Pith. sign in

REVIEW 3 major objections 5 minor 38 references

ProLaViT lets multimodal models solve complex visual reasoning by chaining progressive latent visual thoughts instead of text steps, one-shot latents, or costly image generation.

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

2026-07-12 06:12 UTC pith:NCBUK3J5

load-bearing objection Solid systems paper: progressive endogenous latent thoughts beat one-step and external-expert baselines on their suite, but the fixed synthetic templates are doing a lot of the work. the 3 major comments →

arxiv 2607.02907 v1 pith:NCBUK3J5 submitted 2026-07-03 cs.CV cs.CL

ProLaViT: Learning Progressive Latent Visual Thoughts in Structured Latent Space

Peiming Li , Yifan Wang , Xiaotian Zhang , Zhiyuan Hu , Shiyu Li , Zheng Wei , Yang Tang This is my paper
classification cs.CV cs.CL
keywords Multimodal Large Language ModelsLatent Visual ReasoningProgressive Visual DerivationEndogenous Self-DistillationDistance-Weighted Diversity LossVisual Chain-of-ThoughtStructured Latent Space
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.

Multimodal language models still fail on tasks that need multi-step spatial perception or logical verification because text reasoning loses fine visual detail and one-shot latent guesses overload capacity. ProLaViT claims the fix is to force the model to emit an ordered chain of continuous latent visual thoughts that progressively refine attention—Locate then Focus then Isolate for spatial problems, or Hypothesize then Critique then Verify for logical ones. Those latent steps are trained by endogenous self-distillation: the model’s own frozen vision encoder supplies target features extracted from programmatically synthesized intermediate images, so no external expert is required at train or test time. A distance-weighted diversity loss keeps successive steps from collapsing into identical representations. The result is higher accuracy on vision-centric benchmarks at only modest extra latency, showing that structured progressive derivation inside latent space can internalize algorithmic visual precision without tools.

Core claim

A multimodal model can perform precise multi-step visual reasoning by generating a structured progressive chain of continuous latent thoughts, supervised solely by its own frozen vision encoder on algorithmically synthesized intermediate views and kept topologically distinct by a causal-distance-weighted diversity loss; this combination outperforms both textual chain-of-thought and unstructured one-step latent prediction while remaining efficient.

What carries the argument

Progressive Latent Visual Thought (ordered latent steps such as Locate→Focus→Isolate or Hypothesize→Critique→Verify) trained by Endogenous Self-Distillation against the native encoder’s features of programmatically rendered auxiliary images, regularized by Distance-Weighted Diversity Loss that penalizes cosine similarity between step centroids in proportion to their causal distance.

Load-bearing premise

Matching the frozen encoder features of four fixed, programmatically rendered intermediate images is assumed sufficient to teach genuine multi-step visual reasoning that transfers to real images at test time.

What would settle it

If a single latent step given the same total token budget, or the same four steps presented in random order, matches or exceeds ProLaViT accuracy on VisPuzzle, ChartQA and BLINK-Jigsaw under identical training data and compute, the necessity of structured progressive derivation would be falsified.

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

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

3 major / 5 minor

Summary. ProLaViT equips MLLMs with progressive latent visual thoughts in continuous space. It decomposes reasoning into structured chains (Locate o Focus o Isolate for spatial tasks; Hypothesize o Critique o Verify for logical ones) realized by special tokens. Supervision uses endogenous self-distillation: the model’s own frozen vision encoder extracts features from programmatically synthesized auxiliary images (grid/bbox/crop/seg or edge/arrow/counterfactual/verify) generated via a Gemini-structured + deterministic rendering pipeline. A Distance-Weighted Diversity Loss (Eqs. 7–9) regularizes step centroids by causal distance to avoid latent collapse. Trained with multi-stage curriculum + LoRA on Qwen2.5-VL-7B, it reports SOTA averages (75.11 %) on MMVP/VisPuzzle/VStar/ChartQA/BLINK/CV-Bench versus text CoT, one-step latent, CoVT and LVR, with only 1.21 imes latency.

