Pith. sign in

REVIEW 5 major objections 4 minor 18 references

When and What: Diffusion-Grounded VideoLLM with Entity Aware Segmentation for Long Video Understanding

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A Video-LLM that extracts temporal features with a diffusion denoiser and grounds query entities with pre-reasoning segmentation claims state-of-the-art results on long-video grounding and video QA benchmarks.

desk verdict The architecture combination is plausible, but the paper's own Table I contradicts its state-of-the-art claim, and Table II's baselines are placeholders; the empirical backbone doesn't hold. read the letter →

arxiv 2508.15641 v1 pith:UNYQAZRL submitted 2025-08-21 cs.CV

Pengcheng Fang , Yuxia Chen , Rui Guo This is my paper
classification cs.CV
keywords VideoLLMtemporalgroundinglongunderstandingdiffusionlatententity-awaresegmentationgroundedquestionansweringopen-vocabularydetectionmixed-tokensequence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Understanding a video well enough to answer 'when' and 'what' requires a model to know both the timing of an event and which entities are involved. This paper argues that current Video-LLMs handle timing only implicitly and grounding only after language generation, which makes them imprecise at strict temporal resolutions. To fix this, Grounded-VideoDiT (a) turns a video diffusion model into a temporal feature extractor conditioned on object masks, (b) segments and tracks the nouns in a query before any language reasoning, and (c) inserts discrete timestamp tokens into the LLM input sequence. The paper claims these changes deliver state-of-the-art results on Charades-STA, DiDeMo, NExT-GQA, and multiple open-ended VideoQA benchmarks, with the largest gains at high IoU thresholds. A reader should care because the design points toward video reasoning that is both temporally precise and verifiable, since each answer can be tied to visible mask evidence and a time interval.

What carries the argument

Diffusion Temporal Latent (DTL) encoder — a masked video diffusion backbone (Wan) queried at an early denoising step to produce compact spatiotemporal tokens from frame latents conditioned on object masks. This carries the temporal half of the argument: intermediate denoising states encode boundary-sensitive motion and appearance cues that static image encoders miss. The other half is the entity-grounded token stream: Grounding-SAM2 with DINO-based tracking produces per-noun binary masks that enter the diffusion conditioning and also serve as explicit object tokens in the mixed-token input, carrying the 'what' half. A mixed-token strategy interleaves these object tokens with discrete timesta

What would settle it

Perform an independent evaluation on Charades-STA using the official code and metrics, comparing the released model with LLaVA-ST; if LLaVA-ST's mIoU is still higher, the paper's claim of state-of-the-art temporal grounding is not supported.

Watch

Extended reading notes

Core claim

Grounded-VideoDiT pairs a frozen LLM with a Grounding-SAM2 detector and a video diffusion encoder. Nouns in the query are detected frame-by-frame, gated on co-occurrence, and tracked to produce binary masks that condition the diffusion encoder. That encoder, sampled at an early denoising step, yields a compact temporal latent token for the LLM. Timestamps become sinusoidal embeddings; text, video, mask, and time tokens are interleaved in one mixed sequence. A KL loss aligns the diffusion features with a stronger auxiliary encoder. The paper reports best-in-benchmark results for temporal grounding, grounded video QA, and open-ended VideoQA, attributing the gains to all three components.

Load-bearing premise

If the mask tracking ever drifts from the true entity, every downstream token built from those masks is corrupted, so the whole pipeline's accuracy is capped by the tracker's reliability.

Editorial extensions

If this is right

  • Temporal grounding at strict IoU thresholds becomes practical, so long-video retrieval and highlight detection can pinpoint event boundaries rather than rough segments.
  • Grounded video QA answers come with temporal evidence, making model outputs auditable: a user can jump to the exact frames that justify an answer.
  • Diffusion models can serve as reusable video feature extractors for LLM pipelines, decoupling temporal representation learning from text generation.
  • Because grounding happens before language modeling, the masks themselves can be visualized as the model's entity attention, supporting debugging of multi-entity questions.

