Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Sidewalk Moments: Are Richer Representations Always More Human-Aligned? Evidence from City-Walk Videos

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read In binary judgments of engaging street moments, temporally averaged images match full video clips, and independent human raters agree.

desk verdict A genuinely useful empirical study of video vs. temporally averaged images for urban engagement, but the headline parity claim is confounded by unmatched backbones and an impure human TAI condition. read the letter →

arxiv 2607.20903 v1 pith:FVAXFNNK submitted 2026-07-23 cs.CV cs.HC

Liu Liu , Freya Huying Tan , Fábio Duarte This is my paper · ORCID
classification cs.CVcs.HC
keywords urbanperceptiontemporallyaveragedimagevideorepresentationengagementpredictioncity-walkvideosreplayheatbinaryclassificationensemble
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that more temporally rich video representations are not automatically more human-aligned than cheaper, temporally compressed image summaries when the task is distinguishing engaging from ordinary street moments. Using replay-heat engagement signals from 61 city-walk videos cut into over 50,000 ten-second clips, it shows that full spatiotemporal video features align best when ranking clips continuously, but that temporally averaged images (TAIs) match or beat video in the binary high-versus-low classification paradigm that most urban perception scoring models actually use. A separate human two-alternative forced-choice study supports the point: people identified the more engaging moment from TAIs 86.38% of the time, essentially tied with full video at 85.71%, while text descriptions trailed; audio features, by contrast, were near chance in the machine experiments. The paper concludes that representation choice is a substantive methodological filter, and that perceptually grounded temporal compression can be a principled alternative to full video encoding.

What carries the argument

The paper's central object is the temporally averaged image (TAI): a single composite created by pixel-wise averaging frames sampled at 1 Hz within each 10-second clip. The TAI is presented as a computational analog of the biological smoothing documented in ensemble-perception research, and it carries the argument because it removes transient motion while preserving stable spatial composition, the cue the paper links to composition-driven engagement judgments. Around this object, the argument is carried by three instruments: YouTube replay-heat values as a behavioral engagement signal; a within-video binary classification protocol with five quantile thresholds and six classifiers, which is the paradigm the paper says dominates perceptual scoring; and a gap analysis contrasting Group A clips (TAI correct, video wrong) with Group B clips (video correct, TAI wrong) using semantic segmentation variance and smoothed log-odds keyword contrasts.

What would settle it

Re-run the binary classification with the backbone held fixed: use VideoMAE-v2 on both the actual clip and a synthetic clip made by repeating the TAI ten times, and use MAE ViT-L on both the TAI and the clip's mid-point frame. If the synthetic static clip beats the real clip under VideoMAE, the paper's case is strengthened; if the real clip wins broadly, the TAI-video parity in the paper may be an artifact of comparing two different encoders rather than a fact about temporal richness.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that alignment with human engagement does not increase monotonically with temporal richness once the prediction task is binary. In continuous ranking, spatiotemporal video features lead the ordering (Spearman ρ = 0.626, skip-mid 0.785), followed by TAIs (0.545), text embeddings (0.517), mid-point frames (0.496), and audio (0.205). But across six classifiers and five quantile thresholds, TAIs match or outperform video in binary high-versus-low engagement classification in every linear and ensemble model, and only SVC with an RBF kernel gives video a consistent edge. The paper shows this is not an artifact of the replay-based label: in a separate two-alternative forced-choice study, human participants selected the higher-engagement moment from TAIs (86.38%) about as accurately as from full video (85.71%). The accompanying gap analysis dissociates the two visual representations by scene type—video succeeds in activity-driven scenes with high frame-to-frame variation, while TAIs succeed in composition-driven scenes with stable spatial structure—which the paper reads as evidence that human perceptual judgment is itself selectively temporal.

Load-bearing premise

The comparison depends on the two visual encoders being similar enough in power that the only meaningful difference between them is temporal information; if VideoMAE-v2 and MAE ViT-L differ in capacity or pretraining in ways that matter here, the TAI-video result could be a model artifact rather than a statement about temporal richness.