Significance. If the progressive endogenous mechanism generalizes, the work offers a practical middle path between costly explicit image generation and unstructured latent guessing, removing external expert dependencies while preserving visual fidelity. Strengths include the clean endogenous teacher design, the topology-aware diversity loss, the multi-stage curriculum, and the suite of targeted ablations (progressive vs. one-step, margin vs. distance-weighted, endogenous vs. DINO/VAE, order permutation, text-trajectory control, tool-use latency). These controls make the efficiency and accuracy claims on the chosen suite credible and reproducible in principle.

major comments (3)
  1. [§3.2–3.3, Supp. B, Tables 2–5] §3.2–3.3 and Supplementary B: the four fixed, hand-designed operations (GridView/BBox/Crop/Seg or Edge/Arrow/Counterfactual/Verify) are synthesized from Gemini CoTs on Chart Refocus / Jigsaw / Visual Search and then rendered deterministically. All ablations (Tables 2–5) only permute, remove or re-weight these same templates; none test whether the learned latent chain remains useful when intermediate visual states deviate (different crop ratios, non-grid spatial cues, non-jigsaw counterfactuals, or natural images lacking clean geometric annotations). Consequently the large lifts on BLINK-Jigsaw (+16.67 %) and ChartQA may reflect successful imitation of the training renderings rather than acquisition of a general progressive derivation capability. A load-bearing experiment with out-of-template intermediates or held-out operation types is required to support the central claim.
  2. [Table 1, §4.2] Table 1 and all subsequent tables report single-run point estimates with no error bars, multiple random seeds, or statistical tests. Several claimed gains are small (e.g., MMVP 79.00 vs. 79.33, VStar 80.11 vs. 79.05). Without variance estimates it is impossible to judge whether the 75.11 % average and the ranking over One-step Latent Pred. / CoVT are robust. At minimum, three seeds with standard deviations (or bootstrap CIs) should be supplied for the main table.
  3. [Supp. B, §4.1] Supplementary B describes the programmatic pipeline but never states the number of synthesized trajectories, the success rate of Gemini structuring, the filtering criteria, or the exact train/validation split sizes per domain. Because the same vision encoder that produces the teacher features is also the student backbone, mild domain circularity is possible if the evaluation benchmarks share visual statistics with the three synthesis domains. Explicit data-volume and overlap statistics are needed to assess whether the reported gains are data-scale or method-driven.
minor comments (5)
  1. [Fig. 4, §4.3] Fig. 4 similarity matrices are informative but lack quantitative summary statistics (mean off-diagonal cosine, effective rank) that would make the ‘latent collapse’ claim more precise.
  2. [§3.4, Eq. (8–9)] Eq. (8) introduces a learnable steepness α = 1 + softplus(θα) and a learnable margin τ; the free-parameter count should be listed explicitly alongside λ_div and λ_distill for reproducibility.
  3. [Supp. A, Table 7] The two-round curriculum in Supp. A (Stage 0–2 then restructured stages) is useful but the exact step ranges and loss schedules are only partially tabulated; a single consolidated hyper-parameter table would help.
  4. [Abstract, §1] Several concatenated words appear in the extracted text (“MultimodalLargeLanguageModels”, “EndogenousSelf-Distillation”); if present in the PDF they should be corrected for readability.
  5. [Table 1] BLINK-JA vg. column header in Table 1 is truncated; expand to the full benchmark name for clarity.

Circularity Check

0 steps flagged

No circularity: ordinary supervised endogenous distillation on synthetic trajectories, evaluated on external benchmarks.

full rationale

The paper presents an empirical ML method (ProLaViT) whose core is training latent tokens via MSE distillation (Eq. 6) against features of programmatically rendered auxiliary images extracted by the model's own frozen vision encoder, plus a distance-weighted diversity regularizer (Eq. 9) and standard LM loss. The synthetic operations (GridView/BBox/Crop/Seg or Edge/Arrow/Counterfactual/Verify) and the teacher features are construction inputs to the training objective; they are not redefined in terms of the final benchmark numbers, nor is any parameter fitted to a subset of the evaluation metrics and then re-labeled a prediction. Ablations (Tables 2-5) and comparisons (Table 1) are run on held-out vision-centric benchmarks (MMVP, VisPuzzle, VStar, ChartQA, BLINK, CV-Bench). There are no load-bearing self-citations of uniqueness theorems, no ansatz smuggled via prior author work, and no renaming of a known empirical pattern as a first-principles derivation. The mild dependence on author-chosen templates is a generalization risk, not a circular reduction of the claimed result to its inputs by construction. The derivation chain is therefore self-contained supervised learning with external evaluation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 3 invented entities

