Pith. sign in

REVIEW 3 major objections 4 minor 134 references

FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing a monolithic JEPA predictor with factorized layout, agent, and interaction channels improves future-latent accuracy, intervention sensitivity, and occlusion resilience in dense urban video.

desk verdict Real contributions in the factorized predictor and DENSEWORLD dataset, but the audit-split contradiction must be resolved before Causal L1 is credible. read the letter →

arxiv 2608.01049 v1 pith:ZJA5X4PD submitted 2026-08-02 cs.AI cs.CVcs.LG

Kapil Wanaskar , Gaytri Jena , Aman Chadha , Vinija Jain , Vasu Sharma , Amitava Das This is my paper
classification cs.AIcs.CVcs.LG
keywords worldmodelsJEPAfactorizedpredictiondenseurbanscenespartialobservabilityDENSEWORLDvideoforecastinginterventionsensitivity
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 argues that JEPA (Joint Embedding Predictive Architecture) world models fail in crowded, chaotic cities because their monolithic predictor packs everything into one entangled latent, letting the model exploit visual shortcuts instead of recovering scene structure. To fix this, FactorJEPA decomposes the predicted future into layout, visibility-gated agents, and sparse interactions, each with its own subspace and supervision. On DENSEWORLD, a new 1,000-hour, 22-city dataset of dense Global South urban video, this factorized predictor lowers future-frame L1 and causal intervention L1, flattens the mask-ratio slope, and leaves a reproducible motion trade-off that full-scale training resolves. If the claim holds, explicit factorization, not just more fine-tuning, is the way to make video world models forecast dense, partially observed scenes.

What carries the argument

The load-bearing object is the factorized predictor with block-structured matrix factorization $\hat{Y}= C A^\top = C_L A_L^\top + C_A A_A^\top + C_I A_I^\top$. Each factor has its own coordinate block, synthesis dictionary, and reliability-weighted supervision; a soft visibility gate $v_n^{(i)}$ attenuates uncertain or occluded agents without removing them, and a separation loss penalizes linear and nonlinear cross-channel covariance. This forces the factorization to be architectural rather than a post hoc label: distinct pathways and dictionaries compose the future embedding in a way a monolithic predictor does not.

What would settle it

Train FactorJEPA with the same architecture but feed it temporally shuffled or clip-scrambled layout, agent, and interaction targets. If the predictive advantages over FactorJEPA-RAW survive or grow, the claimed semantic anchoring is not the active mechanism and the gains come from regularization rather than structure.

Watch

Extended reading notes

Core claim

FactorJEPA's central claim is that how a predictor organizes future information matters as much as how well it matches the target. In a conventional JEPA, the future embedding is produced by one monolithic predictor, and in dense scenes that predictor can satisfy the objective by mixing crowd texture, visible appearance, and road geometry into shortcuts. FactorJEPA composes the future latent as $\hat{Y}=C_L A_L^\top + C_A A_A^\top + C_I A_I^\top$, with layout, visibility-gated agent, and sparse interaction coordinates, each anchored by DINOv2-derived targets with reliability weights, and penalizes cross-channel leakage. The paper reports separation from the best matched baseline on Future-fr

Load-bearing premise

The whole gain depends on the DINOv2-derived teacher targets being reliable enough to anchor the three channels; if they miss small or occluded agents or misclassify interactions, the factor supervision can bias the predictor rather than improve it.

Editorial extensions

If this is right

  • If the factorization is what drives the gains, JEPA world models for dense scenes should be built with explicit layout, agent, and interaction channels rather than monolithic predictors.
  • The mask-ratio result implies that factorized structure can soften degradation under partial observability, so occluded agents continue to be predicted through visibility gating and interaction priors.
  • Cross-scale stability means a 1B backbone can stand in for a 2B model when screening adaptation choices, reducing experiment cost.
  • The motion-cosine trade-off, if data-dependent, implies that factorized training prioritizes motion that supports forecasting, and that the trade-off is not an intrinsic cost of factorization.
  • DENSEWORLD offers a benchmark stress test for any world model claiming to handle dense urban dynamics and partial observability.

Reading between the lines

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

  • The factorization principle should transfer beyond Global South streets to other crowded partial-observability settings—indoor crowds, disaster scenes, mixed mobility—provided a teacher can supply layout, agent, and interaction targets.
  • The separation loss may serve as a practical interpretability diagnostic: if off-diagonal leakage is low, individual channels can be inspected or edited with predictable effects on the decoded future.
  • Causal L1 only measures consistency with imposed interventions, not causal identification; the strongest reading of the paper is structure-preserving prediction rather than learned causal mechanisms.
  • A teacher-free variant that discovers layout, agent, and interaction coordinates without DINOv2 targets would test whether the architecture alone, rather than the specific teacher, is the source of the gains.
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

3 major / 4 minor

Summary. The paper introduces DENSEWORLD, a roughly 1,000-hour, 22-city video benchmark for crowded, chaotic Global South urban scenes, and FactorJEPA, a variant of V-JEPA in which the monolithic future predictor is replaced by factorized layout, agent, and interaction channels with a soft visibility gate and separation losses. The central empirical claims are that FactorJEPA improves future-latent accuracy (Future-frame L1), intervention-sensitive prediction (Causal L1), and robustness to reduced evidence (Mask-ratio slope) over conventional fine-tuning and parameter-efficient adaptation baselines, with stable method rankings across 2B and 1B V-JEPA 2.1 backbones. The paper includes a detailed appendix with target-construction contracts, training pseudocode, resource accounting, paired clustered inference, ablations, and a latent-to-RGB decoding analysis, and it releases the dataset and checkpoints.