Editorial extensions

If this is right

  • Perceptual scoring pipelines that binarize engagement can substitute TAIs for full video with little or no loss in accuracy, at a fraction of the compute.
  • Continuous ordinal tasks—ranking many clips, retrieving top moments—still favor spatiotemporal video features, so the choice of representation should depend on whether the downstream task is binary or graded.
  • Representation choice biases measurement: TAIs foreground stable spatial composition and greenery while video amplifies pedestrian flow and activity, so urban perception models should match representations to the construct being measured.
  • Text descriptions carry a moderate semantic signal but cannot replace direct visual access, and audio contributes almost nothing to engagement prediction in city-walk footage.

Reading between the lines

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

  • Beyond the paper, the TAI-video parity suggests that large-scale urban perception surveys could switch from video encoding to aggregated street imagery with major cost and storage savings, provided the target construct is an extreme, binary judgment rather than a fine-grained rating.
  • The paper's ensemble-perception framing implies an untested prediction: TAI alignment with human judgment should peak near the temporal integration window of human summary perception—roughly a few seconds—and degrade for much shorter or much longer averaging windows.
  • The Group A/Group B imbalance (463 vs. 182 clips) hints that a hybrid model, routing stable scenes to TAI features and dynamic scenes to video features, could outperform either representation alone; the paper stops short of proposing such a model.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper asks whether richer visual representations (full video, temporally averaged images, audio, text) yield more human-aligned measures of urban engagement in first-person city-walk videos. Using 61 YouTube videos segmented into 50,380 ten-second clips, the authors compute replay-heat-based engagement labels and evaluate four feature types with Spearman correlation, binary classification, an AMT two-alternative forced-choice study, and a divergence/gap analysis. They report that video features dominate in continuous correlation, but that temporally averaged images (TAIs) match or outperform video features in binary high-vs-low engagement classification, and that human participants identify engaging moments about equally well from TAIs and full video. The paper concludes that richer representations are not inherently more human-aligned and that TAI-style temporal compression can be a principled alternative for urban perception pipelines.

Significance. If the central claim holds, the paper would be a useful challenge to a common implicit assumption in urban perception and perceptual-scoring research: that preserving more temporal detail in visual representations necessarily improves alignment with human judgment. The study's scale (over 50k clips), its use of multiple classifiers and quantile thresholds, the independent AMT validation, and the gap-analysis method are notable strengths; the paper also ships code and a reasonably detailed methods section. However, the empirical support for the headline parity claim is currently incomplete because both pillars of the comparison are confounded: the model-based video-vs-TAI comparison uses two different encoders, and the human 'TAI' condition includes a representative frame alongside the averaged image. The manuscript is therefore a promising but not yet conclusive demonstration of the claimed effect.

