Pith. sign in

REVIEW 4 major objections 6 minor 122 references

Adding parameter-shared parallel branches to both the vision encoder and language decoder improves multimodal benchmark averages over single-branch fine-tuning at the same parameter budget.

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 ↗

ParVL scales MLLM computation by running multiple prefix-conditioned ViT and LLM branches over a shared backbone, improving average benchmark scores by 0.3 to 0.9 points and showing task-dependent vision-language allocation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible extension of parallel scaling to MLLMs; the allocation story is interesting but rests on single runs that undersell the claim. the 4 major comments →

arxiv 2608.04010 v1 pith:SP4RYAHM submitted 2026-08-04 cs.CV cs.CL

ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs

Yang Yang , Qinyu Zhao , Mouxiang Chen , Xiaohui Li , Lixin Gu , Wenhai Wang , Hongjie Zhang , Wenwei Zhang This is my paper
classification cs.CV cs.CL
keywords parallel scalingparameter sharingmultimodal large language modelscompute allocationprefix-conditioned attentionvision-language branchessupervised fine-tuningsparse routing
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.

The reading

ParVL claims that a multimodal model can spend more computation without spending more backbone parameters: instead of training one vision and one language stream, it runs several parallel streams that share the same ViT and LLM weights and are told apart only by small learned KV prefixes. On a fixed recipe of full-parameter supervised fine-tuning, the best parallel configurations beat the same-recipe single-branch baseline by 0.9 points at 1B, 0.3 at 2B, and 0.5 at 8B on a nine-benchmark average, adding at most 4% parameters. The paper's deeper claim is that the vision-language compute split is itself a tunable axis: different tasks prefer vision-heavy, language-heavy, or balanced branch counts, and no configuration wins everywhere. If true, this turns compute allocation between encoder and decoder into a design choice that can be optimized per task, rather than a fixed property of the architecture.

Core claim

The central discovery is that parallel, parameter-shared scaling works for multimodal LLMs and exposes a new allocation dimension. ParVL duplicates neither the ViT nor the LLM: it reuses one backbone per modality across Pv vision branches and Pl language branches, distinguishes branches with learned key/value prefixes, and fuses branch outputs with token-wise MLP aggregation before the connector and language-modeling head. Trained end-to-end on roughly 13B tokens, the framework improves the nine-benchmark average over same-recipe single-branch baselines at every scale tested, and the best Pv:Pl configuration changes by task and domain. The authors conclude that no single branch-count configu

What carries the argument

Prefix-conditioned attention: each parallel branch k shares the backbone query/key/value projections but prepends its own learned KV prefix to the attention keys and values, so branches specialize at negligible parameter cost. A token-wise MLP aggregator fuses branch states per token, and the branch-count pair (Pv, Pl) is the control variable for the compute-allocation question.

Load-bearing premise

The rankings rest on one training run and one evaluation per configuration, so differences of roughly 0.3 to 0.9 points could be noise.

What would settle it

Train the 1B single-branch baseline and the 4:4 configuration at least five times each with different seeds and evaluate deterministically; if the seed-to-seed spread of the nine-benchmark average overlaps the observed 0.9-point gap, the claimed improvement is not established.

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

If this is right

  • At fixed backbone size, accuracy can be scaled through additional parallel computation: the largest configurations raised the nine-benchmark average while increasing total parameters by only 0.6-3.6%.
  • Because vision and language branch counts are controlled independently, the same framework can be tuned toward OCR-heavy or math-heavy workloads by choosing a different Pv:Pl.
  • Parallel branches execute concurrently, so extra inference computation can be hidden when hardware has spare parallelism; profiled 4:4 latency at batch size 1 was only 1.04x the 1:1 baseline.
  • A small learned router can activate one ViT-LLM branch pair per sample, cutting the active configuration from 4:4 to 1:1 while staying within 0.3 points of dense aggregation.
  • The pattern of gains across scales suggests the benefit is not confined to small models: the 8B balanced configuration improved General, Math, and OCR domain groups.

