Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Branch-JEPA replaces a world model's single predicted future with a weighted set of K latent futures, and full-set Energy-Score training makes that set a genuine predictive distribution while preserving distinct alternatives.

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

2026-08-04 04:28 UTC pith:UAB5PKTY

load-bearing objection Careful empirical paper, real but narrow contribution; the 'preserves more distinct futures' claim holds only at K=6 and is undermined by the paper's own stress test. the 3 major comments →

arxiv 2607.05238 v3 pith:UAB5PKTY submitted 2026-07-06 cs.AI

Branch-JEPA: Finite-Support Predictive Distributions for JEPA World Models

Zhi Song , Ximing Xing , Zhenchao Tang , hanbo Huang , Jiehui Huang , Weilong Yan , Tianxu Lv , Minghao Yang
show 4 more authors
Zhongzheng Niu Bing He Lusheng Wang Jianhua Yao
This is my paper
classification cs.AI
keywords joint-embedding predictive architectureJEPA world modelsfinite-support predictive distributionsEnergy Scorelatent branchingmultimodal trajectory predictioneffective supportArgoverse 2
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.

This paper argues that JEPA world models lose information by returning one latent successor, even when several futures are plausible. It introduces Branch-JEPA, which instead predicts a context-weighted finite set of K latent successors, decodes each one independently, and keeps the complete set at deployment. The paper shows that training this set with a full-set Energy Score gives a principled population target: the best K-atom approximation to the true conditional distribution. On official Argoverse 2 validation, full-set Branch-JEPA improves trajectory Energy Score and probability-weighted distance over matched-K assignment and transport objectives, and retains about 5.36 effective branches. It also shows that placing multiplicity in the latent transition beats adding it only at the output decoder, retaining 10.3% more effective modes in a parameter-exact comparison.

Core claim

Branch-JEPA replaces the point-valued transition of a JEPA world model with a context-weighted K-atomic latent measure: one encoded context produces K separate latent successors and K context-only weights; a shared decoder turns each successor into a full trajectory, and the complete weighted set is returned at inference. The paper establishes two things. First, full-set Energy-Score training has a well-defined population target: over K-atomic measures, the minimizer is the energy-discrepancy-optimal weighted atomic approximation to the true conditional future law. Second, this architecture beats matched baselines on official validation: full-set training improves trajectory Energy Score by

What carries the argument

The central object is the K-atomic weighted latent measure Q(c)=Σ_k π_k δ_{u_k}, produced by K separate predictor heads and a context-only router, with a shared decoder applied independently to each atom so branch identity survives to deployment. The argumentative engine is the Energy Score identity E_Y ES_D(Q,Y)=½ ED(P,Q)+constant, which shows that full-set training on the empirical Energy Score seeks the energy-discrepancy-optimal weighted atomic approximation to the true conditional law, making the objective both principled and proper rather than an ad hoc diversity penalty.

Load-bearing premise

The load-bearing premise is that the fixed number of branches K is large enough to cover roughly the number of futures that actually occur; if real conditional distributions contain many separated or continuous futures, the K-atom output is a lossy approximation and the demonstrated advantage over baselines shrinks, as the paper's own stress test near k=4–5 suggests.

What would settle it

A direct test: construct or identify a benchmark where the ground-truth distribution has well-separated future modes numbering well above K (for example, 12 separated trajectories with K=6). If Branch-JEPA's retained effective support stays near 5 and its Energy Score advantage over output-only branching disappears or reverses, the central claim about latent branching preserving distinct futures would be falsified.

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

If this is right

  • JEPA world models can expose an enumerable, finite-support predictive distribution while still predicting in representation space, so downstream planners can consume a set of candidate futures rather than a single point.
  • The Energy-Score identity gives a general recipe: any point-valued JEPA transition can be made distributional by scoring a K-atomic measure with a proper multivariate score, with a clear population interpretation.
  • At matched output cardinality and parameter count, placing multiplicity in the latent transition yields more effective modes and better distribution quality than adding branches only after a singleton latent, suggesting the representation itself should carry the ambiguity.
  • Full-set scoring improves not only dispersion but also probability-weighted observation fidelity, meaning the learned masses assign meaningful probability to the realized future, as shown by improved Brier and expected ADE in all five paired seeds.
  • The architecture retains roughly five distinct drivable-area-contained alternatives on Argoverse 2 validation, indicating that the multi-atom output is not a set of duplicated or map-invalid trajectories.

