belongie, and Oisin Mac Aodha

    Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge J. Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12879–12888,

    2021

  8. [8]

    Open-domain visual entity recognition: Towards recogniz- ing millions of wikipedia entities.2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain visual entity recognition: Towards recogniz- ing millions of wikipedia entities.2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 12031–12041,

    2023

  9. [9]

    Ross, and Alireza Fathi

    Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A. Ross, and Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23369–23379,

    2023

  10. [10]

    Unsupervised dense information retrieval with contrastive learn- ing.Trans

    11 Published as a conference paper at ICLR 2026 Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learn- ing.Trans. Mach. Learn. Res., 2022,

    2026

  11. [11]

    Curriculum guided reinforcement learning for efficient multi hop retrieval augmented generation.ArXiv, abs/2505.17391,

    Yuelyu Ji, Rui Meng, Zhuochun Li, and Daqing He. Curriculum guided reinforcement learning for efficient multi hop retrieval augmented generation.ArXiv, abs/2505.17391,

    Pith/arXiv arXiv

  12. [12]

    Search- r1: Training llms to reason and leverage search engines with reinforcement learning.ArXiv, abs/2503.09516,

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Dong Wang, Hamed Zamani, and Jiawei Han. Search- r1: Training llms to reason and leverage search engines with reinforcement learning.ArXiv, abs/2503.09516,

    Pith/arXiv arXiv

  13. [13]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models.arXiv preprint arXiv:2301.12597,

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models.arXiv preprint arXiv:2301.12597,

    Pith/arXiv arXiv

  14. [14]

    Jabri, Trevor Darrell, and Pulkit Agrawal

    Rich Gang Li, A. Jabri, Trevor Darrell, and Pulkit Agrawal. Towards practical multi-object manipu- lation using relational reinforcement learning.2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 4051–4058,

    2020

  15. [15]

    Retrieval augmented visual question answering with outside knowledge

    Weizhe Lin and Bill Byrne. Retrieval augmented visual question answering with outside knowledge. arXiv preprint arXiv:2210.03809,

    Pith/arXiv arXiv

  16. [16]

    Improved baselines with visual instruction tuning.ArXiv, abs/2310.03744,

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning.ArXiv, abs/2310.03744,

    Pith/arXiv arXiv

  17. [17]

    Ok-vqa: A visual question answering benchmark requiring external knowledge.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3190–3199,

    2019

  18. [18]

    Thomas Mensink, Jasper R. R. Uijlings, Llu´ıs Castrej´on, Arushi Goel, Felipe Cadar, Howard Zhou, Fei Sha, Andre F. de Ara´ujo, and Vittorio Ferrari. Encyclopedic vqa: Visual questions about de- tailed properties of fine-grained categories.2023 IEEE/CVF International Conference on Com- puter Vision (ICCV), pp. 3090–3101,

    2023

  19. [19]

    Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    12 Published as a conference paper at ICLR 2026 Sha Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23507–23517,

    2026

  20. [20]

    Teacher algorithms for cur- riculum learning of deep rl in continuously parameterized environments.ArXiv, abs/1910.07224,

    R´emy Portelas, C´edric Colas, Katja Hofmann, and Pierre-Yves Oudeyer. Teacher algorithms for cur- riculum learning of deep rl in continuously parameterized environments.ArXiv, abs/1910.07224,

    Pith/arXiv arXiv 1910

  21. [21]

    Rora-vlm: Robust retrieval-augmented vision language models.ArXiv, abs/2410.08876,

    Jingyuan Qi, Zhiyang Xu, Rulin Shao, Yang Chen, dingnan jin, Yu Cheng, Qifan Wang, and Lifu Huang. Rora-vlm: Robust retrieval-augmented vision language models.ArXiv, abs/2410.08876,

    Pith/arXiv arXiv

  22. [22]

    Mining fine-grained image-text alignment for zero-shot captioning via text-only training.ArXiv, abs/2401.02347,

    Longtian Qiu, Shan Ning, and Xuming He. Mining fine-grained image-text alignment for zero-shot captioning via text-only training.ArXiv, abs/2401.02347,

    Pith/arXiv arXiv

  23. [23]

    Noisygrpo: Incentivizing multimodal cot reasoning via noise injection and bayesian estimation.arXiv preprint arXiv:2510.21122,

    Longtian Qiu, Shan Ning, Jiaxuan Sun, and Xuming He. Noisygrpo: Incentivizing multimodal cot reasoning via noise injection and bayesian estimation.arXiv preprint arXiv:2510.21122,

    Pith/arXiv arXiv

  24. [24]

    Da-dpo: Cost-efficient difficulty-aware preference optimization for reducing mllm hallucinations.arXiv preprint arXiv:2601.00623,

    Longtian Qiu, Shan Ning, Chuyu Zhang, Jiaxuan Sun, and Xuming He. Da-dpo: Cost-efficient difficulty-aware preference optimization for reducing mllm hallucinations.arXiv preprint arXiv:2601.00623,

    arXiv

  25. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Jun-Mei Song, Mingchuan Zhang, Y . K. Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.ArXiv, abs/2402.03300,

    Pith/arXiv arXiv

  26. [26]

    Efficient reinforcement finetun- ing via adaptive curriculum learning.ArXiv, abs/2504.05520, 2025a

    Taiwei Shi, Yiyang Wu, Linxin Song, Tianyi Zhou, and Jieyu Zhao. Efficient reinforcement finetun- ing via adaptive curriculum learning.ArXiv, abs/2504.05520, 2025a. Taiwei Shi, Yiyang Wu, Linxin Song, Tianyi Zhou, and Jieyu Zhao. Efficient reinforcement finetun- ing via adaptive curriculum learning.ArXiv, abs/2504.05520, 2025b. Krishna Srinivasan, Karthik...