Where Pith is reading between the lines

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

  • A multi-seed replication is the natural next test: the main ranking rests on single training runs and single evaluations, and 0.3-0.9 point gaps are exactly the size where seed noise could change conclusions.
  • If the allocation preferences are stable, scaling studies for multimodal models should treat vision-versus-language compute as a third scaling axis alongside parameters and tokens, with its own task-dependent optimum.
  • Since branches share all backbone weights, the mechanism may be closer to a trained ensemble than to added capacity; measuring representation diversity across branches would test whether gains come from specialization or from a smoothing effect.
  • The paper only fine-tunes on a 1/20 data subsample; scaling the training data upward is a testable way to see whether parallel allocation helps general reasoning as much as it helps math and OCR.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 6 minor

Summary. The paper proposes ParVL, a framework that scales multimodal LLM computation by running multiple parameter-shared, prefix-conditioned branches over the same ViT and LLM backbones during full-parameter SFT. The branch counts Pv and Pl can be set independently, yielding nine vision-language allocation configurations. The main empirical claims are that, under a fixed backbone parameter budget and the same SFT recipe, ParVL improves the nine-benchmark average over a single-branch baseline (49.6→50.5 at 1B, 54.4→54.7 at 2B, 62.5→63.0 at 8B) and that the best allocation varies across tasks rather than being explained by total branch count. The paper also reports latency/memory profiling and a learned sparse router that approximates dense 4:4 aggregation while executing one branch pair. Appendix A.4 states that every model-quality result comes from one training run and one deterministic evaluation per configuration.

Significance. If the empirical claims are supported, ParVL is a meaningful contribution: it formulates an under-explored allocation problem in MLLM scaling, introduces a parameter-efficient mechanism for expanding computation along separate vision and language axes, and provides controlled same-recipe comparisons at three scales. The paper is also carefully scoped in several places: it compares against same-recipe baselines, discloses the single-run protocol, provides detailed evaluation metrics (Appendix A.3), and releases profiling protocols. The strongest limitation is that all headline accuracy claims rest on single training runs, with differences of 0.3–0.9 points on overall averages and 0.1–4.1 points on individual benchmarks. Because these gaps are comparable to typical seed-level variance in full-parameter SFT, the central quantitative conclusions are not yet established to the standard required by the claims.

major comments (4)
  1. [Appendix A.4 / Table 1] Every reported accuracy result comes from one training run per configuration, and deterministic decoding only eliminates evaluation sampling noise, not training stochasticity. The headline claims—improving the overall average from 49.6 to 50.5 at 1B, 54.4 to 54.7 at 2B, and 62.5 to 63.0 at 8B—are gaps of 0.3–0.9 points. These are within the range of run-to-run variation expected for full-parameter SFT at these scales. Without multiple seeds or an explicit variance estimate, the claim that ParVL improves over the same-recipe baseline is not quantitatively supported. The bootstrap analysis in Table S5 is applied only to response length, not to any accuracy comparison.
  2. [Table S4 / Table 2] The task-specific allocation preferences are selected after viewing the full 9-cell heatmap for each benchmark. Differences between adjacent cells are often 0.1–0.4 points (e.g., TextVQA rows/cols in Table S4), and the reported task preferences (e.g., MathVision 1:4, LogicVista 4:1, MathVista 4:4) are argmax choices from the same grid used to report the claim. With one run per cell, no multiple-comparison control, and no held-out confirmation, these preferences may be noise-driven. The paper should either provide seed-averaged heatmaps with error bars or reframe the allocation claims as preliminary descriptive observations rather than conclusions.
  3. [Table 1 and Section 'Vision/language allocation'] The claim that the best vision-language allocation varies across tasks is only fully tested at 1B, where all nine Pv:Pl configurations are available. At 2B and 8B, only the balanced 2:2 configuration is compared to 1:1. Thus the conclusion in the abstract and introduction that ParVL enables a meaningful allocation trade-off is not demonstrated beyond 1B. The paper should either add the missing configurations at larger scales or explicitly restrict the allocation-variation claim to the 1B results.
  4. [Appendix A.1] The exact SFT subset is not released, and the paper states that some component datasets are not redistributable. Since all controlled comparisons use this private 13B-token subsample, independent replication is impossible without the exact data mixture. A data manifest, hash list, or a public substitute mixture would materially improve reproducibility. This is a limitation that should be addressed before the quantitative claims can be independently checked.
minor comments (6)
  1. [Appendix A.4] The single-run protocol is disclosed only in an appendix. Given that it directly qualifies the main experimental claims, it should be stated prominently in the main paper's experimental setup.
  2. [Equation (1)] The notation Q(k) suggests branch-specific query projections, but the text states that queries are not prefixed and projections are shared. Clarify that Q(k) is the same shared projection applied to the branch's hidden states, or simplify the notation.
  3. [Table 3] The row label 'Aggregation' is not immediately self-explanatory. Define it as dense all-branch inference with the learned aggregation weights, to contrast with 'Random pair' and 'Learned router'.
  4. [Figure 2] The architecture figure is dense; in particular, the distinction between 'Prefix KV' and 'Original KV' is hard to follow for readers unfamiliar with prefix-conditioned attention. Consider adding a small inset showing the masked attention matrix and where prefixes enter.
  5. [Section 'Inference Efficiency'] The heading appears as 'Sparserouting'; fix the spacing/typo to 'Sparse routing'.
  6. [Abstract] The abstract states 'modest parameter growth' without a number; the main text reports at most 4% over the evaluated configurations. Including the bound in the abstract would make the efficiency claim more precise.

