Pith. sign in

REVIEW 3 major objections 4 minor 34 references

Temporal-Distance-JEPA mines a directed, asymmetric temporal cost from reward-free demonstration logs and shows that using this cost as the plan-time objective or as representation shaping lets a latent JEPA world-model planner match or exc

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:43 UTC pith:Y6FUM2AU

load-bearing objection Solid incremental extension of LeWM, but the 'every environment' claim only holds under a per-task cost selection made after seeing the results; worth peer review with revision. the 3 major comments →

arxiv 2607.25337 v2 pith:Y6FUM2AU submitted 2026-07-28 cs.CL cs.RO

Temporal-Distance JEPA: Plan-Aware Representation Learning for Latent World Model Predictive Control

Jiaxin Bai , Jiaxuan Xiong This is my paper
classification cs.CL cs.RO
keywords temporal-distance learningJEPA world modelslatent model predictive controlquasimetric costcross-entropy methodoffline reward-free learninggoal-conditioned planningrepresentation learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that JEPA world models trained for latent prediction do not automatically provide a plan-time ranking of imagined futures; planning usually falls back on Euclidean distance in embedding space, which correlates only weakly with actual progress. It introduces Temporal-Distance-JEPA, which keeps the LeWM backbone and adds a directed temporal cost mined from reward-free demonstration logs: same-trajectory step order gives positive targets, cross-trajectory pairs give negative hinges, and a rollout-consistency loss matches the planner horizon. This mined signal plays two roles: it is deployed directly as the planning cost where progress is topological, and used as representation shaping that improves geometric planning where contact dominates. Under locked evaluation with the cost chosen per task, the method matches or exceeds both LeWM and RC-aux on all four environments, including 100.0% Two-Room success and a 14.2-point gain over LeWM on OGB-Cube. A sympathetic reader would care because it offers an offline, reward-free route to a genuine progress signal for latent MPC without learning a value function.

Core claim

Temporal-Distance-JEPA discovers temporal progress structure directly from offline demonstration trajectories and uses it to supervise the same cost that the planner optimizes, rather than letting plan-time ranking be an untrained byproduct of embedding geometry. The central claim is that this mined, directed temporal distance—a quasimetric-style energy with a symmetric component plus an asymmetric residual—outperforms plain latent Euclidean distance on navigation and reaching when deployed directly, and also improves Euclidean planning on contact-rich tasks when used only as representation shaping. Under locked evaluation, the resulting planner matches or exceeds LeWM and RC-aux on every en

What carries the argument

The central object is d_psi, a directed, non-negative temporal cost parameterized by a metric-residual network: d_psi(z_s, z_g) = ||phi_sym(z_s) - phi_sym(z_g)||^2 + max_k ReLU(phi_asym,k(z_s) - phi_asym,k(z_g)). The symmetric term captures configuration similarity; the asymmetric residual encodes directed reachability, so the cost from A to B can differ from the cost from B to A. This cost is calibrated by regression to same-trajectory step counts tau(i,j) = j - i, pushed above a margin on cross-trajectory pairs, and coupled to the planner through an H-step rollout-consistency loss applied to the same open-loop operator used at plan time. At deployment, d_psi is used directly on topology-do

Load-bearing premise

The evaluation lets the authors choose per task whether to plan with the mined temporal cost or with latent L2, and that choice is made from the same locked results used to claim superiority; if the cost had to be fixed in advance, the advantage on Push-T and OGB-Cube would disappear, since pure d_psi yields 69.0% and 77.0% there.

What would settle it

Fix a single plan cost in advance for all four tasks—either always latent L2 or always d_psi—and re-run the locked evaluation; the paper's own Table 11 shows pure d_psi falling to 69.0% on Push-T and 77.0% on OGB-Cube, so any pre-registered single-cost protocol that reproduces those numbers would refute the 'matches or exceeds on every environment' claim. Alternatively, an independent check of Fig. 1's Spearman correlations (0.95 for d_psi versus 0.65 for L2 on held-out Push-T pairs) would test whether the mined cost truly ranks temporal separation as claimed.