Where Pith is reading between the lines

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

  • If this result holds, the point-valued support bottleneck is a structural limitation of standard JEPA predictors, not a tuning issue; any downstream task that needs multiple hypotheses—such as contingency planning or scenario forecasting—could benefit from a finite-set latent transition.
  • The Energy-Score population characterization suggests a testable extension: apply the same K-atomic full-set objective to other latent prediction spaces (e.g., video or robotic state) and measure whether the effective support tracks the true number of separable futures.
  • The paper's own branch-factor stress test hints that a fixed K budget has a cliff near the true support size; a natural next step is to make K context-dependent, allocating more branches when the router entropy or scene ambiguity is high.
  • Because uniform-mass Branch-JEPA comes within about 1.25% of learned-mass full-set in trajectory Energy Score, most of the gain comes from learning the support locations rather than the weights; this suggests that improving how well the K atoms cover the true support is the highest-leverage direction.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Branch-JEPA, a joint-embedding predictive architecture that replaces a single latent successor with a context-weighted finite set of K latent successors, decoded independently and retained at inference. Two training regimes are studied: specialization (hard assignment to recover separated successors) and full-set Energy-Score training (proper scoring of the entire weighted atomic distribution). The central empirical claims are that on the Argoverse 2 official validation split, full-set training improves trajectory Energy Score and probability-weighted distance over matched-K assignment and transport baselines, and that a parameter-exact comparison shows latent branching improves Energy Score, expected ADE, Brier, and effective support over branching only at the output decoder, with all paired 95% intervals excluding zero. Additional OGBench, AntMaze, and RGB diagnostics are used to argue that latent branching preserves more distinct futures across representations. The paper also provides a population characterization (Eq. 8) of the full-set objective as an energy-discrepancy-optimal K-atomic approximation to the true conditional law.

Significance. If the claims hold, the paper makes a useful contribution to JEPA-based world modeling: it identifies and addresses a structural limitation (point-valued support) and provides a principled finite-support training objective. The evaluation protocol is a strength: selection is confined to a train-derived development split, the official validation set is evaluated once with aggregate-only statistics, per-seed results are reported without dropped seeds, and paired confidence intervals are computed from seed-level data. The population result in Eq. 8 is proved in the supplement, and the code archive is described as containing preprocessing scripts, argument hashes, and training manifests. These are significant methodological positives. The main risk is that the paper's broad generalization 'latent branching preserves more distinct futures' goes beyond what its own stress-test evidence supports.

major comments (3)
  1. [§S10, Fig. S4(c), Table S10(b)] The generalized claim 'latent branching preserves more distinct futures' is undermined by the paper's own branch-factor stress test. At K=8, held-out verified-route existence for true successor counts k=4 and k=5 is 13.4±13.5% and 23.3±14.3%, respectively — statistically indistinguishable from zero. The large-map entries in Table S10 are similarly unresolved (Branch-JEPA 13.9±16.7% verified route vs MDN 5.7±3.0%). These are the direct tests of support recovery, and they fail precisely where the claim needs support. The scoped AV2 K=6 result is unaffected, but the abstract and conclusion should be narrowed to the regimes actually tested, or the method must be shown to handle k≥4 (e.g., with larger K or a coverage-promoting objective).
  2. [Table 2 and 'Latent Branching versus Output-Only Branching'] The headline '10.3% more effective modes' uses Neff = expH(pi), a router-weight entropy that does not penalize duplicate trajectories. The paper elsewhere (Section 'Effective support', Table 1) uses endpoint-deduplicated support for exactly this reason; a collision-adjusted count is also defined but not reported for this architecture comparison. Since output-only branching may produce near-duplicate residual heads, the claimed effective-mode advantage could reflect weight concentration rather than distinct futures. Please report the deduplicated/collision-adjusted effective support for both arms, or justify why expH(pi) is sufficient for this load-bearing comparison.
  3. [Eq. (7) vs Table 1] The main matched-K comparison evaluates full-set training on trajectory Energy Score, which is the same objective being optimized in Eq. (7). The 5.8–6.5% ES gain therefore partly reflects fit-to-objective. Expected ADE and Brier provide some independent evidence, but the abstract foregrounds ES. Please state this circularity explicitly, frame the ES rows as an optimization-efficiency result, and add at least one hold-out proper score not used in the training loss (e.g., a variogram or quantile score), or move the Brier/expected-ADE evidence to the front of the claims.
minor comments (5)
  1. [Abstract vs Table 1] The abstract states 'retaining 5.36 endpoint-deduplicated effective branches' while Table 1 reports 5.364±.023; please reconcile notation and make clear that this is the endpoint-deduplicated Neff.
  2. [§'Latent Branching versus Output-Only Branching'] When reporting the parameter-exact comparison, specify that 'effective modes' refers to expH(pi), not the endpoint-deduplicated support used in Table 1, to avoid ambiguity for the reader.
  3. [§S10] The three-seed execution control described in S10 as inconclusive (86.7±11.5 vs 66.7±11.5) should be acknowledged in the main text where the OGBench results are summarized, since it bears on the practical significance of the support gains.
  4. [Table S10(a)] The 'large' row for Branch-JEPA has a very large standard deviation (38.6±50.0). Consider reporting per-seed values or a robust summary such as the median, since the mean is not informative in this regime.
  5. [Eq. (8) and S4] The noncompact Euclidean setting requires a coercivity/compactness condition for existence of an optimal atomic approximation; this is correctly noted but appears late. A one-sentence caution in the main text near Eq. (8) would help.