Circularity Check

0 steps flagged

No circular derivation; central claims are external benchmark measurements, only a non-load-bearing self-citation to ParScale.

full rationale

The paper's central claims—that ParVL improves over same-recipe single-branch baselines and that the best Pv:Pl allocation varies across tasks—are empirical results read from Tables 1, 2, and S4 against external benchmarks (MMMU, MathVista, OCRBench, etc.). No equation or fitted constant is used to derive these scores. The method's equations (prefix-conditioned attention, MLP aggregation, SFT loss) define the architecture but do not encode the conclusions. The only self-referential component is the learned sparse router (Table 3, Appendix E), whose teacher targets are NLLs computed from the same 4:4 checkpoint; however, the paper explicitly treats this as an auxiliary exploration, discloses the procedure, and does not rely on it for the main allocation findings. The ParScale citation (Chen et al. 2025) is a design inspiration for the parallel aggregation and a fixed 48-token prefix hyperparameter; it is not invoked as evidence for ParVL's effectiveness and is not load-bearing. Appendix A.4's disclosure of one training run per configuration is a statistical robustness concern (possible seed noise on 0.3–4.1 point deltas), not circularity. No self-definition, renamed result, or uniqueness-imported-from-authors pattern is present.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The contribution is empirical, so the ledger lists the experimental load-bearing assumptions and hand-chosen hyperparameters rather than derived quantities. The most important entries are the single-run assumption and the data-subsample assumption, both of which directly affect whether the small allocation differences are real.

free parameters (4)
  • visual KV prefix length = 96 tokens per branch
    Selected after ablation over 48, 72, 96, and 128 tokens; 96 is presented as a compute-aware compromise, so the vision-heavy results depend on this choice.
  • language KV prefix length = 48 tokens
    Fixed following ParScale (Chen et al. 2025), not independently tuned here.
  • aggregation smoothing coefficients = lambda_v = 0, lambda_l = 0.1
    Set by hand for the main experiments; affects how strongly the MLP aggregator weights branches and is not independently justified.
  • branch counts Pv, Pl = varied over {1, 2, 4}
    These are the controlled independent variables, but selecting the best configuration per benchmark from observed scores introduces post-hoc choice.
axioms (3)
  • domain assumption Full-parameter SFT on a shared backbone with branch-specific prefixes preserves pretrained capabilities and lets branches specialize.
    This is the architectural premise of ParVL; the experiments test it indirectly but no separate analysis isolates interference between branch gradients.
  • domain assumption A 1/20 subsample of the InternVL3.5 SFT collection is representative enough for controlled branch-count comparisons.
    The paper uses roughly 13B tokens from a subsample; if the subsample shifts domain proportions, the allocation findings may not transfer to full-data training.
  • domain assumption Benchmark scores from one training run and one deterministic evaluation per configuration are stable enough to rank branch configurations.
    Appendix A.4 states each result comes from a single run and evaluation; the small average gains make this assumption load-bearing for the allocation conclusions.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs." pith.science (2026). https://pith.science/paper/SP4RYAHM