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

If this is right

  • Deploying the mined temporal cost on topology-dominated tasks raises Two-Room success to 100.0% and Reacher to 97.0%, exceeding LeWM's geometric planner on both.
  • On contact-rich tasks, temporal supervision improves geometric planning on the same checkpoint: OGB-Cube rises 14.2 points over LeWM and Push-T improves, even though the planner still uses latent L2.
  • The method requires no reward model, value function, or policy: progress supervision comes entirely from reward-free demonstration logs.
  • The directed head, cross-trajectory negative hinge, and rollout-consistency loss each contribute; removing any one degrades all planner settings on Push-T.
  • Better temporal ranking does not by itself mean a better contact-rich plan cost: d_psi tracks step order with Spearman 0.91 versus 0.79 for L2, yet pure d_psi underperforms L2 on Push-T and OGB-Cube.
  • The paper's ablations show that a symmetric Euclidean head, even with the same temporal labels, sharply underperforms the directed MRN head, establishing that cost form and supervision are jointly load-bearing.

Where Pith is reading between the lines

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

  • If the per-task cost-selection rule (d_psi for navigation, L2 for contact-rich manipulation) is not fixed in advance, the headline 'matches or exceeds on every environment' rests on an oracle-like choice; the paper's own Table 11 shows that pure d_psi drops to 69.0% on Push-T and 77.0% on OGB-Cube, so the transferable claim is likely the representation-shaping effect, not the per-task cost deploym
  • The same mined step-count signal could plausibly serve as a dense reward or goal-reaching value surrogate for offline RL, since it calibrates to steps-to-goal along the demonstration manifold; the paper does not test this, but the calibration makes it a natural next probe.
  • A learned gating network that switches between d_psi and L2 based on predicted contact phase would directly test the two-role thesis; the paper's hand-gated variants failing to beat fixed L2 suggests the geometric signal must act throughout CEM search, not only near contact.
  • The 14.2-point OGB-Cube gain under shared L2 planning deserves scrutiny: if it survives a pre-registered single-cost protocol, it isolates a representation-level effect of temporal supervision independent of any plan-cost change.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Temporal-Distance-JEPA, an extension of the LeWM JEPA world-model backbone that mines a directed, quasimetric-style temporal cost d_psi from reward-free demonstration logs. Same-trajectory step counts provide regression targets, cross-trajectory pairs provide hinge negatives, and a horizon-matched rollout-consistency loss couples the learned cost to the open-loop planner rollout. At plan time, the paper deploys d_psi on Two-Room and Reacher and latent Euclidean distance on Push-T and OGB-Cube, claiming that under this locked protocol Temporal-Distance-JEPA matches or exceeds both LeWM and the RC-aux baseline on every environment. The paper also presents ablations (directed head, cross-trajectory hinge, rollout consistency), a plan-cost comparison matrix, and diagnostics explaining why pure d_psi underperforms on contact-rich manipulation.

Significance. The core idea is timely and potentially useful: explicitly mining temporal progress from reward-free logs, rather than relying on embedding geometry, is a sensible way to close the train–plan gap in JEPA planners. The paper is also commendable for reporting matched plan-cost matrices, conducting ablations, and releasing code. However, the headline claim—'matches or exceeds both methods on every environment'—is only true under a per-environment plan-time cost selection that is made after inspecting the locked evaluation outcomes. Because no single cost achieves the claim, and because the navigation/contact taxonomy is applied to the same four test environments, the central comparison is not a fair test of the method's superiority. The contribution remains plausible, but the evaluation protocol must be strengthened before the claim is supported.