Significance. If the empirical claims hold, the paper makes a useful contribution: it identifies a predictive regime that is underrepresented in JEPA evaluations, provides a large public dataset for that regime, and shows that explicitly structured predictive channels can outperform monolithic adaptation for dense, partially observed urban scenes. The paper's strengths include a carefully matched experimental protocol across methods, paired hierarchical bootstrap inference, a clear attribution chain (Auto-RGN -> FactorJEPA-RAW -> FactorJEPA), unusually explicit limitation statements, and public release of data and checkpoints. The main weaknesses are not in the experimental ambition but in unresolved inconsistencies in the definition and provenance of the primary measurement signals, which currently prevent the central claims from being accepted as stated.

major comments (3)
  1. [§4 'Evaluator Independence and Audit Split'; Appendix A.3; Appendix C.1; Limitations (i)] The main text states that the audit split D_audit has agent masks, visibility states, interaction pairs, and intervention regions 'independently annotated' via T_audit = Ψ_human(x), and Limitations (i) relies on this audit split to ensure the interaction generator is not reused as the headline evaluator. However, Appendix A.3 states that 'DENSEWORLD does not use manually annotated factor labels' and Appendix C.1 repeats 'no manually annotated factor labels.' These statements are mutually contradictory. If Ψ_human does not exist, Causal L1 has no human-annotated intervention regions and the stated safeguard against agreement with the DINOv2-based pseudo-label generator collapses. The paper must state clearly whether a human-annotated audit split exists; if it does not, it must specify the actual provenance of the intervention regions (e.g., the automatic G_int in Appendix E.2) and demonst
  2. [§4 'Future-frame L1' vs Appendix E.2 and Table 8] The main text and figures (including Figure 8 and Figure 9) report the headline metric as 'Future-frame L1' and define it as a normalized L1 distance. Appendix E.2, however, defines 'Future-frame MSE' as the executed quantity and explicitly says: 'The executed scorecard reports this MSE quantity; it is not relabeled as L1.' Table 8 also lists 'Future-frame MSE' as the primary metric. Since this metric carries one of the paper's three central claims and the cross-scale correlations, the discrepancy is load-bearing. The authors must specify which quantity is actually computed and reported, and correct the main text, appendix, or tables accordingly. If MSE is used, all reported values and comparisons need to be re-expressed or relabeled consistently.
  3. [§4 'Future-frame L1' and 'Mask-ratio slope'; Protocol, Algorithm .1 step 41; Table 6] The main text says Future-frame L1 and Mask-ratio slope use 'the frozen V-JEPA target encoder,' which suggests a single, method-independent evaluation target. But the training protocol (Algorithm .1 step 41 and Table 6) updates the momentum target encoder by EMA from the online encoder, so after training the target encoder is generally not the original pretrained V-JEPA checkpoint unless a separate frozen copy is used for evaluation. If evaluation uses the EMA-updated target encoder, then each method is evaluated against a different target space and the claimed independence of the headline diagnostics is not established. The paper must specify exactly which target encoder is used at evaluation time: a frozen pretrained V-JEPA encoder, or the per-method momentum encoder. This matters for Future-frame L1, Mask-ratio slope, and Causal L1 alike.
minor comments (4)
  1. [§2 'How Dense is DENSEWORLD?'] The sentence 'As shown in Section , we measure this regime...' has an empty section reference. Please fill in the correct section number.
  2. [Appendix E.5] The text says 'Figures 17 and 17 show all executed diagnostics for the ViT-G and ViT-g backbones.' The figure numbering appears duplicated/broken; please renumber and reference the scorecards uniquely.
  3. [Table 6] Several resource entries are labeled 'prespecified execution ceilings' and 'must not be described as measured consumption unless verified against profiling logs.' As currently presented, the table mixes commitments with measurements. Please separate budget ceilings from profiled results, or state explicitly in the main text that resource figures are planned envelopes, not measured outcomes.
  4. [§4 and Table 8] The metric is called 'Causal L1' in the abstract and main text but 'Intervention L1' in Table 8, where the paper notes the revised name reflects that the metric does not establish causal identification. Align the terminology throughout to avoid confusion.

Circularity Check

2 steps flagged · score 6.0 of 10

Causal L1's independent-audit guarantee is contradicted by the paper's no-manual-annotation appendix, and Future-frame L1/Mask-ratio slope use the method-dependent EMA target encoder despite being called frozen.

  1. self definitional [Main text 'Evaluator Independence and Audit Split'; Appendix A.3 'Partitioning and DINOv2 target provenance'; Appendix C 'Frozen Teacher and Factor-Target Construction']
    "Primary factor-specific evaluation uses a city-disjoint audit split D_audit, whose agent masks, visibility states, interaction pairs, and intervention regions are independently annotated: Dtrain ∩ Daudit = ∅, Taudit = Ψ_human(x). ... DENSEWORLD does not use manually annotated factor labels. For each partition, automated structural targets are generated by the fixed DINOv2 pipeline: T_DINO_s = {Ψ_struct(F_DINOv2(x)) : x ∈ D_s}, s ∈ {train,val,test}."

    The audit split is defined as the guarantee that Causal L1's intervention regions are independent of the DINOv2 teacher ('Intervention regions and types come from independent audit annotations, not DINOv2 predictions'). But the same paper states that DENSEWORLD has no manually annotated factor labels and that every partition's factor targets are generated by the fixed DINOv2 pipeline. If T_audit = Ψ_human(x) exists, the no-manual-labels claim is false; if it does not exist, the audit split cannot supply independently annotated intervention regions, so by the paper's own metric definition they must reduce to DINOv2-generated regions. In that case, Causal L1 evaluates FactorJEPA's consistency with changes defined by the same teacher that produced the interaction supervision, and the stated s

  2. other [Main text 'Evaluator Independence and Audit Split' vs Appendix C.3 'Optimization, Model Selection, and Executed Algorithm' and Algorithm .1 step 41]
    "Future-frame L1 and Mask-ratio slope use the frozen V-JEPA target encoder. ... The target encoder is updated only through the executed exponential-moving-average rule. ... Update the momentum encoder: ¯θ = μ_s ¯θ + (1−μ_s)θ."