Reading between the lines

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

  • If the diffusion-as-feature-extractor idea generalizes, the same early-step denoising trick could be applied to audio or sensor streams for temporal event detection, though the paper only tests video.
  • The AND-gated persistence rule presupposes that every noun in the query is visible simultaneously near the start; for queries where entities interact only after one enters later, the rule would need a relaxation the paper does not explore.
  • The paper's ablation attributes gains to the three components but does not isolate the contribution of mask quality; a study replacing Grounding-SAM2 with oracle masks (or a weaker detector) would reveal how much headroom remains in entity grounding.
  • The claimed state-of-the-art margins depend on the specific baselines listed; re-running with recently updated LLM backbones could shift the comparisons, though the qualitative examples suggest the mechanism itself is the differentiator.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes Grounded-VideoDiT, a Video-LLM for long-video understanding that combines three components: (i) a Diffusion Temporal Latent (DTL) encoder that uses an inpainting-capable latent diffusion model as a temporal feature extractor, (ii) object-grounded representations obtained by parsing noun phrases from the query and running Grounded-SAM2 with DINO-based tracking to produce masks that condition the diffusion encoder and are injected as object tokens, and (iii) a mixed-token input scheme with discrete timestamp tokens, fused with sinusoidal temporal encodings and LoRA-adapted LLM. The central claim, stated in the Abstract and Contribution 4, is state-of-the-art performance on Charades-STA, NExT-GQA, NExT-QA, and other VideoQA benchmarks. Experiments are reported on temporal grounding, grounded VideoQA, and open-ended VideoQA. The paper also includes ablations of components, token budgets, and diffusion hyperparameters.

Significance. If the empirical claims were correct, the paper would make a useful contribution: repurposing a video inpainting diffusion model as a temporal encoder and injecting segmentation masks before language modeling is a timely and potentially impactful idea. The proposed architecture is plausible, and the emphasis on explicit entity tracking and timestamp tokens addresses real limitations of current Video-LLMs. However, the significance is entirely contingent on the validity of the reported results, and the paper's own tables contradict the central SOTA claim. The reported numbers also exhibit internal inconsistencies that prevent the reader from trusting any of the quantitative conclusions. As submitted, the paper does not provide a reliable validation of its proposed method, so its significance to the field is currently not established.