major comments (3)
  1. [§4.1, Table 2 and Appendix Table 11] The claim that Temporal-Distance-JEPA 'matches or exceeds both methods on every environment' is not supported by a single protocol. Table 11 shows pure d_psi yields 69.0% on Push-T (below LeWM 83.6 and RC-aux 81.4) and 77.0% on OGB-Cube (below RC-aux 81.6), while pure latent L2 yields 95.0% on Reacher (below LeWM 96.0). The Table 2 result is therefore obtained by selecting d_psi for Two-Room/Reacher and L2 for Push-T/OGB-Cube. This per-environment selection is made on the same locked outcomes that are then reported as the headline result. Unless the selection rule was fixed before seeing the results, or an independent validation split is used to choose the cost per environment, this is test-set selection and the 'every environment' claim is inflated.
  2. [§4.5, Table 10; §4.3, Fig. 3] The Reacher result depends on a non-default aggregation rule. Table 10 shows that with w=0.3 (primary) Temporal-Distance-JEPA scores 96% vs LeWM 94%, but with terminal-only scoring (w=1) the comparison reverses: 74% vs 86%. This is a second protocol choice that is selected to make the method look favorable. Similarly, the two-role taxonomy (topology -> d_psi, contact -> L2) is plausible but is justified only by the same four locked outcomes in Fig. 3/Table 11. No independent validation, pre-registration, or leave-one-out procedure is provided. This is load-bearing because the central claim is precisely that the method helps 'on every environment'; under fixed costs or terminal-only aggregation the claim fails on at least one environment.
  3. [§4.4, Table 9] The contact-gate experiment tests hand-designed gates (hard and soft) and finds they do not recover pure L2. This is informative, but it does not address the protocol-selection problem. A learned or validated rule that chooses the plan-time cost from training/validation episodes would be needed to establish that the per-environment choice is not an artifact of peeking at the test manifest. As written, the gate results only reinforce that pure d_psi is suboptimal on Push-T and that the choice of planning cost is consequential.
minor comments (4)
  1. [Table 3] The literature-context table lists LeWM as 87/86/96/74 on the four environments, while the matched Table 2 reports 97.4/96.0/83.6/68.0. The text does state these rows are not matched, but the discrepancy is large and may confuse readers; consider moving the table to an appendix or adding a stronger visual warning.
  2. [Table 6] The Spearman correlations for d_psi are partly by construction, since d_psi is regressed directly to temporal separation tau in Eq. (6). The paper already frames this as calibration rather than transfer, but it would help to state explicitly that Table 6 does not provide independent evidence of temporal ranking quality.
  3. [§3.6, Eq. (9)] The loss weights lambda_roll=0.5, lambda_td=1.0, lambda_sig=0.09 and margin scale eta are fixed constants. No sensitivity analysis is reported for these hyperparameters, even though the claimed gains are relatively small on Reacher and Push-T. A small sweep or at least a statement that the results are robust to reasonable variations would strengthen the paper.
  4. [Appendix C] The appendix says RC-aux uses the same locked manifests and latent L2 planning, but it is unclear whether this is the authors' re-implementation or the original code. For reproducibility, specify the source and version of the RC-aux baseline, and similarly for LeWM checkpoints.

Circularity Check

0 steps flagged

No material circularity: d_psi is explicitly trained to temporal step labels, and the headline gains are measured against external LeWM/RC-aux baselines; the per-task plan-cost choice is a protocol/selection concern, not a circular derivation.

full rationale

The derivation chain is not circular. The directed cost d_psi is explicitly fit by Eq. (6) to same-trajectory step counts tau(i,j)=j-i and cross-trajectory hinges; using that cost at plan time (Eq. 7) is the method's stated training target, not a hidden redeployment of the evaluation metric. The success numbers in Table 2 are external goal-reaching rates on locked manifests, compared against LeWM and RC-aux; none of those baselines or the goal-success labels appear in the training objective (Eq. 9). The quasimetric/MRN form is imported from published external work [24], not from a self-citation chain, and no load-bearing claim is justified solely by the authors' prior results. The only caveat is procedural rather than circular: the plan-time cost (d_psi vs l2) and Reacher aggregation weight w=0.3 are selected per environment from the same locked results (Tables 2/11/10); with a single fixed cost, no set of environments satisfies the 'every environment' claim (pure d_psi gives Push-T 69.0 vs LeWM 83.6; pure l2 gives Reacher 95.0 vs LeWM 96.0). This is a test-set selection-bias concern, not an equation-level reduction of the claimed result to its inputs, so it does not constitute circularity under the stated criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

Central claims rest on a learned regression to demonstration step counts plus heuristic negatives; no new physical entities are introduced. The main untested leaps are generalization from on-manifold labels to open-loop rollouts and the per-environment cost selection rule.