    Pith/arXiv arXiv

  27. [27]

    Eva-clip: Improved training techniques for clip at scale.ArXiv, abs/2303.15389,

    Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale.ArXiv, abs/2303.15389,

    Pith/arXiv arXiv

  28. [28]

    Denny Vrandeˇci´c and Markus Kr ¨otzsch

    Accessed: 2025-09-22. Denny Vrandeˇci´c and Markus Kr ¨otzsch. Wikidata: A free collaborative knowledgebase.Commu- nications of the ACM, pp. 78–85, Sep

    2025

  29. [29]

    URLhttp://dx.doi

    doi: 10.1145/2629489. URLhttp://dx.doi. org/10.1145/2629489. 13 Published as a conference paper at ICLR 2026 Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. Fvqa: Fact-based visual question answering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 40: 2413–2427,

    doi:10.1145/2629489 2026

  30. [30]

    Rui Wang, Joel Lehman, Jeff Clune, and Kenneth O. Stanley. Paired open-ended trailblazer (poet): Endlessly generating increasingly complex and diverse learning environments and their solutions. ArXiv, abs/1901.01753,

    Pith/arXiv arXiv 1901

  31. [31]

    Zhang, Zheren Fu, and Zhendong Mao

    Shaohan Wang, L. Zhang, Zheren Fu, and Zhendong Mao. Cl-rag: Bridging the gap in retrieval- augmented generation with curriculum learning.ArXiv, abs/2505.10493, 2025a. Zhenting Wang, Guofeng Cui, Kun Wan, and Wentian Zhao. Dump: Automated distribution-level curriculum learning for rl-based llm post-training.ArXiv, abs/2504.09710, 2025b. Cong Wei, Yang Che...

    arXiv

  32. [32]

    de Ara´ujo, Bingyi Cao, and Jack Sim

    Tobias Weyand, Andre F. de Ara´ujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2 – a large-scale benchmark for instance-level recognition and retrieval.2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2572–2581,

    2020

  33. [33]

    Mmsearch- r1: Incentivizing lmms to search.ArXiv, abs/2506.20670,

    Jinming Wu, Zihao Deng, Wei Li, Yiding Liu, Bo You, Bo Li, Zejun Ma, and Ziwei Liu. Mmsearch- r1: Incentivizing lmms to search.ArXiv, abs/2506.20670,

    Pith/arXiv arXiv

  34. [34]

    A simple base- line for knowledge-based visual question answering.arXiv preprint arXiv:2310.13570,

    Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos. A simple base- line for knowledge-based visual question answering.arXiv preprint arXiv:2310.13570,

    Pith/arXiv arXiv

  35. [35]

    Echosight: Advancing visual-language models with wiki knowledge

    Yibin Yan and Weidi Xie. Echosight: Advancing visual-language models with wiki knowledge. ArXiv, abs/2407.12735,

    Pith/arXiv arXiv

  36. [36]

    Dapo: An open-source llm reinforcement learning system at scale.ArXiv, abs/2503.14476,

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Honglin Yu, Weinan Dai, Yuxuan Song, Xiang Wei, Haodong Zhou, Jingjing Liu, ...

    Pith/arXiv arXiv

  37. [37]

    Vl-cogito: Progressive curricu- lum reinforcement learning for advanced multimodal reasoning.ArXiv, abs/2507.22607,

    Ruifeng Yuan, Chenghao Xiao, Sicong Leng, Jianyu Wang, Long Li, Weiwen Xu, Hou Pong Chan, Deli Zhao, Tingyang Xu, Zhongyu Wei, Hao Zhang, and Yu Rong. Vl-cogito: Progressive curricu- lum reinforcement learning for advanced multimodal reasoning.ArXiv, abs/2507.22607,

    Pith/arXiv arXiv

  38. [38]

    A curriculum learning approach to reinforcement learning: Leveraging rag for multimodal question answering.ArXiv, abs/2508.10337,

    Chenliang Zhang, Lin Wang, Yuanyuan Lu, Yusheng Qi, Kexin Wang, Peixu Hou, and Wenshi Chen. A curriculum learning approach to reinforcement learning: Leveraging rag for multimodal question answering.ArXiv, abs/2508.10337,

    arXiv

  39. [39]

    The CLIP I-I is the retrieval with the visual similarity score from EVQA-CLIP 8B only

    14 Published as a conference paper at ICLR 2026 Table 6:Retrieval results on EVQA test and InfoSeek validation sets.We report Recall@K for K={1,5,10,20}. The CLIP I-I is the retrieval with the visual similarity score from EVQA-CLIP 8B only. Methods Retrieval ModeEVQA Test InfoSeek Val R@1 R@5 R@10 R@20 R@1 R@5 R@10 R@20 CLIP I-I Visual 11.0 26.2 33.8 41.0...

    2026

  40. [40]

    As shown in Table 7, our method requires substantially fewer training samples while achieving superior performance

    A.3 TRAININGDATASCALECOMPARISON In this section, we provide a comparison of the training data scale between our proposed framework and baseline methods. As shown in Table 7, our method requires substantially fewer training samples while achieving superior performance. This highlights the efficiency of Wiki-R1 and demonstrates its applicability in scenario...

    2026

  41. [41]

    The re- sults show that Wiki-R1 achieves competitive or lower total training costs compared to existing approaches. Given that both Wiki-LLaV A and ReflectiV A are derived from the LLaV A-1.5 architecture and did not report their training times, we estimated their training costs based on LLaV A-1.5, using the formula: Training Time≈Baseline Time×(Data Rat...

    2026