@misc{pith2026260804010,
  author       = {Pith},
  title        = {Pith review of: ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SP4RYAHM}},
  note         = {Machine review of arXiv:2608.04010}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. To address this, we introduce the Parallel Vision-Language (ParVL) scaling framework for MLLMs, which scales parallel computation by reusing the existing ViT and LLM backbone parameters across multiple vision and language branches. This framework raises a central question: given a fixed backbone parameter budget, how should additional shared-backbone computation be allocated between the vision and language modalities? We instantiate each parallel computational stream with branch-specific prefix parameters over a shared backbone, and train the entire model end-to-end via full-parameter supervised fine-tuning on roughly 13B tokens. We systematically study the computation-allocation trade-off between the ViT encoder and LLM decoder. ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. Code is available at https://github.com/YangYangGirl/ParVL.

Figures

Figures reproduced from arXiv: 2608.04010 by Hongjie Zhang, Lixin Gu, Mouxiang Chen, Qinyu Zhao, Wenhai Wang, Wenwei Zhang, Xiaohui Li, Yang Yang.

Figure 1
Figure 1. Figure 1: Fixed vs. expandable compute allocation in MLLMs. (Left) Existing multimodal models suffer from a fixed compute allocation between ViT (vision) and LLM (language) components during the SFT phase, which is vi￾sualized as a locked seesaw. (Right) Our ParVL frame￾work enables expandable and flexible compute for the two modules. By leveraging extensive parameter reuse, our par￾allel architecture independently … view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture. ParVL scales multimodal computation withPv parameter-shared ViT branches andPl parameter￾shared LLM branches in parallel. Learnable KV prefixes differentiate branches within each shared backbone. A token-wise MLP aggregator fuses vision states before the connector, whose projected tokens are replicated as common input to all language branches. A separate token-wise MLP aggregator fuse… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of two mechanisms for parame￾terizing parallel ViT branches in ParVL-1B under Pv = 4, Pl = 1: branch-specific KV prefixes in prefix-conditioned attention and branch-specific LoRA modules. Prefix Tokens and LoRA Rank denote the KV-prefix length per ViT branch and the rank of each LoRA module, respectively. The lan￾guage prefix remains fixed at 48 tokens (Chen et al. 2025). Vision/language allocat… view at source ↗
Figure 4
Figure 4. Figure 4: Performance–cost trade-offs at batch size = 1 for inputs containing 768 image-context tokens and 256 text tokens per sample (1024 input tokens total); generation is fixed to 128 tokens. Labels denote Pv:Pl , colors Pv, and marker shapes Pl . Dashed lines trace the performance–cost frontier in each panel. The reported 4:1 configuration uses 96 visual prefix tokens per branch. lection to 50.2, while dense 4:… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

122 extracted references · 24 canonical work pages

  1. [1]

    Wang, Ke and Pan, Junting and Shi, Weikang and Lu, Zimu and Ren, Houxing and Zhou, Aojun and Zhan, Mingjie and Li, Hongsheng , journal=

  2. [2]

    2024 , organization=

    Zhang, Renrui and Jiang, Dongzhi and Zhang, Yichi and Lin, Haokun and Guo, Ziyu and Qiu, Pengshuo and Zhou, Aojun and Lu, Pan and Chang, Kai-Wei and Qiao, Yu and others , booktitle=. 2024 , organization=

    2024

  3. [3]

    Zou, Chengke and Guo, Xingang and Yang, Rui and Zhang, Junyu and Hu, Bin and Zhang, Huan , journal=

  4. [4]

    Qiao, Runqi and Tan, Qiuna and Dong, Guanting and MinhuiWu, MinhuiWu and Sun, Chong and Song, Xiaoshuai and Wang, Jiapeng and Gongque, Zhuoma and Lei, Shanglin and Zhang, Yifan and others , booktitle=

  5. [5]

    Xiao, Yijia and Sun, Edward and Liu, Tianyu and Wang, Wei , journal=

  6. [6]

    Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T , booktitle=

  7. [7]

