Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Diffusion world models run up to 4.92x faster by caching whole latent transitions under a risk gate.

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 →

T0 review · deepseek-v4-flash

2026-08-04 19:44 UTC pith:XCWBSX7I

load-bearing objection WorldDynCache is a well-motivated and ablated inference acceleration method, but the Aether speedup is internally inconsistent with the stated exact-call ratio and the missing artifacts require caution. the 3 major comments →

arxiv 2608.01845 v1 pith:XCWBSX7I submitted 2026-08-03 cs.LG cs.CV

WorldDynCache: Risk-Controlled Latent Dynamics Approximation for Diffusion World Model

classification cs.LG cs.CV
keywords diffusion world modelslatent cachingrisk-controlled inferenceKoopman-inspired dynamicsvideo generation acceleration3D reconstructiontraining-free accelerationtransition approximation
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.

WorldDynCache argues that existing caching methods for diffusion world models decide whether to skip computation by looking at local changes in denoising outputs, features, or token states. The paper identifies two failure modes in that local view: small approximation defects can amplify into large downstream errors after later denoising transitions, and changing denoising phases or camera/geometry conditions can rotate the direction of latent evolution so that short native-space histories mislead. Instead, WorldDynCache approximates the full composed latent transition—from z_t to z_{t-1}—using a condition- and phase-aware lifted representation that retrieves exact-transition increments from memory, and it gates each approximation with an online-calibrated risk estimate that accumulates across a cache streak. On HunyuanVoyager-13B and Aether-5B, the paper reports 4.92x and 2.15x generation speedups respectively while achieving the best generation quality among compared caching methods across WorldScore, PSNR, SSIM, and LPIPS; on Aether 3D reconstruction, it reports a 3.42x speedup with improved depth and pose metrics. The central claim is that risk-controlled latent-transition approximation is a more reliable route to fast world-model inference than local cacheability criteria.

Core claim

The paper claims that the right unit of approximation in diffusion world models is the composed latent transition z_t → z_{t-1} = Ψ_t(z_t, F_θ(z_t, t, c)), not the denoiser output y_t = F_θ(z_t, t, c). Because an approximate latent is then evolved by remaining denoising transitions, reliability depends on the accumulated future impact of approximation defects and on phase- and condition-dependent changes in the direction of latent evolution. WorldDynCache implements this with two components: a latent dynamics surrogate that lifts the current state and condition descriptors, retrieves compatible exact-transition increments from a memory bank in the lifted space, rescales them by a scheduler-p

What carries the argument

The central object is the lifted latent-state representation φ_t = Φ(z_t, H_t) from Eq. (2), which stacks an RMS-normalized current world latent, an anchor-relative displacement, camera/motion and depth/geometry descriptors, a normalized scheduler phase q_t, and a summary of recent exact transitions. Around it, the method uses (i) trajectory-local Koopman-inspired evolution in the lifted space: the predicted next lifted state is φ_t + α_t Σ_i π_i Δφ_i, where π_i are retrieval weights from a memory bank of exact transitions and α_t rescales the increment magnitude by the current-to-anchor scheduler-phase ratio; (ii) kernelized nonlinear observation that maps the predicted lifted state back to

Load-bearing premise

The method's accuracy rests on the assumption in Eq. (4) that a trajectory-local Koopman-inspired linear evolution in the lifted space, using transition increments retrieved from similar exact transitions and rescaled by the scheduler-phase ratio α_t, approximates the true composed latent transition over a short horizon; the paper provides ablations but no theoretical guarantee, and the method depends on hyperparameters (memory capacity, compatibility thresholds, temperatures

What would settle it

Run the same WorldDynCache pipeline on HunyuanVoyager-13B or Aether-5B while replacing the lifted-space increment retrieval with increments drawn from mismatched conditions or phases (or randomly permuted), holding the risk controller fixed; if generation quality does not degrade substantially, the lifted retrieval is not the source of the reported gains. Conversely, if the risk controller is disabled while the exact-call ratio is held constant, the reported quality should drop; if quality stays flat, the risk gate contributes little. A direct check: measure the angular direction error of the