    The evaluator described as 'frozen' is actually the momentum target encoder, which is updated each step from the online encoder being trained. Hence the target Y⋆ = sg(f̄θ(x)) is method-dependent and changes as FactorJEPA (or any compared method) trains. Future-frame L1 is exactly the LJEPA term that FactorJEPA's predictor is trained to minimize, and Mask-ratio slope is derived from that same self-referential error. The claim that 'no headline evaluator reuses FactorJEPA's pseudo-label generator' remains true, but the target is not an independent fixed ground truth; part of the reported improvement on these two primary diagnostics can reflect each model shaping its own target encoder. The prose 'frozen V-JEPA target encoder' is contradicted by the training protocol.

full rationale

The central architectural derivation is not circular: FactorJEPA's predictor factorization, the factor-target supervision, and the FactorJEPA-RAW versus FactorJEPA ablation are genuine design choices, and the paper does not claim to predict the DINOv2 pseudo-labels as its headline result. The circularity problem is confined to the evaluation independence guarantees. The paper's own statements contradict each other on whether a human-annotated audit split exists. Since Causal L1 is one of the three headline improvements and its entire claim to independence rests on D_audit, the contradiction is load-bearing: if the audit does not exist, Causal L1 reduces to measuring agreement with the DINOv2 teacher's intervention regions, exactly the circularity the paper says it avoids. Separately, the 'frozen' target encoder used for Future-frame L1 and Mask-ratio slope is actually the EMA momentum encoder, making these metrics self-referential rather than measured against a fixed external signal. These issues make the headline rankings partially dependent on the model's own outputs, though the method could still be genuinely better. Score 6 indicates partial circularity in the evaluation, not a derivation that is equivalent to its inputs by definition.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests mainly on DINOv2-derived pseudo-labels, hand-chosen factor ranks and loss weights, and an ambiguity about the evaluation target encoder. No new physical or conceptual entity beyond the method's architectural components is introduced.

free parameters (6)
  • Factor ranks r_L, r_A, r_I = 64, 96, 64
    Hand-chosen subspace dimensions; no derivation. The larger agent rank is an ad hoc choice.
  • Loss coefficients lambda_sep, lambda_sparse, lambda_V, lambda_sup, beta_nlin = 0.05, 0.01, 0.25, 1.00, 0.10
    Fixed on validation, not derived; directly control the objective's balance.
  • DINOv2 region retention thresholds = conf 0.55, min area 0.1%, IoU 0.70
    Determine which regions become teacher targets and hence the supervision signal.
  • Track association parameters = 0.45 box IoU, 0.35 mask IoU, 0.20 descriptor cosine, max gap 4f, min length 3f
    Define tracklets and interaction pairs; interaction targets are sensitive to these.
  • Interaction graph radius and max neighbors = radius 0.25, max 12
    Hand-set locality rule for the interaction channel.
  • Top-K encoder blocks = 2 for 2B, 1 for 1B
    Chosen to match trainable-parameter budget of baselines; affects capacity.
assumptions (5)
  • domain assumption DINOv2 features support reliable region proposals and tracklets in dense urban scenes
    Used in Appendix C.1 to construct all factor targets; if false, factor supervision is biased.
  • domain assumption V-JEPA 2.1 pretrained encoder is a meaningful backbone and its latent space is a valid prediction target
    The whole study conditions on V-JEPA 2.1; no independent justification for this choice.
  • domain assumption A frozen or momentum target encoder provides a stable evaluation oracle across methods
    Future-frame L1 and Mask-ratio slope rely on this; the paper is ambiguous about which encoder is used.
  • domain assumption The automatic intervention generator produces meaningful edits for causal sensitivity
    Appendix E.2 defines the generator as automatic and not consuming manual labels; the main text contradicts this by saying interventions are human-annotated.
  • standard math Hierarchical bootstrap over cities and source videos is valid for the reported confidence intervals
    Appendix E.4 uses paired clustered bootstrap; assumes source-video-level dependence is sufficient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds." pith.science (2026). https://pith.science/paper/ZJA5X4PD

@misc{pith2026260801049,
  author       = {Pith},
  title        = {Pith review of: FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZJA5X4PD}},
  note         = {Machine review of arXiv:2608.01049}
}
read the original abstract

World models have attracted significant attention for their ability to capture and predict the structure and dynamics of the physical world. In this emerging landscape, Joint Embedding Predictive Architectures (JEPA) offer a particularly compelling direction. We study a largely unexplored regime: populous, crowded, and chaotic Global South urban environments, which we call DENSEWORLD. Unlike the lower-density, lane-structured settings that dominate existing evaluations, these scenes exhibit soft spatial boundaries, extreme agent heterogeneity, persistent occlusion, and rapid social negotiation under mixed traffic. We introduce the first large-scale dataset for this regime: 1,000 hours of drive-through, walk-through, and aerial video across 22 cities. Existing JEPA formulations struggle to preserve dense interaction dynamics under heterogeneity and partial observability. We introduce FactorJEPA, which makes world structure a first-class predictive primitive. Rather than encoding the future in a monolithic latent, it composes layout, entities, and interactions, using a visibility gate and separated subspaces to preserve partially observed agents and discourage cross-factor shortcuts. FactorJEPA improves (i) future-latent accuracy (Future-frame L1), (ii) intervention-sensitive prediction (Causal L1), and (iii) robustness to reduced visual evidence (Mask-ratio slope), while exposing (iv) a reproducible motion-information trade-off (Motion cosine). Method rankings replicate across 2B and 1B V-JEPA 2.1 backbones, with rho = 0.895 to 0.978. We publicly release the DENSEWORLD-115k dataset (https://huggingface.co/datasets/anonymousML123/denseworld-115k) and the surgery-trained FactorJEPA checkpoints (https://huggingface.co/datasets/anonymousML123/factorjepa-outputs/tree/main/outputs/full/vjepa_2_1_vitg_1B/train/m09c_surgery_3stage_DI_diheavy_encoder).

Figures

Figures reproduced from arXiv: 2608.01049 by the authors.