    Lu, Pan and Gong, Ran and Jiang, Shibiao and Qiu, Liang and Huang, Siyuan and Liang, Xiaodan and Zhu, Song-Chun , booktitle=

  8. [8]

    2021 , eprint=

    Chen, Jiaqi and Tang, Jianheng and Qin, Jinghui and Liang, Xiaodan and Liu, Lingbo and Xing, Eric P and Lin, Liang , booktitle=. 2021 , eprint=

    2021

  9. [9]

    2019 , publisher=

    Huang, Zixian and Shen, Yulin and Li, Xiao and Wei, Yuang and Cheng, Gong and Zhou, Lin and Dai, Xinyu and Qu, Yuzhong , booktitle=. 2019 , publisher=

    2019

  10. [10]

    Chen, Jiacheng and Xia, Tony and Lu, Pan and Li, Chunyuan and Gao, Jianfeng and Wang, Hanlin and Cheng, Hao , journal=

  11. [11]

    Kazemi, Mehran and Alvari, Hamidreza and Anand, Ankit and Wu, Jialin and Chen, Xi and Soricut, Radu , journal=

  12. [12]

    Kembhavi, Aniruddha and Seo, Minjoon and Schwenk, Dustin and Choi, Jonghyun and Farhadi, Ali and Hajishirzi, Hannaneh , booktitle=

  13. [13]

    arXiv preprint arXiv:2208.05358 , year=

    Lindstr. arXiv preprint arXiv:2208.05358 , year=

    Pith/arXiv arXiv

  14. [14]

    2020 , eprint=

    Methani, Nitesh and Ganguly, Pritha and Khapra, Mitesh M and Kumar, Pratyush , booktitle=. 2020 , eprint=

    2020

  15. [15]

    2023 , publisher=

    Masry, Ahmed and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , booktitle=. 2023 , publisher=

    2023

  16. [16]

    2022 , publisher=

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , booktitle=. 2022 , publisher=

    2022

  17. [17]

    2018 , eprint=

    Kafle, Kushal and Price, Brian and Cohen, Scott and Kanan, Christopher , booktitle=. 2018 , eprint=

    2018

  18. [18]

    2021 , eprint=

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C V , booktitle=. 2021 , eprint=

    2021

  19. [19]

    arXiv preprint arXiv:1710.07300 (ICLR Workshop) , year=

    Kahou, Samira Ebrahimi and Michalski, Vincent and Atkinson, Adam and K. arXiv preprint arXiv:1710.07300 (ICLR Workshop) , year=

    Pith/arXiv arXiv

  20. [20]

    Kembhavi, Aniruddha and Ilievski, Minchul and Seo, Minjoon and Farhadi, Ali and Hajishirzi, Hannaneh , booktitle=

  21. [21]

    Singh, Amanpreet and Goswami, Vivek and Natarajan, Vinay and Shah, Meet and Jiang, Yipei and Chen, Xinlei and Parikh, Devi and Batra, Dhruv , booktitle=

  22. [22]

    Lu, Pan and Bansal, Hritik and Xia, Tony and Liu, Jiacheng and Li, Chunyuan and Hajishirzi, Hannaneh and Cheng, Hao and Chang, Kai-Wei and Galley, Michel and Gao, Jianfeng , booktitle=

  23. [23]

    and Webb, W

    Lu, Pan and Mishra, Swaroop and Xia, Tony and Zhang, Liang and Chang, Kai-Wei and Zhu, Song-Chun and Goyal, Naman and Tandon, N. and Webb, W. , booktitle=

  24. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Yue, Xiang and Ni, Yuansheng and Zhang, Kai and Zheng, Tianyu and Liu, Ruoqi and Zhang, Ge and Stevens, Samuel and Jiang, Dongfu and Ren, Weiming and Sun, Yuxuan and Wei, Cong and Yu, Botao and Yuan, Ruibin and Sun, Renliang and Yin, Ming and Zheng, Boyuan and Yang, Zhenzhu and Liu, Yibo and Huang, Wenhao and Sun, Huan and Su, Yu and Chen, Wenhu , title =...