major comments (5)
  1. [Abstract, Contribution 4, §IV-B1, Table I] The central claim of state-of-the-art performance is contradicted by the paper's own Table I. On Charades-STA, LLaVA-ST achieves R@0.3/0.5/0.7 = 63.1/44.8/23.4 and mIoU 42.4, while Ours achieves 58.7/41.2/21.0 and 39.5. On DiDeMo, LLaVA-ST achieves 56.2/39.8/20.1 and 37.6 versus Ours 53.0/37.0/18.5 and 35.2. Thus the statement in §IV-B1 that Ours 'consistently improves over prior Video-LLMs' is false on the reported numbers. Since the SOTA claim is load-bearing for the paper, this internal contradiction is a fundamental flaw.
  2. [§IV-A, Table II] Table II explicitly labels all NExT-QA entries except Ours as 'placeholder values pending actual evaluation.' This means the claimed 56.9 Acc on NExT-QA has no verified baselines in the table, so the statement in §IV-B3 that 'our model achieves the best overall performance across four benchmarks' is unsupported for NExT-QA. Furthermore, Sec. IV promises 'mean ± std over 3 seeds,' but no variance is reported in any table, and the Appendix is referenced multiple times but is absent from the manuscript. These omissions make the empirical evaluation impossible to verify.
  3. [Table I vs. Tables IV–VI] There are direct numerical inconsistencies between the main results and the ablations. Table I reports Ours as R@0.3=58.7, R@0.5=41.2, mIoU=39.5 on Charades-STA. Table IV's full model (DTL+Obj+Time) matches these values. However, Table V reports that the recommended 4 Obj / 8 Time token budget yields R@0.3=53.8, R@0.5=36.9, mIoU=34.5 on the same dataset, and Table VI reports the default S=4, cosine, GS=1.0 configuration as R@0.5=36.9 and mIoU=34.5. Unless different evaluation protocols are implied but not disclosed, the same configuration cannot produce two different scores. This inconsistency undermines the reliability of the reported experimental results as a whole.
  4. [§IV-B2, Table III] The text states that Ours achieves 'the highest IoP@0.3 (45.1),' but Table III contains no IoP@0.3 column; it reports IoP@0.5 and mIoP. The cited value 45.1 is therefore unverifiable from the presented data. Additionally, the claimed best Acc@GQA of 28.4 is only 1.7 points above Grounded-VideoLLM, and no variance or significance testing is provided despite the promise of 3-seed mean±std. This further weakens confidence in the grounded-QA result.
  5. [§III-A, Eqs. (1)–(6)] The entire pipeline depends on the assumption that Grounding-SAM2 with DINO-based tracking produces correct, temporally consistent masks for every noun extracted from the query. These masks are used as hard inpainting conditions (Eq. 8) and as object tokens in the mixed sequence. If tracking drifts—which is common in long videos with occlusion, reappearance, or multiple similar entities—the diffusion encoder and object tokens are corrupted for all downstream reasoning. The paper provides no analysis of tracking failure rates, no ablation of mask quality, and no discussion of how tracking errors propagate. This is a correctness risk that should be addressed even if the empirical numbers were consistent.
minor comments (4)
  1. [Fig. 1, Sec. IV-A] There is a typo in the teaser: 'When dose the dog' should be 'When does the dog.' Also, the implementation states the LLM backbone is Phi-3.5-Vision-Instruct-3.8B, but reference [17] is cited as 'Phi-4 Technical Report'; please align the citation.
  2. [Table II] The baseline 'Momcnet' appears to be a misspelling of 'Momenter' (as in Table I). Please check the model names used in the tables.
  3. [§IV-C1] Table IV is missing column headers for the first three columns on the left; the checkmarks and blank entries are ambiguous without explicit column labels. Also, the text says 'Removing either object or time tokens from the full model leads to consistent degradation,' but the table rows with only one component removed are not clearly identified.
  4. [§III-C, Eq. (13)] The fusion equation concatenates [zt; etext; etime(t)], where zt is defined in Sec. III-B as a per-frame embedding. However, Eq. (9) defines z as a single vector after pooling over the whole video, creating ambiguity about how per-frame embeddings are obtained. Please clarify whether zt is a frame-level feature or a duplicated global vector.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the central claims rest on empirical benchmarking, not derivation-by-construction.

full rationale

The paper contains no self-citation chain and no equation reduces a claimed prediction to its input by construction. The Diffusion Temporal Latent encoder, Grounding-SAM2 masks, and mixed token scheme are all input-side features fed into a frozen LLM; the grounding outputs (start/end times, QA answers) are produced by learned classifiers over the LLM's hidden states, not by re-reading the masks or prompt. Hyperparameters such as S=4, GS=1.0, and the 4-obj/8-time token budget are selected via ablations on evaluation benchmarks (Tables V-VI), which is standard empirical tuning, not a fitted parameter being renamed as a prediction. The main red flags are correctness/consistency issues, not circularity: Table I shows LLaVA-ST outperforming 'Ours' on Charades-STA and DiDeMo at every IoU threshold, contradicting the Abstract's 'state-of-the-art' and §IV-B1's 'consistently improving over prior Video-LLMs'; Table II explicitly labels all non-Ours NExT-QA entries as 'placeholder values pending actual evaluation'; Table III text cites IoP@0.3=45.1 though no such column is shown; and the promised Appendix with mean±std over 3 seeds is absent. These undermine the empirical claim but do not constitute circular reasoning.