Figure 1
Figure 1. Same question, two encoders. Each row is a three￾frame filmstrip of one held-out clip posed as a multiple￾choice question, answered from an identical probe head over the frozen V-JEPA 2.1 encoder vs. ours (factor-view predictor surgery); only the backbone differs. Top: motion speed (ours 69.8% vs frozen 60.9%). Bottom: turn direction. and chaotic Global South urban scenes. It is defined not only by geography, but by… view at source ↗
Figure 2
Figure 2. DENSEWORLD 1.0 scene-type coverage. Representative examples from the dataset illustrate the breadth of urban environments covered by DENSEWORLD 1.0, including market, residential, commercial, promenade, transit, highway, heritage, junction, flyover, and beach scenes. This diversity reflects the spatial, social, and infrastructural heterogeneity of populous, crowded, and chaotic urban environments. geneity, with pede… view at source ↗
Figure 3
Figure 3. DENSEWORLD exhibits higher multi-agent density than standard driving benchmarks. We compare agent count density and agent occupancy across matched scene types. DENSEWORLD shows large gaps in mar￾ket, commercial, and flyover/underpass scenes, indicating stronger interaction pressure and heavier visual competition. ing predictive world models. Can Fine-Tuning Close the DENSEWORLD Gap? Before introducing FactorJEPA, we… view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: DENSEWORLD 1.0 scene-type coverage. Twelve representative scene types (market, temple, commercial, transit, residential, promenade, ghat, heritage, highway, junction, flyover, beach) across 22 Indian cities, shown from ground-level pedestrian viewpoints. A deterministi…
Figure 5
Figure 5. Figure 5: Factorized decomposition of the same urban scene. (a) Agents are isolated from scene context. (b) Layout highlights persistent spatial structure while retaining suppressed agent silhouettes. (c) Interactions show tracklets, residual motion, and sparse pairwise coupling…
Figure 6
Figure 6. Figure 6: Factorized prediction. FactorJEPA composes the target embedding from layout, agents, and interactions, with a soft visibility gate applied only to entity terms. Block-Structured Matrix Factorization For each factor k ∈ {L, A, I}, we stack all target-token coordinates a…
Figure 7
Figure 7. Figure 7: Depth-wise factor realization. Lightweight probes show a staged profile: layout emerges early, agents grow gradually, and interactions peak in deeper layers. and Yk′ . We penalize linear and nonlinear leakage through Lsep = 2 X k<k′ ∥Γkk′∥ 2 F | {z } Lcov +βnlin X k<k′…
Figure 8
Figure 8. Figure 8: FactorJEPA vs. the strongest competitor across scale and data regimes. Bars report FactorJEPA’s advantage in units of the paired-difference 95% confidence interval; the dashed line marks statistical separation at 1×. Under matched stratified 10k protocol, FactorJEPA se…
Figure 9
Figure 9. Figure 9: Evaluation scorecard across 2B and 1B scales. Frozen, conventionally adapted, parameter-efficient, and factorized V-JEPA 2.1 variants are compared across predictive, motion, semantic, and temporal diagnostics. Bars report performance with 95% BCa intervals where availa…
Figure 10
Figure 10. Figure 10: Causal future-block rankings replicate across model scales. Each point represents one adaptation method, comparing its causal future-block L1 score with the ViT-G 2B backbone on the horizontal axis and the ViT-g 1B back￾bone on the vertical axis; lower values are bett…
Figure 11
Figure 11. Figure 11: Staged factor-curriculum surgery on the full corpus. Training JEPA loss for FactorJEPA predictor surgery on the full DENSEWORLD corpus (∼115k clips; ViT-g 1B backbone, batch 32, learning rate 5×10−5 ). Surgery proceeds in four shaded phases (dashed boundaries): a head…
Figure 12
Figure 12. Figure 12: Factorized predictive channels. Context is split into layout, visibility-gated agent, and sparse interaction coordinates, then recomposed in the future JEPA embedding, limiting cross-factor leakage [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Automatic factor-target construction. Per clip, the pipeline derives layout DL, agent DA, and interaction DI supervision by masking foreground/background and linking agent tubes. agents, or uncertainty collapsed into a single future. Visual plausibility should therefo…
Figure 14
Figure 14. Figure 14: City-level composition by capture mode. The upper panel shows metropolitan collection sites and the lower panel shows additional urban sites. Each bar decomposes a city’s model-ready clips into drive-through, walk-through, and aerial capture. The value at the right gi…
Figure 15
Figure 15. Figure 15: Illustrative DENSEWORLD dynamic-agent taxonomy. The examples communicate the semantic scope of the automated taxonomy, including conventional transport, intermediate mobility, mobile vendors, animal-drawn transport, and independently moving animals. They are illustrat…
Figure 16
Figure 16. Figure 16: Conceptual end-to-end construction of FactorJEPA targets. The upper band shows the frozen teacher-side pipeline: a privacy-filtered clip is processed by DINOv2 to obtain region evidence—masks, boxes, descriptors, and confidence— which is linked into tracklets by tempo…
Figure 17
Figure 17. Figure 17: Complete ViT-G attribution and ablation scorecard. The comparison includes frozen and continually adapted V-JEPA, full and parameter-efficient fine-tuning, Auto-RGN, FactorJEPA-RAW, the executed FactorJEPA objective variants, and the WiseFT encoder-scope variants. Bar…
Figure 18
Figure 18. Figure 18: Illustrative layout for future-consistency inspection. The top row compares the observed context with decoded futures from frozen V-JEPA, V-JEPA Auto-RGN, FactorJEPA-RAW, and FactorJEPA, followed by the target-latent oracle reconstruction and observed future. The iden…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

134 extracted references · 50 canonical work pages

  1. [1]

    arXiv preprint arXiv:2501.03575 , year =

    Cosmos World Foundation Model Platform for Physical AI , author =. arXiv preprint arXiv:2501.03575 , year =

    arXiv

  2. [2]

    Transactions on Machine Learning Research , year =

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year =

  3. [3]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =. 2022 , url =

    2022

  4. [4]

    2024 , publisher =

    Liu, Shih-Yang and Wang, Chien-Yi and Yin, Hongxu and Molchanov, Pavlo and Wang, Yu-Chiang Frank and Cheng, Kwang-Ting and Chen, Min-Hung , booktitle =. 2024 , publisher =

    2024

  5. [5]

    International Conference on Learning Representations , year =

    Surgical Fine-Tuning Improves Adaptation to Distribution Shifts , author =. International Conference on Learning Representations , year =

  6. [6]

    2022 , month = jun, note =

    A Path Towards Autonomous Machine Intelligence , author =. 2022 , month = jun, note =

    2022

  7. [7]

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

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2301.08243 , archivePrefix=

    arXiv

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    V-JEPA: Latent Video Prediction for Visual Representation Learning , author =. International Conference on Learning Representations (ICLR) , year =

Show all 134 references
  1. [9]