Circularity Check

0 steps flagged

No significant circularity: empirical claims rest on held-out official validation with locked protocols; Eq. 8 is a standard proper-scoring identity; no load-bearing self-citation chain.

full rationale

The paper's derivation chain is self-contained rather than circular. The formal population statement (Eq. 8, S4) is a standard identity for energy scores on strong-negative-type metric spaces, proved in the supplement and attributed to Gneiting–Raftery and Lyons; it does not assume the conclusion it is used to justify. The empirical comparisons are evaluated on the 24,988-scene official Argoverse 2 validation split, which is explicitly excluded from model, recipe, and checkpoint selection, with locked seeds, matched parameter counts, and aggregate-only one-shot evaluation. The fact that full-set training is evaluated on trajectory Energy Score, the same objective it optimizes, is not circular: the official validation split is held out, so the comparison is a legitimate test of whether optimizing that objective transfers; moreover, the paper also reports Brier, ECE, expected ADE, and effective-support metrics that are not part of the full-set objective. The OGBench branch-factor stress test showing unresolved verified-route existence at k=4/5 is a limitation of finite-budget support recovery, not a circularity, and the paper acknowledges it. The only potential self-citation of note, M3-JEPA (Lei et al., 2025), is used merely to motivate a fused-output control, not as load-bearing evidence for Branch-JEPA's claims. No equation is defined in terms of the result it is said to predict, and no fitted parameter is renamed as a prediction. The central claims therefore have independent content supported by held-out external evaluation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on a standard proper-scoring-rule framework (energy score), standard JEPA EMA-target machinery, and a small set of hand-chosen constants (K, lambdas, beta, dedup thresholds). No invented physical entities. The main empirically tested-elsewhere premises are the stability and informativeness of the EMA target latent and the adequacy of the fixed branch budget K for the true support; the paper's own stress tests show the K-budget premise degrades near the boundary.