Assumptions & free parameters 8 free parameters · 7 assumptions · 2 invented entities

The central claims depend on a chain of hand-designed heuristics (detection thresholds, persistence window, early diffusion step, token budgets) and on external modules whose reliability is assumed. Most thresholds are never reported. These choices are load-bearing but unquantified.

free parameters (8)
  • Object detection confidence threshold tau_i = not reported
    Used in Eq. 2 to decide when a noun is detected; value is presumably tuned but never specified.
  • Persistence window K = not reported
    Eq. 4 requires all nouns detected over K consecutive frames; K is a hand-chosen heuristic and never reported.
  • Minimum span length L = not reported
    Step 5 filters candidate intervals by minimum length; L is never specified.
  • Early diffusion step tau0 = not reported
    Section III-B queries the denoiser at an early step; this critical hyperparameter is never reported.
  • Denoising steps S = S=4
    Selected via ablation on Charades-STA in Table VI as the best trade-off.
  • Guidance scale GS = 1.0
    Selected via ablation on Charades-STA in Table VI.
  • Object/time token budget = 4 object, 8 time
    Selected via ablation on Charades-STA in Table V.
  • LoRA rank r and alpha = r=64, alpha=128
    Chosen for efficient adaptation; no justification or sensitivity analysis is provided.
assumptions (7)
  • domain assumption Grounding-SAM2 open-vocabulary scores rank the correct object proposal highest for each noun phrase.
    Eq. 2 assumes argmax over scores selects the true entity; no evidence is given for long videos with occlusion or compositional prompts.
  • ad hoc to paper AND-gated persistence over K frames prevents premature or partial tracking starts.
    Eq. 3-4 is a hand-designed heuristic; the paper provides no validation that this rule is superior to alternatives.
  • domain assumption Early-step denoising features from a masked inpainting diffusion model encode temporal dynamics and object-focused appearance.
    Section III-B asserts this property without a theorem or a controlled experiment isolating the mechanism.
  • domain assumption Mask conditioning in the diffusion encoder preserves object-centric focus without destroying temporal continuity.
    Section III-B(c) assumes compatibility between binary masks and inpainting-conditioned diffusion latents.
  • ad hoc to paper KL alignment to an auxiliary feature extractor improves discriminability for temporal grounding.
    Section III-D motivates the regularization only by claimed empirical gains; no analytic or external evidence is provided.
  • domain assumption A frozen Phi-3.5-Vision-Instruct-3.8B LLM adapted with LoRA is sufficient for temporal reasoning.
    Section IV-A fixes the backbone without comparing alternatives such as larger LLMs or different vision backbones.
  • domain assumption Official splits, evaluation scripts, and baseline results are correctly reproduced.
    Section IV states official scripts are used, but Table II contains placeholder baseline values, undermining this assumption.
invented entities (2)
  • Diffusion Temporal Latent (DTL) encoder
    purpose: Produces temporal latent tokens by querying a masked inpainting diffusion model at an early denoising step, then pooling and projecting.
    Its claimed benefit (boundary sensitivity, temporal consistency) is only supported by the paper's own ablations, which are not externally verifiable and are partly contradicted by the main results.
  • Segmentation-guided object embeddings (Obj-Seg)
    purpose: Binds query noun phrases to segmentation masks and feeds temporally tracked object tokens into the LLM.
    Depends on Grounding-SAM2 and DINO-based tracking; the embedding scheme itself has no separate falsifiable handle outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When and What: Diffusion-Grounded VideoLLM with Entity Aware Segmentation for Long Video Understanding." pith.science (2026). https://pith.science/paper/UNYQAZRL

@misc{pith2026250815641,
  author       = {Pith},
  title        = {Pith review of: When and What: Diffusion-Grounded VideoLLM with Entity Aware Segmentation for Long Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNYQAZRL}},
  note         = {Machine review of arXiv:2508.15641}
}
read the original abstract