    2023 , eprint =

    MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features , author =. 2023 , eprint =

    2023

  2. [10]

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

    Masked Autoencoders Are Scalable Vision Learners , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2111.06377 , archivePrefix =

    arXiv

  3. [11]

    2020 , eprint =

    Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning , author =. 2020 , eprint =

    2020

  4. [12]

    2021 , eprint =

    Emerging Properties in Self-Supervised Vision Transformers , author =. 2021 , eprint =

    2021

  5. [13]

    2022 , eprint =

    data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language , author =. 2022 , eprint =

    2022

  6. [14]

    Advances in Neural Information Processing Systems , volume =

    Object-Centric Learning with Slot Attention , author =. Advances in Neural Information Processing Systems , volume =. 2020 , eprint =

    2020

  7. [15]

    2018 , eprint =

    Relational inductive biases, deep learning, and graph networks , author =. 2018 , eprint =

    2018

  8. [16]

    2019 , eprint =

    Invariant Risk Minimization , author =. 2019 , eprint =

    2019

  9. [17]

    Algorithmic Learning Theory (ALT) , year =

    Measuring statistical dependence with Hilbert-Schmidt norms , author =. Algorithmic Learning Theory (ALT) , year =

  10. [18]

    Geva, Mor and Khashabi, Daniel and Segal, Elad and Khot, Tushar and Roth, Dan and Berant, Jonathan , journal =. Did. 2021 , publisher =

    2021

  11. [19]

    Patel, Arkil and Bhattamishra, Satwik and Goyal, Navin , booktitle =. Are. 2021 , publisher =

    2021

  12. [20]

    Advances in Neural Information Processing Systems , year =

    Training Verifiers to Solve Math Word Problems , author =. Advances in Neural Information Processing Systems , year =

  13. [21]

    Findings of the Association for Computational Linguistics: NAACL 2024 , year =

    Benchmarking Generation and Evaluation Capabilities of Large Language Models for Instruction Controllable Summarization , author =. Findings of the Association for Computational Linguistics: NAACL 2024 , year =

    2024

  14. [22]

    2025 , address =

    Srirag, Dipankar and Joshi, Aditya and Painter, Jordan and Kanojia, Diptesh , booktitle =. 2025 , address =

    2025

  15. [23]

    arXiv preprint arXiv:1904.09751 , year =

    The Curious Case of Neural Text Degeneration , author =. arXiv preprint arXiv:1904.09751 , year =

    1904 arXiv

  16. [24]

    Advances in Neural Information Processing Systems , year =

    Large Language Models are Zero-Shot Reasoners , author =. Advances in Neural Information Processing Systems , year =

  17. [25]

    arXiv preprint arXiv:2304.15004 , year =

    The Many Faces of Emergent Abilities in Large Language Models , author =. arXiv preprint arXiv:2304.15004 , year =

    arXiv

  18. [26]

    Advances in Neural Information Processing Systems , volume =

    Are Emergent Abilities of Large Language Models a Mirage? , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  19. [27]

    Advances in Neural Information Processing Systems , volume =

    Language Models are Few-Shot Learners , author =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  20. [28]

    Transactions on Machine Learning Research , year =

    Emergent Abilities of Large Language Models , author =. Transactions on Machine Learning Research , year =

  21. [29]

    , booktitle =

    Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R. , booktitle =. 2018 , address =

    2018

  22. [30]

    Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages =

    Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank , author =. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages =. 2013 , address =

    2013

  23. [31]

    Advances in Neural Information Processing Systems , volume =

    Character-level Convolutional Networks for Text Classification , author =. Advances in Neural Information Processing Systems , volume =. 2015 , url =

    2015

  24. [32]

    Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

    A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference , author =. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2018 , address =

    2018

  25. [33]

    CoRR , volume =

    Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond , author =. CoRR , volume =. 2016 , url =

    2016

  26. [34]

    Advances in Neural Information Processing Systems , volume =

    Teaching Machines to Read and Comprehend , author =. Advances in Neural Information Processing Systems , volume =. 2015 , url =

    2015

  27. [35]

    Dear Sir or Madam, May I Introduce the

    Rao, Sudha and Tetreault, Joel , booktitle =. Dear Sir or Madam, May I Introduce the. 2018 , address =

    2018

  28. [36]

    Proceedings of the 2nd Workshop on Neural Machine Translation and Generation , pages =

    Controllable Abstractive Summarization , author =. Proceedings of the 2nd Workshop on Neural Machine Translation and Generation , pages =. 2018 , address =

    2018

  29. [37]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    He, Junxian and Kry. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =. 2020 , address =

    2020

  30. [38]

    Transactions of the Association for Computational Linguistics , volume =

    Controllable Summarization with Constrained Markov Decision Process , author =. Transactions of the Association for Computational Linguistics , volume =. 2021 , url =

    2021

  31. [39]

    arXiv preprint arXiv:2412.14093 , year =

    Alignment Faking in Large Language Models , author =. arXiv preprint arXiv:2412.14093 , year =

    arXiv

  32. [40]

    , booktitle =

    Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R. , booktitle =. 2019 , note =

    2019

  33. [41]

    Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22) , pages =

    Predictability and Surprise in Large Generative Models , author =. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22) , pages =. 2022 , publisher =

    2022

  34. [42]

    , booktitle =

    Wang, Alex and Pruksachatkun, Yada and Nangia, Nikita and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R. , booktitle =. 2019 , note =

    2019

  35. [43]

    Nature Machine Intelligence , volume =

    Shortcut Learning in Deep Neural Networks , author =. Nature Machine Intelligence , volume =. 2020 , doi =

    2020

  36. [44]

    Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

  37. [45]

    International Conference on Learning Representations (ICLR) , year =

    Self-Consistency Improves Chain of Thought Reasoning in Large Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  38. [46]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , year =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. Proceedings of the 41st International Conference on Machine Learning (ICML) , year =

  39. [47]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Hierarchical Neural Story Generation , author =. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  40. [48]

    International Conference on Learning Representations (ICLR) , year =

    The Curious Case of Neural Text Degeneration , author =. International Conference on Learning Representations (ICLR) , year =

  41. [49]

    Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

    Discovering Language Model Behaviors with Model-Written Evaluations , author =. Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