free parameters (5)
  • Loss weights lambda_roll=0.5, lambda_td=1.0, lambda_sig=0.09 = 0.5 / 1.0 / 0.09
    Hand-selected balance of rollout-consistency, temporal-distance, and anti-collapse terms in Eq. (9); no sensitivity sweep reported.
  • Margin scale eta in m = eta*(T-1) = unspecified
    Margin for cross-trajectory hinge in Eq. (6); T=8 is given but eta is not, so the threshold is under-specified.
  • Planner horizon H=5 and prediction window N_pred=5 = 5 / 5
    Choice determines L_roll and the CEM rollout length; matched to each other but not swept.
  • Per-environment plan cost selection = d_psi on Two-Room/Reacher; L2 on Push-T/OGB-Cube
    Chosen from the matched cost matrix (Table 11 / Fig. 3); drives headline gains and is a post-hoc model selection.
  • Blend weight alpha=0.10 = 0.10
    Used for diagnostic blend only, not the main protocol; fixed without a sweep.
axioms (5)
  • domain assumption Demonstration step count tau = j-i is an on-manifold proxy for steps-to-goal: D*(o_i,o_j) <= tau and d_psi approximately tau ranks progress (Eq. 4).
    Authors state this is an empirical ranking-calibration step, not a shortest-path metric; if the proxy is poor in a task, d_psi planning fails (as seen on Push-T).
  • domain assumption Cross-trajectory pairs at equal time indices are sufficiently valid negatives; false negatives are tolerated (Eq. 6).
    The hinge pushes off-trajectory d_psi above the margin; if many cross-trajectory pairs are actually reachable, the learned cost is distorted.
  • domain assumption The LeWM predictor remains accurate over H=5 open-loop rollouts after adding L_td and L_roll, so minimizing d_psi(z_hat_H, z_g) is meaningful.
    L_roll reduces but does not eliminate compounding error; no formal bound is provided.
  • ad hoc to paper Task structure maps to cost form: d_psi for topology-dominated tasks (Two-Room, Reacher), L2 for contact-rich tasks (Push-T, OGB-Cube).
    The mapping is asserted from the matched cost matrix rather than a pre-registered rule; it is load-bearing for the headline 'every environment' claim.
  • standard math Quasimetric/energy-based background: under deterministic unit-cost dynamics, D* is a quasimetric and the MRN form can represent it (Sec. 3.2, [24]).
    Standard theoretical grounding for asymmetric goal-reaching costs; not proven for the learned latent space.

pith-pipeline@v1.3.0-alltime-deepseek · 14929 in / 15836 out tokens · 147017 ms · 2026-08-01T02:43:03.218712+00:00 · methodology

0 comments
read the original abstract

Joint-Embedding Predictive Architectures (JEPAs) learn world models by predicting in representation space rather than reconstructing pixels, making them a natural backbone for latent model predictive control from offline demonstration logs. JEPA-style training optimizes short-horizon latent prediction, whereas planning requires a multi-step ranking of imagined futures by goal progress. Prior JEPA planners often inherit that ranking from embedding geometry, typically latent Euclidean distance, which arises as a byproduct of representation learning rather than as a progress cost mined from the logs. We propose Temporal-Distance-JEPA, which retains the LeWM encoder--predictor backbone and mines a directed temporal cost from reward-free trajectories: same-trajectory step order supplies positive targets, cross-trajectory pairs act as heuristic negatives, and a rollout-consistency term matches the planner horizon. The mined supervision serves two roles: as the deployed planning cost when progress is topological, and as a representation signal that improves Euclidean planning when contact geometry dominates. Under locked evaluation, deploying the mined cost raises Two-Room success to 100.0% versus LeWM's 97.4%, while shared Euclidean planning on the same temporally trained checkpoint raises OGB-Cube by 14.2 points over LeWM and improves Push-T. Against LeWM and the concurrent RC-aux baseline under locked evaluation, Temporal-Distance-JEPA matches or exceeds both methods on every environment. Ablations show that the directed head, cross-trajectory negatives, and rollout consistency each contribute. Temporal-Distance-JEPA narrows the train--plan gap for JEPA world-model planners by discovering temporal progress structure in offline logs and co-designing cost form with plan-time deployment. Code is available at https://github.com/HKBU-KnowComp/Temporal-Distance-JEPA.

Figures

Figures reproduced from arXiv: 2607.25337 by Jiaxin Bai, Jiaxuan Xiong.