free parameters (4)
  • Branch budget K = 6 (AV2), 8 (OGBench), 4 (Figure 1 illustration)
    Hand-chosen finite support cardinality. The paper's own stress test (Fig. S4c) shows the verified-route advantage becomes unresolved at true branch factors 4-5, so K-budget sensitivity is empirical and load-bearing for the generalized claim.
  • Loss weights lambda_z, lambda_y, lambda_rec = 1, 1, 1 (Eq. 9)
    Set by hand and shared by every full-set arm. The five-seed ablation (Table S6) shows moderate insensitivity, but the values are not derived.
  • Time-decay beta in D_beta = 1 (uniform time weighting)
    Hand-chosen in Eq. 5 for the AV2 instantiation; affects the trajectory distance defining the Energy Score and all decoded-space metrics.
  • Effective-support and dedup thresholds = 1m endpoint dedup; epsilon=.10 latent dedup; 0.5m trajectory-RMS dedup; 95% containment; 3m proxy
    Protocol constants defining the reported headline values (5.36 endpoint-deduplicated effective branches; collision-adjusted N_eff 4.26). Different thresholds would change the support numbers.
axioms (5)
  • standard math Energy Score is strictly proper and Euclidean/chord metrics have strong negative type, so ED(P,Q) >= 0 with equality iff P=Q.
    Invoked in S4 to justify the population target (Eq. 8). Standard results (Gneiting-Raftery 2007; Lyons 2013), with the authors correctly noting the identification claim fails for squared Euclidean distance.
  • domain assumption The EMA target encoder provides a stable, non-collapsing training target (stop-gradient, EMA momentum .996).
    Standard JEPA practice (Eq. 1). If the EMA tower collapsed, the latent Energy Score would be uninformative; the frozen-anchor gates (S7, r_ent 12.86-13.20) are empirical support, not proof.
  • domain assumption The target latent representation retains branch-relevant information needed to decode distinct futures.
    If the EMA encoder were invariant to future variability, branch decoding could not recover true multimodality. The parameter-exact 6-to-6 vs 1-to-6 ladder (Table S5) and frozen-anchor results mitigate but do not prove the premise.
  • standard math Existence of the population minimizer over A_<=K on the decoded space requires compactness or coercivity of D_beta.
    Stated in S4; the authors flag that the identification claim would not survive replacement by squared Euclidean distance. The compact unit-sphere latent space is automatic; decoded space needs a finite-first-moment condition.
  • domain assumption The AV2 context representation (16 nearest neighbors, 48 polylines, focal-centric frame, GRU+attention encoder) is an adequate representation of the observed scene.
    Preprocessing choices in S2. All results are conditional on this encoding; no ablations vary the context encoder.

pith-pipeline@v1.3.0-alltime-deepseek · 25325 in / 25430 out tokens · 237558 ms · 2026-08-04T04:28:32.019984+00:00 · methodology

0 comments
read the original abstract

Joint-embedding predictive architectures (JEPAs) learn dynamics by predicting future observations in representation space. Yet most JEPA world models return one latent successor, even when hidden intent, partial observation, or stochastic dynamics make several futures plausible. We introduce Branch-JEPA, which replaces this point-valued transition with a context-weighted finite set of latent successors. Every branch is decoded independently, and the complete set is retained at inference. The architecture supports two complementary training regimes: specialization for recovering separated successors and full-set Energy-Score training for distributional fidelity. In a locked five-seed evaluation on the Argoverse~2 official validation split, full-set training improves trajectory Energy Score by $5.8$--$6.5\%$ and probability-weighted trajectory distance by $9.3$--$10.4\%$ over matched-$K{=}6$ assignment and transport objectives, while retaining $5.36$ endpoint-deduplicated effective branches. In a parameter-exact official-validation comparison, latent branching retains $10.3\%$ more effective modes and improves Energy Score, expected ADE, and Brier in all five paired seeds over branching only at the output decoder; every paired 95\% interval excludes zero. In an OGBench graph audit, Branch-JEPA increases teleport verified-route existence to $19.2\%$ versus $3.9\%$ for the MDN. Its raw-support advantage also persists with 29-D state and RGB observations. Together, latent branching preserves more distinct futures, while full-set scoring improves the quality of the resulting predictive distribution.

Figures

Figures reproduced from arXiv: 2607.05238 by Bing He, Hanbo Huang, Jianhua Yao, Jiehui Huang, Lusheng Wang, Minghao Yang, Tianxu Lv, Weilong Yan, Ximing Xing, Zhenchao Tang, Zhi Song, Zhongzheng Niu.