  42. [50]

    arXiv preprint , year =

    Red Teaming Language Models with Language Models , author =. arXiv preprint , year =

  43. [51]

    Efficient Reproducible Floating Point Summation and

    Demmel, James and Ahrens, Willow and Nguyen, Hong Diep , institution =. Efficient Reproducible Floating Point Summation and. 2016 , url =

    2016

  44. [52]

    ACM Transactions on Mathematical Software , volume =

    Algorithms for Efficient Reproducible Floating Point Summation , author =. ACM Transactions on Mathematical Software , volume =. 2020 , doi =

    2020

  45. [53]

    Pattern Recognition , volume =

    Randomness in Neural Network Training: Characterizing the Impact of Weight Initialization and Data Shuffling , author =. Pattern Recognition , volume =. 2022 , doi =

    2022

  46. [54]

    2024 , note =

    Atil, Berk and Chittams, Alexander and Fu, Lei and Ture, Ferhan and Xu, Liyan and Baldwin, Tim , journal =. 2024 , note =

    2024

  47. [55]

    2024 , howpublished =

    Reproducible Outputs , author =. 2024 , howpublished =

    2024

  48. [56]

    2020 , note =

    Reproducibility --- Controlling Sources of Randomness in. 2020 , note =

    2020

  49. [57]

    Defeating Nondeterminism in

    He, Russell and others , howpublished =. Defeating Nondeterminism in. 2025 , note =

    2025

  50. [58]

    arXiv preprint arXiv:2305.10601 , year =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. arXiv preprint arXiv:2305.10601 , year =

    arXiv

  51. [59]

    arXiv preprint arXiv:2401.00998 , year =

    On the Unreliability of Zero-Temperature Decoding in Large Language Models , author =. arXiv preprint arXiv:2401.00998 , year =

    arXiv

  52. [60]

    2025 , note =

    Defeating Nondeterminism in LLM Inference , author =. 2025 , note =

    2025

  53. [61]

    Proceedings of SC , year =

    Deterministic Replay for Scientific Debugging of Large-Scale Parallel Programs , author =. Proceedings of SC , year =

  54. [62]

    arXiv preprint arXiv:2501.00000 , year =

    The Silent Culprit: Floating-Point Non-Determinism in Large Language Models , author =. arXiv preprint arXiv:2501.00000 , year =

  55. [63]

    arXiv preprint arXiv:2404.00000 , year =

    Are Large Language Models Deterministic? A Study of Output Variance and Semantic Stability , author =. arXiv preprint arXiv:2404.00000 , year =

  56. [64]

    arXiv preprint arXiv:2405.00000 , year =

    Variance, Seeds, and Stability: Evaluating LLMs Under Repeated Sampling , author =. arXiv preprint arXiv:2405.00000 , year =

  57. [65]

    2024 , note =

    Deterministic Inference in SGLang , author =. 2024 , note =

    2024

  58. [66]

    Le and Ed H

    Xuezhi Wang and Jason Wei and Dale Schuurmans and Quoc V. Le and Ed H. Chi and Sharan Narang and Aakanksha Chowdhery and Denny Zhou , title =. arXiv preprint , volume =. 2023 , url =

    2023

  59. [67]

    2013 IEEE 21st Symposium on Computer Arithmetic (ARITH) , year =

    James Demmel and Hong Diep Nguyen , title =. 2013 IEEE 21st Symposium on Computer Arithmetic (ARITH) , year =

    2013

  60. [68]

    Patterns , year =

    Philipp Heumos and others , title =. Patterns , year =

  61. [69]

    2024 , note =

    Reproducibility --- Controlling Sources of Non-determinism in. 2024 , note =

    2024

  62. [70]

    arXiv preprint , volume =

    Jungseob Kwon and others , title =. arXiv preprint , volume =. 2023 , url =

    2023

  63. [71]

    Gonzalez and Clark Barrett and Ying Sheng , title =

    Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Chuyue Sun and Jeff Huang and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark Barrett and Ying Sheng , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  64. [72]

    Passonneau and Evan Radcliffe and Guru Rajan Rajagopal and Adam Sloan and Tomasz Tudrej and Ferhan Ture and Zhe Wu and Lixinyu Xu and Breck Baldwin , title =

    Berk Atil and Sarp Aykent and Alexa Chittams and Lisheng Fu and Rebecca J. Passonneau and Evan Radcliffe and Guru Rajan Rajagopal and Adam Sloan and Tomasz Tudrej and Ferhan Ture and Zhe Wu and Lixinyu Xu and Breck Baldwin , title =. arXiv preprint , volume =. 2024 , url =

    2024

  65. [73]

    Proceedings of the 2025 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , year =

    Lanqing Song and others , title =. Proceedings of the 2025 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , year =

    2025

  66. [74]

    Larger and More Instructable Language Models Become Less Reliable , journal =

    Lexin Zhou and Wout Schellaert and Fernando Mart. Larger and More Instructable Language Models Become Less Reliable , journal =. 2024 , doi =

    2024

  67. [75]

    arXiv preprint , volume =

    Jie Huang and others , title =. arXiv preprint , volume =. 2023 , url =

    2023

  68. [76]

    Nature , year =

    Sebastian Farquhar and others , title =. Nature , year =

  69. [77]

    arXiv preprint , volume =

    Momchil Manakul and Yarin Gal , title =. arXiv preprint , volume =. 2023 , url =

    2023

  70. [78]

    Artificial Intelligence Review , year =

    Xiaowei Huang and others , title =. Artificial Intelligence Review , year =

  71. [79]

    arXiv preprint , volume =

    Jared Kaplan and Sam McCandlish and Tom Henighan and others , title =. arXiv preprint , volume =. 2020 , url =

    2020

  72. [80]

    Advances in Neural Information Processing Systems , volume =

    Long Ouyang and Jeffrey Wu and Xu Jiang and others , title =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  73. [81]

    Khan , title =

    Aaqib Kaikaus and Adeel Anwar and Yusuf M. Khan , title =. arXiv preprint , volume =. 2024 , url =

    2024

  74. [82]