The central empirical claim rests on a small set of free hyper-parameters (loss weights, margin, step count, token count), standard MLLM training assumptions, and two invented constructs (the progressive latent thought tokens and the distance-weighted diversity loss) whose only evidence is the reported ablations. No external physical or mathematical constants are introduced.

free parameters (6)
  • lambda_div = 0.2
    Weight of the diversity loss; set to 0.2 by authors.
  • lambda_distill = 1.0
    Weight of the self-distillation MSE; set to 1.0.
  • margin tau = 0.8 (init)
    Cosine-similarity threshold in diversity loss; initialized to 0.8 and described as learnable.
  • K (reasoning steps) = 4
    Number of progressive latent steps; fixed at 4.
  • N_tokens per step = 4
    Number of special latent tokens per reasoning step; fixed at 4.
  • alpha (distance steepness) = 1 + softplus(theta_alpha)
    Learnable exponent in the causal-distance weight w_ij = epsilon + (1-epsilon)*(d/dmax)^alpha.
axioms (4)
  • domain assumption Matching features of the model's own frozen vision encoder on programmatically transformed images is a valid supervisory signal for intermediate latent thoughts.
    Core of the endogenous self-distillation mechanism (§3.3).
  • ad hoc to paper The four hand-designed visual operations (grid/bbox/crop/seg or edge/arrow/counterfactual/verify) form a sufficient causal chain for the evaluated spatial and logical tasks.
    Introduced in §3.2 and Fig. 3; no proof of completeness.
  • domain assumption Standard cross-entropy language-modeling loss plus MSE distillation plus diversity loss is an adequate multi-task objective for joint reasoning and answering.
    Eq. (10) and curriculum phases.
  • domain assumption LoRA fine-tuning of the LLM backbone while freezing the vision encoder preserves general VQA capability when mixed with general data in Phase 4.
    Training strategy §3.5 and Supplementary A.
invented entities (3)
  • Progressive Latent Visual Thought tokens (<|vit_pad_latent|>) no independent evidence
    purpose: Continuous latent states that stand in for intermediate visual operations inside the LLM embedding space.
    Defined in §3.2; no independent existence outside the training objective.
  • Distance-Weighted Diversity Loss no independent evidence
    purpose: Topology-aware regularizer that penalizes cosine similarity between step centroids weighted by causal distance to prevent latent collapse.
    Eqs. (7)–(9); introduced specifically for this paper.
  • Endogenous Self-Distillation teacher (same frozen ViT) no independent evidence
    purpose: Provides supervision without external experts by reusing the MLLM's own vision encoder on synthetic images.
    §3.3; the 'endogenous' framing is the paper's contribution relative to CoVT-style external experts.

pith-pipeline@v1.1.0-grok45 · 19465 in / 3281 out tokens · 26059 ms · 2026-07-12T06:12:05.959706+00:00 · methodology

0 comments
read the original abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but still struggle with complex visual reasoning tasks requiring multi-step perception and logical deduction. While explicit visual generation incurs prohibitive computational costs, existing latent approaches often rely on external experts or lack rigorous cognitive logic. In this paper, we introduce ProLaViT (Progressive Latent Visual Thought), a framework empowering MLLMs to perform structured visual derivation in the continuous latent space. Unlike works dependent on heterogeneous external models, ProLaViT leverages an endogenous self-distillation mechanism, utilizing the model's own visual encoder to supervise latent thoughts. To facilitate this, we construct a scalable programmatic synthesis pipeline enabling the model to internalize algorithmic precision without inference time tools. We design two reasoning paradigms: (1) Coarse-to-Fine Causal Chain for spatial tasks, guiding attention from global context to local targets. (2) Dialectical Reasoning Chain for logical tasks, incorporating counter-factual thinking for verification. Furthermore, we propose a Distance-Weighted Diversity Loss to impose topology-aware constraints, preventing feature degeneration by enforcing semantic distinctiveness. Extensive experiments demonstrate that ProLaViT outperforms baselines on vision-centric benchmarks, achieving superior accuracy and interpretability with high efficiency.