major comments (3)
  1. [§5.5, Table 1, Fig. 2, Table S3] The central model-based comparison between video and TAI features is confounded by encoder choice. Video features are extracted with VideoMAE-v2, while TAI features are extracted with MAE ViT-L; these encoders differ in pretraining data and objective, architecture, depth, and feature dimensionality. The statement in §5.5 that both are 'part of the same masked autoencoder family' does not control for capacity or pretraining differences. The internal TAI-vs-midpoint comparison uses the same MAE ViT-L backbone and does show a benefit of temporal averaging for image features, but it cannot validate the cross-encoder video-vs-TAI comparison. To support the claim that temporal compression, rather than model choice, drives the observed parity, the authors need to match the backbones (e.g., use the same ViT architecture for video frames with a temporal pooling operation, or apply VideoMAE-v2 to a single averaged frame) or otherwise demonstrate that the TAI-vs-video gap is robust to encoder choice.
  2. [§5.7, §2.3] The human validation does not actually test the pure TAI representation. The image-preference task showed 'a temporally averaged image alongside a representative frame' (Methods 5.7), so the reported 86.38% accuracy for the image condition could plausibly be driven by the representative frame rather than by the temporal average. Without separate conditions for pure TAI, representative-frame-only, and video, the claim that human judgments 'confirm' TAI-video parity is not supported. In addition, the 86.38% vs 85.71% comparison is reported without confidence intervals or an equivalence test, so the 'parity' is an informal claim rather than a statistically established one. The authors should either add the missing conditions or substantially weaken the claim about independent human validation of TAIs.
  3. [§5.2, §5.6] The clip-segmentation scheme uses a 10-second window with a 5-second stride, so adjacent clips share 5 seconds of content. Methods 5.6 states that 'identical train and test splits' were used across representations but does not describe how those splits were constructed. If the split is random across clips, overlapping time intervals will appear in both training and testing, inflating AUC values for all representations and potentially altering the TAI-vs-video ordering. The authors should specify the exact split procedure (e.g., a per-video temporal split with no overlap, or a clip-level split with deduplication of overlapping regions) and report whether the headline results persist under non-overlapping 10-second windows or with a stride of 10 seconds.
minor comments (5)
  1. [§2.1, Table 1] Table 1 reports means across 61 videos but does not report per-video variability (e.g., standard deviation or percentile intervals) for the correlation metrics; this would help assess whether the ordering is consistent across videos rather than driven by a few outliers.
  2. [Fig. 2] The figure legend labels the TAI feature as 'MAE-ViT-Large' while the Methods and text use 'MAE ViT-L' and 'MAE ViT-Large' in different places; this should be standardized for clarity.
  3. [§5.4] The replay-heat smoothing spline uses s = 2, a free parameter; the paper does not report sensitivity to this choice. Since the heat signal is the ground-truth proxy for all downstream comparisons, a brief robustness check (or an explicit statement in the limitations) would strengthen the analysis.
  4. [§2.3] The claim that the difficult-pair subset has N ≈ 46 per condition is under-specified; the authors should state the total number of pairs per condition and how the 0.3–0.5 heat-gap subset was sampled, so the human-study sample sizes are transparent.
  5. [§5.5] The feature-extraction details for VideoMAE-v2 and MAE ViT-L (input resolution, feature dimension, which layer is used) are not reported; providing these would improve reproducibility and would also make the backbone-capacity gap between the two encoders explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the engagement proxy is externally validated by human judgments and no fitted quantity is relabeled as a prediction.

full rationale

The paper's central comparisons are empirical rather than definitional. Replay heat values are an externally observed behavioral signal extracted from YouTube, not a function of the TAI or video features; binary classification labels are quantiles of that signal, and classifiers are evaluated on within-video splits with no representation-specific tuning. The AMT study is an independent data source: participants chose between stimuli based on their own engagement judgments, and replay heat values served only as reference labels for scoring accuracy, so the observed 86.38% versus 85.71% parity is an empirically contingent result, not a construction. The TAI representation (pixel-wise averaging of sampled frames) does not encode the replay-heat target, and the TAI-versus-midpoint comparison uses the same MAE ViT-L backbone, isolating temporal averaging. The main validity threats—different encoders for video versus TAI (VideoMAE-v2 versus MAE ViT-L) and the TAI human condition showing a representative frame alongside the average—are experimental confounds that bear on whether the manipulation isolates temporal averaging, but they are not cases where a prediction reduces to its input by definition or where a fitted parameter is renamed as a result. No load-bearing self-citation chain is present; the cited cognitive-science results motivate the TAI probe but do not supply the measured outcomes.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The empirical comparison rests on external behavioral data (replay heat) and an external human validation, but several hand-chosen thresholds and a strong backbone-comparability assumption are load-bearing.