    Lipton and Srinivasan Parthasarathy and Weiwei Tu , title =

    Yang Zhuang and Zachary C. Lipton and Srinivasan Parthasarathy and Weiwei Tu , title =. Proceedings of the 3rd Conference on Machine Learning and Systems (MLSys) , year =

  75. [83]

    Edward Suh and Christina Delimitrou and Christopher Batten and Edward F

    Yan Chen and Yu Huang and Xiang Li and Houyi Li and Yanjie Zhao and Zhiru Zhang and G. Edward Suh and Christina Delimitrou and Christopher Batten and Edward F. Redmond and others , title =. Proceedings of the 44th International Conference on Software Engineering (ICSE) , year =

  76. [84]

    arXiv preprint , volume =

    Ziyang Zhang and Xinheng Ding and Jiayi Yuan and Rixin Liu and Huizi Mao and Jiarong Xing and Zirui Liu , title =. arXiv preprint , volume =. 2025 , doi =

    2025

  77. [85]

    Proceedings of the 8th International Conference on Learning Representations (ICLR) , year =

    Ari Holtzman and Jan Buys and Li Du and Maxwell Forbes and Yejin Choi , title =. Proceedings of the 8th International Conference on Learning Representations (ICLR) , year =

  78. [86]

    Gonzalez and Clark Barrett and Ying Sheng , title =

    Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Chuyue Sun and Jeff Huang and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark Barrett and Ying Sheng , title =. 2024 , url =

    2024

  79. [87]

    Machine Learning and Knowledge Extraction , year =

    Mostafa Shahriari and Rudolf Ramler and Lukas Fischer , title =. Machine Learning and Knowledge Extraction , year =. doi:10.3390/make4040045 , url =

    doi

  80. [88]

    arXiv preprint , volume =

    Prabhat Nagarajan and Garrett Warnell and Peter Stone , title =. arXiv preprint , volume =. 2018 , url =

    2018

  81. [89]

    Proceedings of the 2023

    Odd Erik Gundersen and Saeid Shamsaliei and Hakon Sletten Kjaernli and Helge Langseth , title =. Proceedings of the 2023. 2023 , publisher =. doi:10.1145/3589806.3600044 , url =

    2023

  82. [90]

    Evan Hubinger and Carson Denison and Jesse Mu and Mike Lambert and Meg Tong and Monte MacDiarmid and Tamera Lanham and Daniel M. Ziegler and Tim Maxwell and Newton Cheng and Adam Jermyn and Amanda Askell and Ansh Radhakrishnan and Cem Anil and David Duvenaud and Deep Ganguli a...

    2024

  83. [91]

    , title =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning (ICML) , year =

  84. [92]

    arXiv preprint arXiv:1912.02280 , year =

    Natural Alpha Embeddings , author =. arXiv preprint arXiv:1912.02280 , year =

    1912 arXiv

  85. [93]

    Information Geometry and Its Applications , author =

  86. [94]

    2021 , month = oct, doi =

    Seven Transformations for More Equitable and Sustainable Cities , author =. 2021 , month = oct, doi =

    2021

  87. [95]

    2016 , doi =

    Leveraging Urbanization in South Asia: Managing Spatial Transformation for Prosperity and Livability , author =. 2016 , doi =

    2016

  88. [96]

    Transportation Research Record: Journal of the Transportation Research Board , volume =

    Modeling Heterogeneous Traffic Flow , author =. Transportation Research Record: Journal of the Transportation Research Board , volume =. 1999 , publisher =. doi:10.3141/1678-28 , url =

    1999 doi

  89. [97]

    European Transport Research Review , volume =

    Flexible Car-Following Models for Mixed Traffic and Weak Lane-Discipline Conditions , author =. European Transport Research Review , volume =. 2018 , publisher =. doi:10.1186/s12544-018-0338-0 , url =

    2018 doi

  90. [98]

    Transportation Research Record: Journal of the Transportation Research Board , volume =

    Characteristics of Mixed Traffic on Urban Arterials with Significant Volumes of Motorized Two-Wheelers: Role of Composition, Intraclass Variability, and Lack of Lane Discipline , author =. Transportation Research Record: Journal of the Transportation Research Board , volume =....

    2012 doi

  91. [99]

    Varma, Girish and Subramanian, Anbumani and Namboodiri, Anoop and Chandraker, Manmohan and Jawahar, C. V. , booktitle =. 2019 , doi =. 1811.10200 , archivePrefix =

    2019 arXiv

  92. [100]

    Dokania, Shubham and Hafez, A. H. Abdul and Subramanian, Anbumani and Chandraker, Manmohan and Jawahar, C. V. , booktitle =. 2023 , doi =. 2210.12878 , archivePrefix =

    2023 arXiv

  93. [101]

    2020 , eprint =

    Yu, Fisher and Chen, Haofeng and Wang, Xin and Xian, Wenqi and Chen, Yingying and Liu, Fangchen and Madhavan, Vashisht and Darrell, Trevor , booktitle =. 2020 , eprint =

    2020

  94. [102]

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

    nuScenes: A Multimodal Dataset for Autonomous Driving , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2020 , doi =. 1903.11027 , archivePrefix =

    2020 arXiv

  95. [103]

    OpenReview , year =

    A Path Towards Autonomous Machine Intelligence , author =. OpenReview , year =

  96. [104]

    Advances in Neural Information Processing Systems 37 , year =

    SAFE: Slow And Fast parameter-Efficient tuning for Continual Learning with Pre-Trained Models , author =. Advances in Neural Information Processing Systems 37 , year =

  97. [105]

    arXiv preprint arXiv:2304.07193 , year =

    DINOv2: Learning Robust Visual Features without Supervision , author =. arXiv preprint arXiv:2304.07193 , year =

    arXiv

  98. [106]

    arXiv preprint arXiv:2103.00020 , year =

    Learning Transferable Visual Models From Natural Language Supervision , author =. arXiv preprint arXiv:2103.00020 , year =

    arXiv

  99. [107]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    SAPT: A Shared Attention Framework for Parameter-Efficient Continual Learning of Large Language Models , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2024 , address =. doi:10.18653/v1/2024....