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, aggressive inference acceleration for diffusion world models is better achieved by approximating composed latent transitions than by reusing denoising outputs or token-level features.
  • Local instantaneous defect metrics are insufficient for deciding cacheability; controllers should weight defects by future sensitivity and accumulate risk across consecutive surrogate steps.
  • The framework transfers across tasks beyond video generation: the same risk-gated latent surrogate improves depth and pose metrics on Aether 3D reconstruction while delivering a 3.42x speedup.
  • The method is training-free, requires no extra transformer forward passes on accepted surrogate steps, and leaves peak memory at roughly native levels, so it can be dropped into existing world-model pipelines.
  • Because quality gains are reported at fixed exact-call ratios, a practical consequence is that the risk threshold τ_risk becomes a single knob for the latency-fidelity tradeoff.
  • None

Where Pith is reading between the lines

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

  • Editorial extension: the risk-control pattern—approximate a composed transition and gate it with accumulated future-weighted risk—could generalize to other sequential generative models such as long-horizon autoregressive video or planning models, where local cacheability may similarly misjudge downstream impact.
  • Editorial extension: the counterfactual calibration scheme suggests a generic online safety mechanism for any learned surrogate—run the surrogate at trusted anchors, measure its error against the exact result, and inflate subsequent risk estimates by that measured bias; this could be applied beyond Koopman-style surrogates.
  • Editorial extension: a testable follow-up would replace the fixed non-learned lifting map with a learned observable map trained to preserve transition directions; if direction error is the bottleneck, such a map should reduce it further and extend achievable cache streaks.
  • Editorial extension: since the paper reports fixed target exact-call ratios, sweeping the risk threshold on each backbone would produce a speedup-quality frontier; comparing these curves against WorldCache and other baselines would quantify the advantage of risk-gated placement independently of the chosen operating point.
  • None

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 WorldDynCache, a training-free caching framework for diffusion world models. Instead of reusing intermediate transformer features or denoising outputs, WorldDynCache approximates the composed latent transition z_t → z_{t-1} using a condition- and phase-aware lifted surrogate built from a memory bank of exact transitions, and a risk controller that predicts and accumulates a future-weighted defect proxy before deciding whether to commit the surrogate or fall back to exact transformer evaluation. The risk proxy is calibrated online at exact anchors through counterfactual defects. Experiments on HunyuanVoyager-13B and Aether-5B report generation speedups of 4.92× and 2.15×, respectively, with quality gains over WorldCache on WorldScore, PSNR, SSIM, and LPIPS, and a 3.42× speedup on Aether 3D reconstruction with improved depth and pose metrics. Ablations under matched budgets separate the contributions of the controller and surrogate and show consistent ordering across both backbones.

Significance. If the reported speedups and quality comparisons are reliable, the framework is practically significant for diffusion world-model inference: it is training-free, avoids extra transformer evaluations on accepted surrogate steps, and shows consistent gains across two world-model backbones and a reconstruction task. The paper's strengths include the matched-budget ablations that isolate controller and surrogate effects, the cross-model validation, and the empirical diagnostics that motivate the risk controller. The main limitations are that the method is heuristic and lacks formal guarantees for the lifted-surrogate approximation, and several load-bearing empirical numbers are not yet fully supported: the Aether speedup is inconsistent with the stated 47% exact-call ratio unless the realized ratio is lower, and all results are single runs without error bars. These issues are fixable but need to be addressed before the central claim can be fully accepted.