Figures

Figures reproduced from arXiv: 2607.02907 by Peiming Li, Shiyu Li, Xiaotian Zhang, Yang Tang, Yifan Wang, Zheng Wei, Zhiyuan Hu.

Figure 1
Figure 1. Figure 1: Comparison of multimodal reasoning paradigms. Top: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ProLaViT. The framework generates progressive latent thoughts (green tokens) supervised by an Endogenous Self-Distillation mechanism (Right), where latent states are aligned with visual features from synthesized auxil￾iary images via LMSE. To prevent representation degeneration, a Distance-Weighted Diversity Loss (Ldiv, Left) penalizes excessive similarity between reasoning steps, en￾forcing to… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the two reasoning paradigms in ProLaViT. Top: [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Such convergence risks degrading the progressive reasoning chain into [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of latent thought similarity. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] 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

38 extracted references · 15 linked inside Pith

  1. [1]

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)

    2023

  2. [2]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S.: Qwen2.5-vl technical report (2025)

    2025

  3. [3]

    arXiv preprint arXiv:2407.06135 (2024)

    Chern, E., Su, J., Ma, Y., Liu, P.: Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135 (2024)

    Pith/arXiv arXiv 2024

  4. [4]

    arXiv preprint arXiv:2505.14683 (2025)

    Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., et al.: Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    In: The Twelfth International Conference on Learning Representations (2024)

    Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., Kong, X., Zhang, X., Ma, K., Yi, L.: DreamLLM: Synergistic multimodal comprehension and creation. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  6. [6]

    arXiv preprint arXiv:2512.05665 (2025)

    Dong, S., Wang, S., Liu, X., Li, C., Hou, H., Wei, Z.: Interleaved latent visual rea- soning with selective perceptual modeling. arXiv preprint arXiv:2512.05665 (2025)

    arXiv 2025

  7. [7]

    ArXivabs/2404.12390(2024)

    Fu, X., Hu, Y., Li, B., Feng, Y., Wang, H., Lin, X., Roth, D., Smith, N.A., Ma, W.C., Krishna, R.: Blink: Multimodal large language models can see but not per- ceive. ArXivabs/2404.12390(2024)

    Pith/arXiv arXiv 2024

  8. [8]

    arXiv preprint arXiv:2310.01218 (2023)

    Ge, Y., Zhao, S., Zeng, Z., Ge, Y., Li, C., Wang, X., Shan, Y.: Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218 (2023)

    Pith/arXiv arXiv 2023

  9. [9]

    Gu, J., Hao, Y., Wang, H.W., Li, L., Shieh, M.Q., Choi, Y., Krishna, R., Cheng, Y.: Thinkmorph: Emergent properties in multimodal interleaved chain-of-thought reasoning (2025)

    2025

  10. [10]

    ArXiv abs/2412.06769(2024)

    Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J.E., Tian, Y.: Training large language models to reason in a continuous latent space. ArXiv abs/2412.06769(2024)

    Pith/arXiv arXiv 2024

  11. [11]

    arXiv preprint arXiv:2501.05444 (2025)

    Hao, Y., Gu, J., Wang, H.W., Li, L., Yang, Z., Wang, L., Cheng, Y.: Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark. arXiv preprint arXiv:2501.05444 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    arXiv preprint arXiv:2502.09621 (2025)

    Jiang, D., Zhang, R., Guo, Z., Li, Y., Qi, Y., Chen, X., Wang, L., Jin, J., Guo, C., Yan, S., et al.: Mme-cot: Benchmarking chain-of-thought in large mul- timodal models for reasoning quality, robustness, and efficiency. arXiv preprint arXiv:2502.09621 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.B.: Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 3992– 4003 (2023)

    2023

  14. [14]

    Koh, J.Y., Fried, D., Salakhutdinov, R.: Generating images with multimodal lan- guage models (2023)

    2023

  15. [15]

    arXiv preprint arXiv:2507.16746 (2025)

    Li, A., Wang, C., Fu, D., Yue, K., Cai, Z., Zhu, W.B., Liu, O., Guo, P., Neiswanger, W., Huang, F., et al.: Zebra-cot: A dataset for interleaved vision language reason- ing. arXiv preprint arXiv:2507.16746 (2025)

    arXiv 2025

  16. [16]

    ArXivabs/2509.24251(2025)