    2024 doi

  100. [108]

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

    Semantically-Shifted Incremental Adapter-Tuning is A Continual ViTransformer , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2024 , eprint =

    2024

  101. [109]

    Proceedings of the IEEE International Conference on Multimedia and Expo (ICME) , year =

    Audio-JEPA: Joint-Embedding Predictive Architecture for Audio Representation Learning , author =. Proceedings of the IEEE International Conference on Multimedia and Expo (ICME) , year =. 2507.02915 , archivePrefix =

    arXiv

  102. [110]

    arXiv preprint arXiv:2404.16432 , year =

    Point-JEPA: A Joint Embedding Predictive Architecture for Self-Supervised Learning on Point Cloud , author =. arXiv preprint arXiv:2404.16432 , year =. 2404.16432 , archivePrefix =

    arXiv

  103. [111]

    2024 , eprint =

    Hu, Naiwen and Cheng, Haozhe and Xie, Yifan and Li, Shiqi and Zhu, Jihua , journal =. 2024 , eprint =

    2024

  104. [112]

    arXiv preprint arXiv:2501.14622 , year =

    ACT-JEPA: Joint-Embedding Predictive Architecture Improves Policy Representation Learning , author =. arXiv preprint arXiv:2501.14622 , year =. 2501.14622 , archivePrefix =

    arXiv

  105. [113]

    arXiv preprint arXiv:2602.11389 , year =

    Causal-JEPA: Learning World Models through Object-Level Latent Interventions , author =. arXiv preprint arXiv:2602.11389 , year =. 2602.11389 , archivePrefix =

    arXiv

  106. [114]

    arXiv preprint arXiv:2511.08544 , year =

    LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics , author =. arXiv preprint arXiv:2511.08544 , year =. 2511.08544 , archivePrefix =

    arXiv

  107. [115]

    arXiv preprint arXiv:2603.19312 , year =

    LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels , author =. arXiv preprint arXiv:2603.19312 , year =. 2603.19312 , archivePrefix =

    arXiv

  108. [116]

    arXiv preprint arXiv:2506.09985 , year =

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning , author =. arXiv preprint arXiv:2506.09985 , year =

    arXiv

  109. [117]

    2026 , eprint =

    Mur-Labadia, Lorenzo and Muckley, Matthew and Bar, Amir and Assran, Mido and Sinha, Koustuv and Rabbat, Mike and LeCun, Yann and Ballas, Nicolas and Bardes, Adrien , journal =. 2026 , eprint =

    2026

  110. [118]

    arXiv preprint arXiv:2603.22281 , year =

    ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model , author =. arXiv preprint arXiv:2603.22281 , year =. 2603.22281 , archivePrefix =

    arXiv

  111. [119]

    2024 , publisher =

    He, Jinghan and Guo, Haiyun and Zhu, Kuan and Zhao, Zihan and Tang, Ming and Wang, Jinqiao , booktitle =. 2024 , publisher =. doi:10.18653/v1/2024.emnlp-main.190 , eprint =

    2024 doi

  112. [120]

    Advances in Neural Information Processing Systems , volume =

    Interaction Networks for Learning about Objects, Relations and Physics , author =. Advances in Neural Information Processing Systems , volume =. 2016 , eprint =

    2016

  113. [121]

    Proceedings of the 38th International Conference on Machine Learning , series =

    Barlow Twins: Self-Supervised Learning via Redundancy Reduction , author =. Proceedings of the 38th International Conference on Machine Learning , series =. 2021 , eprint =

    2021

  114. [122]

    2022 , eprint =

    Bardes, Adrien and Ponce, Jean and LeCun, Yann , booktitle =. 2022 , eprint =

    2022

  115. [123]

    Computer Vision -- ECCV 2016 Workshops , series =

    Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking , author =. Computer Vision -- ECCV 2016 Workshops , series =. 2016 , doi =. 1609.01775 , archivePrefix =

    2016 arXiv

  116. [124]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    The ``Something Something'' Video Database for Learning and Evaluating Visual Common Sense , author =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =. 2017 , doi =. 1706.04261 , archivePrefix =

    2017 arXiv

  117. [125]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    Resound: Towards Action Recognition without Representation Bias , author =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =. 2018 , publisher =. doi:10.1007/978-3-030-01231-1_32 , url =

    2018 doi

  118. [126]

    2020 , doi =

    Teed, Zachary and Deng, Jia , booktitle =. 2020 , doi =. 2003.12039 , archivePrefix =

    2020 arXiv

  119. [127]

    Computer Vision -- ECCV 2016 , series =

    Shuffle and Learn: Unsupervised Learning Using Temporal Order Verification , author =. Computer Vision -- ECCV 2016 , series =. 2016 , doi =. 1603.08561 , archivePrefix =

    2016 arXiv

  120. [128]

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

    Self-Supervised Video Representation Learning with Odd-One-Out Networks , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2017 , eprint =

    2017

  121. [129]

    Advances in Neural Information Processing Systems , volume =

    Masked Autoencoders As Spatiotemporal Learners , author =. Advances in Neural Information Processing Systems , volume =. 2022 , eprint =

    2022

  122. [130]

    International Journal of Computer Vision , volume =

    The Pascal Visual Object Classes (VOC) Challenge , author =. International Journal of Computer Vision , volume =. 2010 , publisher =. doi:10.1007/s11263-009-0275-4 , url =

    2010 doi

  123. [131]

    arXiv preprint arXiv:2511.16719 , year =

    Carion, Nicolas and Gustafson, Laura and Hu, Yuan-Ting and Debnath, Shoubhik and Hu, Ronghang and Suris, Didac and Ryali, Chaitanya and Alwala, Kalyan Vasudev and Khedr, Haitham and Huang, Andrew and Lei, Jie and Ma, Tengyu and Guo, Baishan and Kalla, Arpit and Marks, Markus a...

    arXiv

  124. [132]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

    Segment Anything , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

  125. [133]

    arXiv preprint arXiv:2408.00714 , year=

    SAM 2: Segment Anything in Images and Videos , author=. arXiv preprint arXiv:2408.00714 , year=

    arXiv

  126. [134]

    Naval Research Logistics Quarterly , volume =

    The Hungarian Method for the Assignment Problem , author =. Naval Research Logistics Quarterly , volume =. 1955 , publisher =. doi:10.1002/nav.3800020109 , url =

    1955 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.