major comments (3)
  1. [Experimental Settings / Table 1] For Aether-5B generation the paper states a target exact-call ratio of 47% and reports a 2.15× speedup. If the exact transformer transition is the dominant per-step cost and the surrogate branch has non-negative wall-clock cost, the maximum speedup is the reciprocal of the realized exact-call ratio. For r=0.47, the bound is 1/0.47=2.128×, so the reported 2.15× is unattainable with a realized ratio of 47%. The realized ratio is reported only for the reconstruction task (14%), not for generation. Please report the realized exact-call ratios for both generation tasks, define the speedup measurement including surrogate overhead, and confirm consistency (e.g., realized r ≤ 1/speedup). A similar check should be shown for HunyuanVoyager-13B (20% target; bound 5.00× vs 4.92×). This is a load-bearing point for the headline speedup claim.
  2. [Experimental Settings / Tables 1-2] All speedup and quality numbers are single-run point estimates with no error bars or significance tests. Several gains over WorldCache are small in absolute terms (e.g., WorldScore-S 65.23 vs 64.89; LPIPS 0.158 vs 0.176 on Voyager; LPIPS 0.061 vs 0.066 on Aether). Since adaptive exact-call decisions introduce run-to-run variability, the reader cannot determine whether the reported quality and speedup differences are stable. Please report means and standard deviations over multiple independent rollouts or seeds, and include per-run exact-call ratios and speedups. This also bears on the speedup-consistency issue raised above.
  3. [Method, Eqs. (2)-(10); Experimental Settings] A large number of hyperparameters and algorithmic details are deferred to a supplementary file that is not available with the preprint: memory capacity and replacement, compatibility thresholds, temperatures (τ_in, τ), clip bounds (α_min, α_max), the fixed non-learned map Φ and summary h(H_t), the definitions of d_cam/d_depth and u_cam/u_depth/u_phase, the feature normalizations, the support constraint in Available_t, and the complete inference algorithm. These details are necessary to reproduce the method and to assess parameter sensitivity. Without them the empirical claims are under-specified. Please include the missing details in the paper or make the supplementary file available with the submission.
minor comments (5)
  1. [Abstract and general text] There are spacing/typo artifacts in the PDF text (e.g., 'criteriacanmiss', 'worldstate'). A careful proofread is needed.
  2. [Figure 2 and Figure 4 captions] The captions refer to thresholds and binning details in the supplementary. Please make the axis definitions and the meaning of 'Low/Medium/High dynamics' self-contained, or at least summarize the thresholds in the main text.
  3. [Table 1] The native row lists PSNR=∞, SSIM=1.000, LPIPS=0.000. This is expected since native is the reference, but state this explicitly so readers do not misinterpret it as an unbounded metric.
  4. [Figure 6] The axes are 'normalized within each model block among the five plotted methods'. The normalization is not defined; please specify whether normalization is min-max or z-score, and over which values.
  5. [References] Reference [Liu et al. 2025a] includes authors who are also authors of the current paper. If this is a self-citation, a disclosure statement should be added.

Circularity Check

0 steps flagged

No circular derivation; method is empirical and self-contained, with a minor numerical consistency caveat.

full rationale

WorldDynCache is an empirically engineered caching framework; it does not derive a prediction from an input by construction. The latent surrogate (Eqs. 2-5) is a fixed non-learned memory-based regressor: it retrieves compatible exact transitions and reconstructs the next latent by kernel-weighted interpolation. The risk controller (Eqs. 6-9) calibrates a scalar eta_d online against counterfactual defects, but this scalar is an internal control signal used for fallback decisions, not the paper's headline prediction; the main claims are end-to-end speedup and quality comparisons against external baselines and native models, which are not forced by the fit. The Koopman language is explicitly non-committal ('Koopman-inspired... rather than an identified or globally learned Koopman operator'), so no uniqueness or first-principles authority is imported. The only self-citation (Liu et al. 2025a includes co-authors Shaoqiu Zhang and Yulun Zhang) supports a generic cost assertion and a TeaCache baseline; it is not load-bearing. One non-circularity concern: for Aether generation, the 2.15x speedup exceeds the theoretical maximum 1/0.47 = 2.128x implied by the stated 47% target exact-call ratio if surrogate overhead is non-negative; this is an internal numerical consistency issue, not a circular reduction, and it could be resolved by reporting the realized exact-call ratio for generation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The framework's behavior depends on many hand-set weights, thresholds, and memory parameters, most of which are only defined in the supplementary material. The target exact-call ratios (20% for Voyager, 47% for Aether) are chosen to achieve the reported speedups. The core modeling assumption is that latent dynamics in a lifted space can be locally approximated by retrieved exact-transition increments rescaled by scheduler phase.