free parameters (5)
  • YouTube heat spline smoothing s = 2
    Hand-chosen smoothing factor for UnivariateSpline used to interpolate replay heat anchor points to per-second values; affects the engagement target (Methods 5.4).
  • Editing-transition z-score threshold = 2
    Frames with normalized cosine similarity above 2 are flagged as cuts and their clips removed (Methods 5.3).
  • Skip-mid exclusion band = 30% of median heat
    Clips near the median heat value are excluded in the skip-mid correlation analysis; a robustness check, not the central result (Table 1).
  • Quantile thresholds = 10% to 50%
    Within-video heat quantiles define high- and low-engagement labels; the 10% and 20% thresholds dominate the reported findings.
  • VLM keyword count = 5 keywords per frame
    The Intern-VL3 prompt requests exactly five nouns per frame, shaping the text modality's information content (Methods 5.5).
assumptions (5)
  • domain assumption Replay heat values from YouTube are a valid behavioral proxy for moment-level human engagement.
    Used as prediction target throughout; AMT validation supports it at ~86% accuracy, but it remains platform-specific (Methods 5.4, Limitations).
  • domain assumption VideoMAE-v2 and MAE ViT-L encoders are sufficiently matched to attribute performance differences to temporal richness rather than architecture or capacity.
    The two backbones differ in architecture, pretraining data, and feature dimension, yet the paper treats them as comparable because both are masked autoencoders (Methods 5.5).
  • domain assumption A 10-second window with 5-second stride approximates the time over which humans form stable scene impressions.
    Motivates the clip segmentation and the TAI design; cited to ensemble perception (Methods 5.2, 5.5).
  • domain assumption City-walk YouTube videos represent typical first-person pedestrian experience.
    Dataset is limited to a genre with consistent recording conventions; acknowledged in Limitations.
  • domain assumption Classifiers trained without representation-specific hyperparameter tuning provide a fair comparison.
    No per-representation tuning was performed; differences in feature scale or dimensionality are not otherwise controlled (Methods 5.6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sidewalk Moments: Are Richer Representations Always More Human-Aligned? Evidence from City-Walk Videos." pith.science (2026). https://pith.science/paper/FVAXFNNK

@misc{pith2026260720903,
  author       = {Pith},
  title        = {Pith review of: Sidewalk Moments: Are Richer Representations Always More Human-Aligned? Evidence from City-Walk Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FVAXFNNK}},
  note         = {Machine review of arXiv:2607.20903}
}
read the original abstract

We examine whether richer visual representations yield more human-aligned measures of urban engagement, using 61 first-person city-walk videos from YouTube segmented into over 50,000 ten-second clips and represented across four modalities: spatiotemporal video features, temporally averaged images (TAIs), audio embeddings, and text-based semantic descriptions. Spearman correlation analysis reveals the expected ordering along the temporal-richness continuum, with video features showing the strongest continuous alignment. However, this ordering breaks down under binary classification of high- versus low-engagement moments (the paradigm most commonly used to train perceptual scoring models), where TAIs consistently match or outperform video across most classifiers and quantile thresholds. An independent two-alternative forced-choice study on Amazon Mechanical Turk confirms that this parity reflects human judgment: participants identified engaging moments with comparable accuracy from TAIs and full video clips, while text performed substantially worse and audio remained near chance. Gap analysis reveals a functional dissociation: video features are advantaged in activity-driven scenes with dynamic content, whereas TAIs better align with human judgments in composition-driven scenes dominated by stable spatial structure. These findings challenge the assumption that richer representations are inherently more human-aligned, and suggest that perceptually grounded temporal compression can be a principled alternative to full video encoding.

Figures

Figures reproduced from arXiv: 2607.20903 by the authors.