Understanding videos requires more than answering open ended questions, it demands the ability to pinpoint when events occur and how entities interact across time. While recent Video LLMs have achieved remarkable progress in holistic reasoning, they remain coarse in temporal perception: timestamps are encoded only implicitly, frame level features are weak in capturing continuity, and language vision alignment often drifts from the entities of interest. In this paper, we present Grounded VideoDiT, a Video LLM designed to overcome these limitations by introducing three key innovations. First, a Diffusion Temporal Latent (DTL) encoder enhances boundary sensitivity and maintains temporal consistency. Second, object grounded representations explicitly bind query entities to localized visual evidence, strengthening alignment. Third, a mixed token scheme with discrete temporal tokens provides explicit timestamp modeling, enabling fine grained temporal reasoning. Together, these designs equip Grounded VideoDiT with robust grounding capabilities, as validated by state of the art results on Charades STA, NExT GQA, and multiple VideoQA benchmarks.

Figures

Figures reproduced from arXiv: 2508.15641 by the authors.

Figure 1
Figure 1. Teaser of our video–language model. Given a video and a natural-language instruction, we extract key nouns (e.g., dog, frisbee) and introduce special timestamp tokens (e.g., <24>, <96>--<120>) to enable precise temporal grounding. The nouns are linked to segmentation-guided object embeddings (Obj-Seg row), while video frames provide visual evidence along the timeline. The model handles three query types: (1) event l… view at source ↗
Figure 2
Figure 2. Model overview. A user instruction with a query and an accompanying video are provided as inputs. Key nouns extracted from the query are passed to a frozen Grounded-SAM2 module to obtain object-level segmentation masks and embeddings (left). In parallel, a frozen video diffusion encoder processes the video together with the masks and produces multi-scale features; we retain the intermediate representation at 10% of … view at source ↗
Figure 3
Figure 3. Qualitative comparisons on three tasks. We show results on Temporal Grounding, Grounded VideoQA, and Open-ended VideoQA. Given “Video with Segmentation” and timestamp tokens, our model (OURS) produces more precise localizations (e.g., 32.0s--58.0s) and more specific, object-aware answers (e.g., “a small red airplane”) than prior methods (LLaVA-ST, VTimeLLM, TimeChat, VideoLLaMA). Checkmarks indicate correct predicti… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 9 canonical work pages

  1. [1]

    Vtimellm: Empower llm to grasp video moments,

    B. Huang, X. Wang, H. Chen, Z. Song, and W. Zhu, “Vtimellm: Empower llm to grasp video moments,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 14 271–14 280. II-A, IV

    work page 2024

  2. [2]

    Videotree: Adaptive tree-based video representation for llm reasoning on long videos,

    Z. Wang, S. Yu, E. Stengel-Eskin, J. Yoon, F. Cheng, G. Bertasius, and M. Bansal, “Videotree: Adaptive tree-based video representation for llm reasoning on long videos,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 3272–3283. II-A

    work page 2025

  3. [3]

    Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding,

    Y . Guo, J. Liu, M. Li, D. Cheng, X. Tang, D. Sui, Q. Liu, X. Chen, and K. Zhao, “Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 3, 2025, pp. 3302–

    work page 2025

  4. [4]

    Flamingo: a visual language model for few-shot learning,

    J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y . Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds et al. , “Flamingo: a visual language model for few-shot learning,” Advances in neural information processing systems, vol. 35, pp. 23 716–23 736, 2022. II-A

    work page 2022

  5. [5]

    Video-llama: An instruction-tuned audio-visual language model for video understanding,

    H. Zhang, X. Li, and L. Bing, “Video-llama: An instruction-tuned audio-visual language model for video understanding,” arXiv preprint arXiv:2306.02858, 2023. II-A

    arXiv 2023

  6. [6]