    2024

  25. [25]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  26. [26]

    URL https://github

    OpenCompass: A universal evaluation platform for foundation models, 2023 , author=. URL https://github. com/open-compass/opencompass , volume=

    2023

  27. [27]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    Suhr, Alane and Lewis, Mike and Szedl. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL) , year=

  28. [28]

    Goyal, Yash and Khot, Tejas and Summers-Stay, Douglas and Batra, Dhruv and Parikh, Devi , booktitle=

  29. [29]

    Hudson, Drew A and Manning, Christopher D , booktitle=

  30. [30]

    Qiu, Liang and Liang, Yizhou and Zhao, Yizhou and Lu, Pan and Peng, Baolin and Yu, Zhou and Wu, Ying Nian and Zhu, Song-Chun , booktitle=

  31. [31]

    arXiv preprint arXiv:2411.10442 , year=

    Enhancing the reasoning ability of multimodal large language models via mixed preference optimization , author=. arXiv preprint arXiv:2411.10442 , year=

    Pith/arXiv arXiv

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Parallel Scaling Law for Language Models , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    International Conference on Machine Learning , pages=

    Scaling laws for generative mixed-modal language models , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  34. [34]

    arXiv preprint arXiv:2010.14701 , year=

    Scaling laws for autoregressive generative modeling , author=. arXiv preprint arXiv:2010.14701 , year=

    Pith/arXiv arXiv 2010

  35. [35]

    5: Advancing open-source multimodal models in versatility, reasoning, and efficiency , author=

    Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency , author=. arXiv preprint arXiv:2508.18265 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2509.17238 , year=

    MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE , author=. arXiv preprint arXiv:2509.17238 , year=

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2509.22044 , year=

    A2R: An Asymmetric Two-Stage Reasoning Framework for Parallel Reasoning , author=. arXiv preprint arXiv:2509.22044 , year=

    work page arXiv

  38. [38]

    arXiv preprint arXiv:2509.26488 , year=

    dParallel: Learnable Parallel Decoding for dLLMs , author=. arXiv preprint arXiv:2509.26488 , year=

    arXiv

  39. [39]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    Deepprune: Parallel scaling without inter-trace redundancy , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  40. [40]

    arXiv preprint arXiv:2309.17453 , year=

    Efficient streaming language models with attention sinks , author=. arXiv preprint arXiv:2309.17453 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2510.24698 , year=

    ParallelMuse: Agentic Parallel Thinking for Deep Information Seeking , author=. arXiv preprint arXiv:2510.24698 , year=

    arXiv

  42. [42]

    arXiv preprint arXiv:2510.24824 , year=

    Parallel Loop Transformer for Efficient Test-Time Computation Scaling , author=. arXiv preprint arXiv:2510.24824 , year=

    arXiv

  43. [43]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling laws for neural language models , author=. arXiv preprint arXiv:2001.08361 , year=

    Pith/arXiv arXiv 2001

  44. [44]

    arXiv preprint arXiv:2203.15556 , year=

    Training compute-optimal large language models , author=. arXiv preprint arXiv:2203.15556 , year=

    Pith/arXiv arXiv

  45. [45]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  46. [46]

    arXiv preprint arXiv:2505.23751 , year=

    REOrdering Patches Improves Vision Models , author=. arXiv preprint arXiv:2505.23751 , year=

    work page arXiv

  47. [47]

    Machine learning , volume=

    Simple statistical gradient-following algorithms for connectionist reinforcement learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  48. [48]

    Advances in neural information processing systems , volume=

    Vmamba: Visual state space model , author=. Advances in neural information processing systems , volume=

  49. [49]

    European Conference on Computer Vision , pages=

    Motion mamba: Efficient and long sequence motion generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  50. [50]

    arXiv preprint arXiv:2401.09417 , year=

    Vision mamba: Efficient visual representation learning with bidirectional state space model , author=. arXiv preprint arXiv:2401.09417 , year=

    Pith/arXiv arXiv

  51. [51]

    European Conference on Computer Vision , pages=

    Localmamba: Visual state space model with windowed selective scan , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  52. [52]

    arXiv preprint arXiv:2403.17695 , year=

    Plainmamba: Improving non-hierarchical mamba in visual recognition , author=. arXiv preprint arXiv:2403.17695 , year=