Figure 1
Figure 1. Overview of the study framework. City-walk videos are segmented into short clips and linked to replay-based heat values as a behavioral signal of engagement. Multiple representations are extracted and evaluated through within-video prediction and human preference validation, with a gap analysis examining conditions under which representations diverge. 3 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Model-wise AUC (top) and F1 (bottom) across quantile splits for different feature representations. Solid lines represent mean values, with shaded areas indicating 95% confidence intervals (CIs). Despite preserving fine-grained temporal information, video-based representations do not consistently outperform temporally compressed images across classifiers and thresholds. This pattern suggests that increased representa… view at source ↗
Figure 3
Figure 3. Overview of the human validation pipeline. From left to right: multimodal repre￾sentations (video, TAI, and keywords); construction of within-video high-heat vs. low-heat pairs; two-alternative forced-choice (2AFC) evaluation on AMT; and aggregated accuracies across representational conditions. Across representational conditions, human judgments showed systematic correspondence with the replay-based engagement signa… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparative analysis of top-20 distinctive keywords for GA (where TAI excels) and GB (where video excels). Keywords are ranked by smoothed log-odds ratio [30], with bar length indicating the degree of association with each group. Left (blue) bars represent features dom…
Figure 5
Figure 5. Figure 5: Illustrative examples from the four subgroups in the divergence analysis. Group A (GA) contains clips for which the TAI-based prediction is correct and the video-based predic￾tion is incorrect, while Group B (GB) contains clips for which the video-based prediction is c…
Figure 6
Figure 6. Figure 6: Geographic distribution of the CityWalk Corpus. Red markers indicate cities represented in the dataset, spanning North America, Europe, Asia, and the Middle East. The diversity of urban contexts supports the generalizability of findings across different street environm…
Figure 7
Figure 7. Figure 7: Examples of pedestrian video footage with replay heat values. Each row shows a representative clip from the CityWalk Corpus; the heat value curve (shown in the progress bar below each frame) reflects the relative frequency with which viewers replayed that segment, with…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    The image of the city (MIT press, 1960)

    Lynch, K. The image of the city (MIT press, 1960). 19

    work page 1960

  2. [2]

    & Malcolm, R

    Lynch, K. & Malcolm, R. A walk around the block (Massachusetts Institute of Technol- ogy, Department of City Planning, 1959)

    work page 1959

  3. [3]

    Why buildings are known: A predictive tool for architects and planners

    Appleyard, D. Why buildings are known: A predictive tool for architects and planners. Environment and Behavior 1, 131–156 (1969)

    work page 1969

  4. [4]

    Styles and methods of structuring a city

    Appleyard, D. Styles and methods of structuring a city. Environment and behavior 2, 100–117 (1970)

    work page 1970

  5. [5]

    & Kaplan, R

    Kaplan, S. & Kaplan, R. Cognition and Environment: Functioning in an Uncertain World (Praeger, 1982)

    work page 1982

  6. [6]

    The psychology of place

    Canter, D. The psychology of place. (St Martin’S Press, 1977)

    work page 1977

  7. [7]

    Nasar, J. L. The evaluative image of the city. Journal of the American Planning Association 56, 41–53 (1990)

    work page 1990

  8. [8]

    Human aspects of urban form: towards a man—environment approach to urban form and design (Elsevier, 2013)

    Rapoport, A. Human aspects of urban form: towards a man—environment approach to urban form and design (Elsevier, 2013)

    work page 2013

Show all 34 references
  1. [9]

    Salesses, M. P. Place Pulse: Measuring the collaborative image of the city . Ph.D. thesis, Massachusetts Institute of Technology (2012)

    2012

  2. [10]

    & Hidalgo, C

    Naik, N., Philipoom, J., Raskar, R. & Hidalgo, C. Streetscore-predicting the perceived safety of one million streetscapes. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops , 779–785 (2014)

    2014

  3. [11]

    & Hidalgo, C

    Dubey, A., Naik, N., Parikh, D., Raskar, R. & Hidalgo, C. A. Deep learning the city: Quantifying urban perception at a global scale. In European conference on computer vision, 196–212 (Springer, 2016)

    2016

  4. [12]

    & Sabel, C

    Zhang, Z., Mˇ echurov´ a, K., Resch, B., Amegbor, P. & Sabel, C. E. Assessing the associ- ation between overcrowding and human physiological stress response in different urban contexts: a case study in salzburg, austria. International Journal of Health Geographics 22, 15 (2023)

    2023

  5. [13]

    Mavros, P., J W¨ alti, M., Nazemi, M., Ong, C. H. & H¨ olscher, C. A mobile eeg study on the psychophysiological effects of walking and crowding in indoor and outdoor urban environments. Scientific reports 12, 18476 (2022)