Figure 1
Figure 1. Figure 1: Push-T temporal-gap alignment on held-out [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Temporal-Distance-JEPA training and latent MPC pipeline. Observations are encoded into latents; the LeWM pre [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Two-Room planning success versus training epoch [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Offline CEM diagnostics on manipulation. (a) OGB-Cube success across training epochs under offline-CEM replan [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learned 𝑑𝜓 versus held-out temporal separation. C Training and Evaluation Protocol Training. The short-window recipe uses 𝑁pred=5 (window 𝑇=8 with history size 3), rollout horizon 5, 𝜆roll=0.5, 𝜆td=1.0, and 𝜆sig=0.09. Directed head: hidden dimension 512; 𝜙sym, 𝜙asym ∈ R 128 . All runs share a ViT-tiny encoder (∼15M parameters, patch 14, image 224), 192-dimensional latent, history size 3, AdamW with learnin… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 12 linked inside Pith

  1. [1]

    Junik Bae, Kwanyoung Park, and Youngwoon Lee. 2024. TLDR: Unsupervised Goal-Conditioned Reinforcement Learning via Temporal Distance-Aware Rep- resentations. InProceedings of the Conference on Robot Learning

    2024

  2. [2]

    Randall Balestriero and Yann LeCun. 2025. LeJEPA: Provable and Scalable Self- Supervised Learning Without the Heuristics. arXiv:2511.08544 [cs.LG]

    Pith/arXiv arXiv 2025

  3. [3]

    Kroese, Shie Mannor, and Reuven Y

    Pieter-Tjerk de Boer, Dirk P. Kroese, Shie Mannor, and Reuven Y. Rubinstein

  4. [4]

    Matthieu Destrade, Oumayma Bounou, Quentin Le Lidec, Jean Ponce, and Yann LeCun. 2026. Value-Guided Action Planning with JEPA World Models. arXiv:2601.00844 [cs.LG]

    arXiv 2026

  5. [5]

    Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Ben- jamin Eysenbach, and Sergey Levine. 2020. Learning to Reach Goals via Iterated Supervised Learning. arXiv:1912.06088 [cs.LG] https://arxiv.org/abs/1912.06088

    Pith/arXiv arXiv 2020

  6. [6]

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. 2019. Learning Latent Dynamics for Plan- ning from Pixels. InProceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). PMLR, 2555–2565

    2019

  7. [7]

    Hansen, Hao Su, and Xiaolong Wang

    Nicklas A. Hansen, Hao Su, and Xiaolong Wang. 2022. Temporal Difference Learning for Model Predictive Control. InProceedings of the 39th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 162). PMLR, 8387–8406

    2022

  8. [8]

    Anthony Kobanda, Waris Radji, Mathieu Petitbois, Odalric-Ambrym Maillard, and Remy Portelas. 2025. Offline Goal-Conditioned Reinforcement Learning with Projective Quasimetric Planning. arXiv:2506.18847 [cs.LG]

    arXiv 2025

  9. [9]

    Ilya Kostrikov, Ashvin Nair, and Sergey Levine. 2021. Offline Reinforcement Learning with Implicit Q-Learning. arXiv:2110.06169 [cs.LG]

    Pith/arXiv arXiv 2021

  10. [10]

    Yann LeCun. 2022. A Path Towards Autonomous Machine Intelligence. Open- Review preprint. Version 0.9.2. https://openreview.net/forum?id=BZ5a1r-kVsf

    2022

  11. [11]

    Yann LeCun, Sumit Chopra, Raia Hadsell, Marc’Aurelio Ranzato, and Fu Jie Huang. 2006. A Tutorial on Energy-Based Learning. InPredicting Structured Data. MIT Press, Cambridge, MA, USA

    2006

  12. [12]

    Fei-Fei Li. 2026. A Functional Taxonomy of World Models: Renderers, Simula- tors, Planners, and the Loop That Connects Them. Substack. Published June 3,

    2026

  13. [13]

    Wenyuan Li, Guang Li, Keisuke Maeda, Takahiro Ogawa, and Miki Haseyama

  14. [14]

    Yecheng Jason Ma, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Ja- yaraman. 2023. LIV: Language-Image Representations and Rewards for Robotic Control. InProceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202). PMLR, 23301–23320

    2023

  15. [15]

    Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. 2023. VIP: Towards Universal Visual Reward and Repre- sentation via Value-Implicit Pre-Training. InInternational Conference on Learn- ing Representations

    2023

  16. [16]

    arXiv:2605.07278 [cs.LG]

    Predictive but Not Plannable: RC-aux for Latent World Models. arXiv:2605.07278 [cs.LG]

    Pith/arXiv arXiv

  17. [17]

    Sergi Masip, Jonathan Swinnen, Yutong Hu, Renaud Detry, and Tinne Tuyte- laars. 2026. FF-JEPA: Long-Horizon Planning in World Models with Latent Plan- ners. arXiv:2606.09311 [cs.RO]

    Pith/arXiv arXiv 2026

  18. [18]