    Li, B., Sun, X., Liu, J., Wang, Z., Wu, J., Yu, X., Chen, H., Barsoum, E., Chen, M., Liu, Z.: Latent visual reasoning. ArXivabs/2509.24251(2025)

    Pith/arXiv arXiv 2025

  17. [17]

    ArXiv abs/2304.08485(2023) ProLaViT: Progressive Latent Visual Thoughts 17

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. ArXiv abs/2304.08485(2023) ProLaViT: Progressive Latent Visual Thoughts 17

    Pith/arXiv arXiv 2023

  18. [18]

    ArXivabs/2304.09842(2023)

    Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. ArXivabs/2304.09842(2023)

    Pith/arXiv arXiv 2023

  19. [19]

    ArXiv abs/2203.10244(2022)

    Masry, A., Long, D.X., Tan, J.Q., Joty, S.R., Hoque, E.: Chartqa: A benchmark for question answering about charts with visual and logical reasoning. ArXiv abs/2203.10244(2022)

    Pith/arXiv arXiv 2022

  20. [20]

    OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L.: Gpt-4 technical report (2024)

    2024

  21. [21]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y.B., Li, S.W., Misra, I., Rabbat, M.G., Sharma, V., Syn- naeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Di- nov2:Learningrobustvisual...

    Pith/arXiv arXiv 2023

  22. [22]

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)

    2023

  23. [23]

    Qin, Y., Wei, B., Ge, J., Kallidromitis, K., Fu, S., Darrell, T., Wang, X.: Chain- of-visual-thought: Teaching vlms to see and think better with continuous visual tokens (2025)

    2025

  24. [24]

    Advances in Neural Information Processing Systems 37 (2024)

    Shao, H., Qian, S., Xiao, H., Song, G., Zong, Z., Wang, L., Liu, Y., Li, H.: Vi- sual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. Advances in Neural Information Processing Systems 37 (2024)

    2024

  25. [25]

    Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face (2023)

    2023

  26. [26]

    arXiv preprint arXiv:2502.21074 (2025)

    Shen, Z., Yan, H., Zhang, L., Hu, Z., Du, Y., He, Y.: Codi: Compressing chain-of- thought into continuous space via self-distillation. arXiv preprint arXiv:2502.21074 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning (2023)

    2023

  28. [28]

    Team, C.: Cvbench: A benchmark for cross-video multimodal reasoning (2025)

    2025

  29. [29]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Tong,S.,Liu,Z.,Zhai,Y.,Ma,Y.,LeCun,Y.,Xie,S.:Eyeswideshut?exploringthe visual shortcomings of multimodal llms. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9568–9578 (2024)

    2024

  30. [30]

    arXiv preprint arXiv:2511.21395 (2025)

    Wang, Q., Shi, Y., Wang, Y., Zhang, Y., Wan, P., Gai, K., Ying, X., Wang, Y.: Monet: Reasoning in latent visual space beyond images and language. arXiv preprint arXiv:2511.21395 (2025)

    arXiv 2025

  31. [31]

    ArXiv abs/2201.11903(2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E.H., Xia, F., Le, Q., Zhou, D.: Chain of thought prompting elicits reasoning in large language models. ArXiv abs/2201.11903(2022)

    Pith/arXiv arXiv 2022

  32. [32]

    Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models (2023)

    2023

  33. [33]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Wu, P., Xie, S.: V*: Guided visual search as a core mechanism in multimodal llms. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 13084–13094 (2023)

    2024

  34. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Xu, G., Jin, P., Wu, Z., Li, H., Song, Y., Sun, L., Yuan, L.: Llava-cot: Let vision lan- guage models reason step-by-step. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2087–2098 (October 2025) 18 P. Li et al

    2087

  35. [35]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10371–10381 (2024)

    2024

  36. [36]

    ArXivabs/2506.17218(2025)

    Yang, Z., Yu, X., Chen, D., Shen, M., Gan, C.: Machine mental imagery: Empower multimodal reasoning with latent visual tokens. ArXivabs/2506.17218(2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action (2023)

    2023

  38. [38]

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain- of-thought reasoning in language models (2024) ProLaViT: Progressive Latent Visual Thoughts 1 ProLaViT: Learning Progressive Latent Visual Thoughts in Structured Latent Space Supplementary Material A Implementation Details Base Model and Parameter-Efficient Tuning.We adopt Q...

    2024