    2022

  6. [14]

    Yang, N. et al. Urban perception assessment from street view images based on a mul- tifeature integration encompassing human visual attention. Annals of the American Association of Geographers 114, 1424–1442 (2024)

    2024

  7. [15]

    & Tong, T

    Li, C., Du, C., Ge, S. & Tong, T. An eye-tracking study on visual perception of vegetation permeability in virtual reality forest exposure. Frontiers in Public Health 11, 1089423 (2023)

    2023

  8. [16]

    Zhang, F. et al. Measuring human perceptions of a large-scale urban region using machine learning. Landscape and Urban Planning 180, 148–160 (2018). 20

    2018

  9. [17]

    Quintana, M. et al. Global urban visual perception varies across demographics and personalities. Nature Cities 1–15 (2025)

    2025

  10. [18]

    De Nadai, M. et al. The death and life of great italian cities: A mobile phone data perspective. Proceedings of the National Academy of Sciences 113, 11547–11552 (2016)

    2016

  11. [19]

    & Biljecki, F

    Ito, K., Kang, Y., Zhang, Y., Zhang, F. & Biljecki, F. Understanding urban perception with visual data: A systematic review. Cities 152, 105169 (2024)

    2024

  12. [20]

    Arnab, A. et al. Vivit: A video vision transformer. In IEEE/CVF (ed.) Proceedings of the IEEE/CVF international conference on computer vision , 6836–6846 (2021)

    2021

  13. [21]

    & Torresani, L

    Bertasius, G., Wang, H. & Torresani, L. Is space-time attention all you need for video understanding? In Meila, M. & Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, vol. 139 of Proceedings of Machine Learning Research, 849–859 (PMLR, 2021...

    2021

  14. [22]

    & Wang, L

    Tong, Z., Song, Y., Wang, J. & Wang, L. Videomae: Masked autoencoders are data- efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078–10093 (2022)

    2022

  15. [23]

    Wang, L. et al. Videomae v2: Scaling video masked autoencoders with dual masking. In IEEE/CVF (ed.) Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14549–14560 (2023)

    2023

  16. [24]

    Girdhar, R. et al. Imagebind: One embedding space to bind them all. In IEEE/CVF (ed.) Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, 15180–15190 (2023)

    2023

  17. [25]

    Li, B. et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024)

    2024 arXiv

  18. [26]

    & Torralba, A

    Oliva, A. & Torralba, A. Building the gist of a scene: the role of global image features in recognition. Progress in brain research 155, 23–36 (2006)

    2006

  19. [27]

    & Yamanashi Leib, A

    Whitney, D. & Yamanashi Leib, A. Ensemble perception. Annual Review of Psychology 69, 105–129 (2018)

    2018

  20. [28]

    & Whitney, D

    Manassi, M. & Whitney, D. Illusion of visual stability through active perceptual serial dependence. Science Advances 8, eabk2480 (2022)

    2022

  21. [29]

    Langlois, J. H. & Roggman, L. A. Attractive faces are only average. Psychological Science 1, 115–121 (1990)

    1990

  22. [30]

    L., Colaresi, M

    Monroe, B. L., Colaresi, M. P. & Quinn, K. M. Fightin’words: Lexical feature selection and evaluation for identifying the content of political conflict. Political Analysis 16, 372–403 (2008). 21

    2008

  23. [31]

    & Lee, Y

    Liu, H., Li, C., Wu, Q. & Lee, Y. J. Visual instruction tuning. Advances in neural information processing systems 36, 34892–34916 (2023)

    2023

  24. [32]

    He, K. et al. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16000–16009 (2022)

    2022

  25. [33]

    Wang, W. et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    2025 arXiv

  26. [34]

    & Cao, Y

    Sun, Q., Fang, Y., Wu, L., Wang, X. & Cao, Y. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023). 22 Supplementary Information Sidewalk Moments: Are Richer Representations Always More Human-Aligned? Evidence from City-Walk Videos S...

    2023 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.