    Vivek Myers, Bill Chunyuan Zheng, Benjamin Eysenbach, and Sergey Levine

  19. [19]

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. 2026. LeWorldModel: Stable End-to-End Joint-Embedding Predic- tive Architecture from Pixels. arXiv:2603.19312 [cs.LG]

    Pith/arXiv arXiv 2026

  20. [20]

    Cristina Pinneri, Shambhuraj Sawant, Sebastian Blaes, Jan Achterhold, Joerg Stueckler, Michal Rolinek, and Georg Martius. 2020. Sample-Efficient Cross- Entropy Method for Real-Time Planning. InProceedings of the Conference on Robot Learning

    2020

  21. [21]

    Vlad Sobal, Wancong Zhang, Kyunghyun Cho, Randall Balestriero, Tim G. J. Rudner, and Yann LeCun. 2025. Stress-Testing Offline Reward-Free Reinforce- ment Learning: A Case for Planning with Latent Dynamics Models. In7th Robot Learning Workshop: Towards Robots with Human-Level Abilities. https: //openreview.net/forum?id=jON7H6A9UU

    2025

  22. [22]

    Lucas Thil, Jesse Read, Rim Kaddah, and Guillaume Doquet. 2026. Subspace- Decomposed JEPAs: Disentangling Progression and Content in Latent World Models. arXiv:2605.31111 [cs.LG]

    Pith/arXiv arXiv 2026

  23. [23]

    Heejeong Nam, Quentin Le Lidec, Lucas Maes, Yann LeCun, and Randall Balestriero. 2026. Causal-JEPA: Learning World Models through Object-Level Latent Masking. arXiv:2602.11389 [cs.AI] https://arxiv.org/abs/2602.11389

    Pith/arXiv arXiv 2026

  24. [24]

    Tongzhou Wang, Antonio Torralba, Phillip Isola, and Amy Zhang. 2023. Opti- mal Goal-Reaching Reinforcement Learning via Quasimetric Learning. InPro- ceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202). PMLR, Honolulu, HI, USA, 36411–36430

    2023

  25. [25]

    Ying Wang, Oumayma Bounou, Yann LeCun, and Mengye Ren. 2026. AdaJEPA: An Adaptive Latent World Model. arXiv:2606.32026 [cs.LG]

    Pith/arXiv arXiv 2026

  26. [26]

    Ying Wang, Oumayma Bounou, Gaoyue Zhou, Randall Balestriero, Tim G. J. Rudner, Yann LeCun, and Mengye Ren. 2026. Temporal Straightening for Latent Planning. arXiv:2603.12231 [cs.LG]

    arXiv 2026

  27. [27]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs.LG]

    Pith/arXiv arXiv 2018

  28. [28]

    Chongyi Zheng, Ruslan Salakhutdinov, and Benjamin Eysenbach. 2024. Con- trastive Difference Predictive Coding. International Conference on Learning Representations

    2024

  29. [29]

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. 2025. DINO- WM: World Models on Pre-trained Visual Features enable Zero-shot Planning. arXiv:2411.04983 [cs.RO] https://arxiv.org/abs/2411.04983

    Pith/arXiv arXiv 2025

  30. [30]

    Silin Zhou, Yao Chen, Shuo Shang, Lisi Chen, Bingsheng He, and Ryosuke Shibasaki. 2025. Blurred Encoding for Trajectory Representation Learning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, New York, NY, USA, 4132–4143. doi:10.1145/3711896. 3736861 A Detailed Plan-Cost Results Table 11 lists the exact values...

    doi:10.1145/3711896 2025

  31. [31]

    Wancong Zhang, Basile Terver, Artem Zholus, Soham Chitnis, Harsh Su- taria, Mido Assran, Randall Balestriero, Amir Bar, Adrien Bardes, Yann LeCun, and Nicolas Ballas. 2026. Hierarchical Planning with Latent World Models. arXiv:2604.03208 [cs.LG]

    Pith/arXiv arXiv 2026

  32. [2005]

    A Tutorial on the Cross-Entropy Method.Annals of Operations Research 134 (2005), 19–67

    2005

  33. [2025]

    Advances in Neural Information Processing Systems

    Offline Goal-Conditioned Reinforcement Learning with Quasimetric Rep- resentations. Advances in Neural Information Processing Systems

  34. [2026]

    https://drfeifei.substack.com/p/a-functional-taxonomy-of-world-models