Figure 1
Figure 1. Figure 1: Argument outline. Stochastic branchings create [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Branch-JEPA architecture. The deployable path encodes the observed context once, produces K weighted latent successors, and decodes each one independently with the same decoder; candidates are never averaged. During training, the EMA target encoder represents the realized future, which supervises the branches through either specialization (Eq. 6) or full-set scoring (Eq. 9). The target path is removed at i… view at source ↗
Figure 3
Figure 3. Figure 3: The collapse and the fix on official data (OGBench [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: MoP-JEPA architecture. The unchanged backbone encodes [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Branch-JEPA adapts its finite support to contextual ambiguity on AV2. Six target-blind cases from the frozen￾anchor control are shown in a common agent-centric frame. Each adjacent pair contrasts Standard JEPA (orange, K=1) with all six raw Branch-JEPA trajectories (colored). The black dashed trajectory is the realized future, added only after scenario identifiers were locked; it was unavailable to either … view at source ↗
Figure 2
Figure 2. Figure 2: Two-step successor beliefs at high-branching cells. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Decoded predictions on OGBench pointmaze [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evidence chain before benchmark scores. Left: beyond 2-D mazes, MoP improves the appropriate domain metric (coverage for pixel/SVHN, inverse ADE for ETH/UCY); in DINO-WM, the winning head is hundreds of times closer than the mean head, showing specialization in the official world-model code. Middle: as the true number of modes grows in a controlled conformer branch sweep, dense coverage degrades fast, MDN … view at source ↗
Figure 5
Figure 5. Figure 5: planAll by arm (bars: mean; dots: seeds). The [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results across domains and branch counts. Left: domain-specific metrics beyond 2-D mazes. Middle: coverage as the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Verified (realroute) success vs. the true shortest [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Realroute success by true goal distance (3 seeds). [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Raw (light) vs. router-gated (solid) coverage on [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Raw (light) vs. router-gated (solid) coverage on [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: K-sweep on real data (teleport COV-R): coverage rises until K equals the number of modes, then plateaus, with no penalty for over-provisioning, as Prop. 3 predicts. ( [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

Review history (4 revisions) →

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

69 extracted references

  1. [1]

    CVPR , year=

    Self-supervised learning from images with a joint-embedding predictive architecture , author=. CVPR , year=

  2. [2]

    Transactions on Machine Learning Research , year=

    Revisiting Feature Prediction for Learning Visual Representations from Video , author=. Transactions on Machine Learning Research , year=

  3. [3]

    2025 , url=

    Assran, Mahmoud and Bardes, Adrien and Fan, David and Garrido, Quentin and Howes, Russell and Komeili, Mojtaba and Muckley, Matthew and Rizvi, Ammar and Roberts, Claire and Sinha, Koustuv and Zholus, Artem and Arnaud, Sergio and Gejji, Abha and Martin, Ada and Robert Hogan, Francois and Dugas, Daniel and Bojanowski, Piotr and Khalidov, Vasil and Labatut, ...

    2025

  4. [4]

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

  5. [5]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    DINO-WM: World models on pre-trained visual features enable zero-shot planning , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , url=

    2025

  6. [6]

    Transactions on Machine Learning Research , year=

    What Drives Success in Physical Planning with Joint-Embedding Predictive World Models? , author=. Transactions on Machine Learning Research , year=

  7. [7]

    2022 , note=

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

    2022

  8. [8]

    2025 , url=

    Lei, Hongyang and Cheng, Xiaolong and Qin, Qi and Wang, Dan and Huang, Huazhen and Gu, Qingqing and Wu, Yetao and Ji, Luo , booktitle=. 2025 , url=

    2025

  9. [9]

    Huang, Yongchao , booktitle=

  10. [10]

    International Conference on Machine Learning , year=

    Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture---Bridging Predictive and Generative Self-Supervised Learning , author=. International Conference on Machine Learning , year=

  11. [11]

    Radha, Santosh Kumar and Goktas, Oktay , journal=

  12. [12]

    International Conference on Learning Representations , year=

    Denoising with a Joint-Embedding Predictive Architecture , author=. International Conference on Learning Representations , year=

  13. [13]

    Lim, Jing Yu and Shah, Rushi and Ikram, Zarif and Yu, Samson and Ma, Haozhe and Leong, Tze-Yun and Liu, Dianbo , journal=

  14. [14]

    NeurIPS , year=

    Multiple choice learning: Learning to produce multiple structured outputs , author=. NeurIPS , year=

  15. [15]

    NeurIPS , year=

    Stochastic multiple choice learning for training diverse deep ensembles , author=. NeurIPS , year=

  16. [16]

    ICLR , year=

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer , author=. ICLR , year=

  17. [17]

    Journal of Machine Learning Research , volume=

    Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity , author=. Journal of Machine Learning Research , volume=