    Video-chatgpt: Towards detailed video understanding via large vision and language models,

    M. Maaz, H. Rasheed, S. Khan, and F. S. Khan, “Video-chatgpt: Towards detailed video understanding via large vision and language models,” arXiv preprint arXiv:2306.05424 , 2023. II-A

    arXiv 2023

  7. [7]

    Timechat: A time-sensitive multimodal large language model for long video understanding,

    S. Ren, L. Yao, S. Li, X. Sun, and L. Hou, “Timechat: A time-sensitive multimodal large language model for long video understanding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 313–14 323. II-A, IV

    work page 2024

  8. [8]

    Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models,

    H. Wang, Z. Xu, Y . Cheng, S. Diao, Y . Zhou, Y . Cao, Q. Wang, W. Ge, and L. Huang, “Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models,” arXiv preprint arXiv:2410.03290, 2024. II-A

    arXiv 2024

Show all 18 references
  1. [9]

    Videogpt+: Integrating image and video encoders for enhanced video understanding,

    M. Maaz, H. Rasheed, S. Khan, and F. Khan, “Videogpt+: Integrating image and video encoders for enhanced video understanding,” arXiv preprint arXiv:2406.09418, 2024. II-A

    2024 arXiv

  2. [10]

    Longvlm: Efficient long video understanding via large language models,

    Y . Weng, M. Han, H. He, X. Chang, and B. Zhuang, “Longvlm: Efficient long video understanding via large language models,” in European Conference on Computer Vision . Springer, 2024, pp. 453–470. II-A

    2024

  3. [11]

    Video summarization using denoising diffusion probabilistic model,

    Z. Shang, Y . Zhu, H. Li, S. Yang, and X. Wu, “Video summarization using denoising diffusion probabilistic model,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 7, 2025, pp. 6776–6784. II-B

    2025

  4. [12]

    Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos,

    H. Yuan, X. Li, T. Zhang, Z. Huang, S. Xu, S. Ji, Y . Tong, L. Qi, J. Feng, and M.-H. Yang, “Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos,” arXiv preprint arXiv:2501.04001,

    arXiv

  5. [13]

    Videoglamm: A large multimodal model for pixel-level visual grounding in videos,

    S. Munasinghe, H. Gani, W. Zhu, J. Cao, E. Xing, F. S. Khan, and S. Khan, “Videoglamm: A large multimodal model for pixel-level visual grounding in videos,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 19 036–19 046. II-C

    2025

  6. [14]

    Vace: All-in- one video creation and editing,

    Z. Jiang, Z. Han, C. Mao, J. Zhang, Y . Pan, and Y . Liu, “Vace: All-in- one video creation and editing,” arXiv preprint arXiv:2503.07598, 2025. III-B0d

    2025 arXiv

  7. [15]

    Wan: Open and advanced large-scale video generative models,

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang et al. , “Wan: Open and advanced large-scale video generative models,” arXiv preprint arXiv:2503.20314 , 2025. III-B0d, IV-A

    2025 arXiv

  8. [16]

    Can i trust your answer? visually grounded video question answering,

    J. Xiao, A. Yao, Y . Li, and T.-S. Chua, “Can i trust your answer? visually grounded video question answering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 13 204–13 214. IV

    2024

  9. [17]

    Phi- 4 technical report,

    M. Abdin, J. Aneja, H. Behl, S. Bubeck, R. Eldan, S. Gunasekar, M. Harrison, R. J. Hewett, M. Javaheripi, P. Kauffmann et al. , “Phi- 4 technical report,” arXiv preprint arXiv:2412.08905 , 2024. IV-A

    2024 arXiv

  10. [18]

    Grounded sam: Assembling open-world models for diverse visual tasks,

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yanet al., “Grounded sam: Assembling open-world models for diverse visual tasks,” arXiv preprint arXiv:2401.14159 , 2024. IV-A

    2024 arXiv

Pith tools

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