    Pith/arXiv arXiv

  53. [53]

    European conference on computer vision , pages=

    Zigma: A dit-style zigzag mamba diffusion model , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  54. [54]

    arXiv preprint arXiv:2403.13600 , year=

    Vl-mamba: Exploring state space models for multimodal learning , author=. arXiv preprint arXiv:2403.13600 , year=

    Pith/arXiv arXiv

  55. [55]

    Wang, Weiyun and Gao, Zhangwei and Chen, Lianjie and Chen, Zhe and Zhu, Jinguo and Zhao, Xiangyu and Liu, Yangzhou and Cao, Yue and Ye, Shenglong and Zhu, Xizhou and others , journal=

  56. [56]

    Wang, Peng and Bai, Shuai and Tan, Sinan and Wang, Shijie and Fan, Zhihao and Bai, Jinze and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and others , journal=

  57. [57]

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and others , journal=

  58. [58]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  59. [59]

    Science China Information Sciences , volume=

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites , author=. Science China Information Sciences , volume=. 2024 , publisher=

    2024

  60. [60]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    Pith/arXiv arXiv

  61. [61]

    arXiv preprint arXiv:2504.10479 , year=

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models , author=. arXiv preprint arXiv:2504.10479 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:2410.18558 , year=

    Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data , author=. arXiv preprint arXiv:2410.18558 , year=

    Pith/arXiv arXiv

  63. [63]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Infographicvqa , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  64. [64]

    arXiv preprint arXiv:2305.07895 , volume=

    On the hidden mystery of ocr in large multimodal models , author=. arXiv preprint arXiv:2305.07895 , volume=

    Pith/arXiv arXiv

  65. [65]

    arXiv preprint arXiv:2404.16790 , year=

    Seed-bench-2-plus: Benchmarking multimodal large language models with text-rich visual comprehension , author=. arXiv preprint arXiv:2404.16790 , year=

    Pith/arXiv arXiv

  66. [66]

    Advances in Neural Information Processing Systems , volume=

    Charxiv: Charting gaps in realistic chart understanding in multimodal llms , author=. Advances in Neural Information Processing Systems , volume=

    work page

  67. [67]

    2024 , eprint=

    LLaVA-OneVision: Easy Visual Task Transfer , author=. 2024 , eprint=

    2024

  68. [68]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2405.20797 , year=

    Ovis: Structural Embedding Alignment for Multimodal Large Language Model , author=. arXiv preprint arXiv:2405.20797 , year=

    Pith/arXiv arXiv

  70. [70]

    The Claude 3 Model Family: Opus, Sonnet, Haiku , year =

    work page

  71. [71]

    arXiv preprint arXiv:2406.06462 , year=

    VCR: Visual Caption Restoration , author=. arXiv preprint arXiv:2406.06462 , year=

    Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2409.17146 , year=

    Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models , author=. arXiv preprint arXiv:2409.17146 , year=

    Pith/arXiv arXiv

  73. [73]

    2024 , eprint=

    How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites , author=. 2024 , eprint=

    work page 2024

  74. [74]

    2024 , eprint=

    Benchmarking Object Detectors with COCO: A New Path Forward , author=. 2024 , eprint=

    work page 2024

  75. [75]

    2019 , eprint=

    A Corpus for Reasoning About Natural Language Grounded in Photographs , author=. 2019 , eprint=

    work page 2019

  76. [76]

    2025 , eprint=

    MoEs are Stronger than you Think: Hyper-Parallel Inference Scaling with RoE , author=. 2025 , eprint=

    work page 2025

  77. [77]

    International Conference on Learning Representations (ICLR) , year=

    StagFormer: A Layer-wise Staggered Transformer for Efficent Long Sequence Modeling , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  78. [78]

    Neural computation , volume=

    Adaptive mixtures of local experts , author=. Neural computation , volume=. 1991 , publisher=

    1991

  79. [79]

    arXiv preprint arXiv:1701.06538 , year=

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer , author=. arXiv preprint arXiv:1701.06538 , year=

    Pith/arXiv arXiv

  80. [80]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the Knowledge in a Neural Network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.