  18. [18]

    Mixture density networks , author=

  19. [19]

    Park, Seohong and Frans, Kevin and Eysenbach, Benjamin and Levine, Sergey , booktitle=

  20. [20]

    Park, Seohong and Ghosh, Dibya and Eysenbach, Benjamin and Levine, Sergey , booktitle=

  21. [21]

    ICML , year=

    Optimal goal-reaching reinforcement learning via quasimetric learning , author=. ICML , year=

  22. [22]

    NeurIPS , year=

    Contrastive learning as goal-conditioned reinforcement learning , author=. NeurIPS , year=

  23. [23]

    ICLR , year=

    Learning to reach goals via iterated supervised learning , author=. ICLR , year=

  24. [24]

    Bardes, Adrien and Ponce, Jean and LeCun, Yann , booktitle=

  25. [25]

    Least squares quantization in

    Lloyd, Stuart , journal=. Least squares quantization in

  26. [26]

    ECCV , year=

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data , author=. ECCV , year=

  27. [27]

    ICCV , pages=

    Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses , author=. ICCV , pages=

  28. [28]

    ICCV , year=

    You'll never walk alone: Modeling social behavior for multi-target tracking , author=. ICCV , year=

  29. [29]

    Computer Graphics Forum , volume=

    Crowds by Example , author=. Computer Graphics Forum , volume=. 2007 , doi=

    2007

  30. [30]

    NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning , year=

    Reading Digits in Natural Images with Unsupervised Feature Learning , author=. NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning , year=

  31. [31]

    Advances in Neural Information Processing Systems 31 , pages=

    Recurrent World Models Facilitate Policy Evolution , author=. Advances in Neural Information Processing Systems 31 , pages=. 2018 , url=

    2018

  32. [32]

    ICML , pages=

    Learning Latent Dynamics for Planning from Pixels , author=. ICML , pages=

  33. [33]

    Numerische Mathematik , volume=

    A Note on Two Problems in Connexion with Graphs , author=. Numerische Mathematik , volume=. 1959 , doi=

    1959

  34. [34]

    Annals of Operations Research , volume=

    A Tutorial on the Cross-Entropy Method , author=. Annals of Operations Research , volume=. 2005 , doi=

    2005

  35. [35]

    The Annals of Statistics , volume=

    Bootstrap Methods: Another Look at the Jackknife , author=. The Annals of Statistics , volume=. 1979 , doi=

    1979

  36. [36]

    Mastering

    Hafner, Danijar and Lillicrap, Timothy and Norouzi, Mohammad and Ba, Jimmy , booktitle=. Mastering

  37. [37]

    Nature , volume=

    Mastering Diverse Control Tasks through World Models , author=. Nature , volume=. 2025 , doi=

    2025

  38. [38]

    Diffusion for World Modeling: Visual Details Matter in

    Alonso, Eloi and Jelley, Adam and Micheli, Vincent and Kanervisto, Anssi and Storkey, Amos and Pearce, Tim and Fleuret, Fran. Diffusion for World Modeling: Visual Details Matter in. Advances in Neural Information Processing Systems , year=

  39. [39]

    ICLR , year=

    Discrete codebook world models for continuous control , author=. ICLR , year=

  40. [40]

    Proceedings of the 43rd International Conference on Machine Learning , series=

    Test-Time Graph Search for Goal-Conditioned Reinforcement Learning , author=. Proceedings of the 43rd International Conference on Machine Learning , series=. 2026 , url=

    2026

  41. [41]

    Journal of the American Statistical Association , volume=

    Strictly Proper Scoring Rules, Prediction, and Estimation , author=. Journal of the American Statistical Association , volume=. 2007 , doi=

    2007

  42. [42]

    The Annals of Probability , volume=

    Distance Covariance in Metric Spaces , author=. The Annals of Probability , volume=. 2013 , doi=

    2013

  43. [43]

    CVPR , pages=

    Overcoming Limitations of Mixture Density Networks: A Sampling and Fitting Framework for Multimodal Future Prediction , author=. CVPR , pages=

  44. [44]

    and Beijbom, Oscar and Wolff, Eric M

    Phan-Minh, Tung and Grigore, Elena Corina and Boulton, Freddy A. and Beijbom, Oscar and Wolff, Eric M. , booktitle=

  45. [45]

    NeurIPS , volume=

    Trajectory-wise Multiple Choice Learning for Dynamics Generalization in Reinforcement Learning , author=. NeurIPS , volume=

  46. [46]

    CVPR , pages=

    Divide-and-Conquer for Lane-Aware Diverse Trajectory Prediction , author=. CVPR , pages=

  47. [47]

    WACV , pages=

    Hierarchical Light Transformer Ensembles for Multimodal Trajectory Forecasting , author=. WACV , pages=

  48. [48]

    Cui, Alexander and Casas, Sergio and Sadat, Abbas and Liao, Renjie and Urtasun, Raquel , booktitle=

  49. [49]

    Chen, Yuxiao and Ivanovic, Boris and Pavone, Marco , booktitle=

  50. [50]

    2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Tree-Structured Policy Planning with Learned Behavior Models , author=. 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2023 , doi=

    2023

  51. [51]

    Artificial Intelligence , volume=

    Planning and Acting in Partially Observable Stochastic Domains , author=. Artificial Intelligence , volume=. 1998 , doi=

    1998

  52. [52]

    Mathematical Programming , volume=

    Scenario Reduction in Stochastic Programming: An Approach Using Probability Metrics , author=. Mathematical Programming , volume=. 2003 , doi=

    2003

  53. [53]

    Zhou, Zikang and Zhou, Hengjian and Hu, Haibo and Wen, Zihao and Wang, Jianping and Li, Yung-Hui and Huang, Yu-Kai , booktitle=

  54. [54]

    Proceedings of the 41st International Conference on Machine Learning , series=

    Evaluation of Trajectory Distribution Predictions with Energy Score , author=. Proceedings of the 41st International Conference on Machine Learning , series=. 2024 , url=

    2024

  55. [55]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Distributional Diffusion Models with Scoring Rules , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , url=

    2025

  56. [56]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Continuous Visual Autoregressive Generation via Score Maximization , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , url=

    2025

  57. [57]

    Wilson, Benjamin and Qi, William and Agarwal, Tanmay and Lambert, John and Singh, Jagjeet and Khandelwal, Siddhesh and Pan, Bowen and Kumar, Ratnesh and Hartnett, Andrew and Pontes, Jhony Kaesemodel and Ramanan, Deva and Carr, Peter and Hays, James , booktitle=

  58. [58]

    Cheng, Jie and Mei, Xiaodong and Liu, Ming , booktitle=

  59. [59]

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

    Query-Centric Trajectory Prediction , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  60. [60]

    Monthly Weather Review , volume=

    Verification of Forecasts Expressed in Terms of Probability , author=. Monthly Weather Review , volume=

  61. [61]

    Proceedings of the 34th International Conference on Machine Learning , series=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning , series=. 2017 , url=

    2017

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    DISCO Nets: DISsimilarity COefficient Networks , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Estimating Regression Predictive Distributions with Sample Networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  64. [64]

    Proceedings of the 25th International Conference on Artificial Intelligence and Statistics , volume=

    Multivariate Quantile Function Forecaster , author=. Proceedings of the 25th International Conference on Artificial Intelligence and Statistics , volume=

  65. [65]

    2020 , url=

    Chai, Yuning and Sapp, Benjamin and Bansal, Mayank and Anguelov, Dragomir , booktitle=. 2020 , url=

    2020

  66. [66]

    Advances in Neural Information Processing Systems , volume=

    Motion Transformer with Global Intention Localization and Local Movement Refinement , author=. Advances in Neural Information Processing Systems , volume=. 2022 , url=

    2022

  67. [67]

    Yuan, Ye and Kitani, Kris , booktitle=

  68. [68]

    2023 , doi=

    Jiang, Chiyu Max and Cornman, Andre and Park, Cheolho and Sapp, Ben and Zhou, Yin and Anguelov, Dragomir , booktitle=. 2023 , doi=

    2023

  69. [69]

    and Sapp, Benjamin , booktitle=

    Nayakanti, Nigamaa and Al-Rfou, Rami and Zhou, Aurick and Goel, Kratarth and Refaat, Khaled S. and Sapp, Benjamin , booktitle=