free parameters (7)
  • eta_d (defect scale) = updated online via EMA (Eq. 8)
    Scales the predicted defect proxy to match observed counterfactual defects at exact anchors.
  • omega_z, omega_age, omega_phase, omega_cam, omega_depth = fixed nonnegative weights (supplementary)
    Weights in the defect proxy pt (Eq. 6).
  • alpha_cam, alpha_depth, alpha_phase = fixed weights (supplementary)
    Future-sensitivity weights in wt (Eq. 9).
  • lambda (risk decay) = scheduled (supplementary)
    Decay for accumulated risk after exact transitions.
  • tau_risk = risk threshold (supplementary)
    Decision threshold for committing a surrogate transition (Eq. 10).
  • Memory capacity, compatibility thresholds, tau_in, tau, alpha_min/alpha_max, gamma = set in supplementary
    Hyperparameters for lifted-state retrieval and kernelized observation.
  • Target exact-call ratios = 20% for Voyager, 47% for Aether
    Chosen evaluation budgets that, together with the risk threshold, determine the measured speedup.
axioms (5)
  • standard math Diffusion world model transition is Markovian: z_{t-1} = Psi_t(z_t, F_theta(z_t,t,c))
    Assumed architecture of the world model, stated in Eq. (1).
  • domain assumption Local linearity in lifted space: a trajectory-local Koopman-inspired linear evolution can approximate latent dynamics
    Motivates the surrogate in Eq. (4); standard in Koopman-based methods but not proven for diffusion latents.
  • domain assumption Scheduler-phase scaling alpha_t linearly rescales transition increments
    Assumed in Eq. (4) to adapt retrieved increments to the current scheduler interval.
  • ad hoc to paper The defect proxy pt is a linear combination of distance, age, phase, camera, and depth signals
    Defined in Eq. (6) without derivation; the weights are hand-set.
  • ad hoc to paper The future-sensitivity weight wt is a capped linear function of camera/depth/phase features
    Defined in Eq. (9) to approximate downstream impact; no external validation beyond the paper's own correlations.

pith-pipeline@v1.3.0-daily-deepseek · 12788 in / 10934 out tokens · 107605 ms · 2026-08-04T19:44:23.744878+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of WorldDynCache: Risk-Controlled Latent Dynamics Approximation for Diffusion World Model." pith.science (2026). https://pith.science/paper/XCWBSX7I

@misc{pith2026260801845,
  author       = {Pith},
  title        = {Pith review of: WorldDynCache: Risk-Controlled Latent Dynamics Approximation for Diffusion World Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XCWBSX7I}},
  note         = {Machine review of arXiv:2608.01845}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion world models generate high-quality futures, but re- peated transformer evaluations make inference prohibitively slow. Existing caches reuse intermediate features, selectively update tokens, or reuse and extrapolate denoising outputs ac- cording to local drift or short native-space histories. These criteria can miss both approximation-induced latent transition defects that accumulate across skipped steps and phase- or condition-dependent changes in the direction of latent evo- lution. We propose WorldDynCache, a risk-controlled latent dynamics approximation framework with two core compo- nents. First, a lightweight latent-transition risk estimator tracks the accumulated future impact of approximation defects and calibrates its predictions against counterfactual defects ob- served at exact anchors. Second, a condition- and phase- aware lifted latent surrogate approximates latent evolution without extra transformer evaluations. On HunyuanVoyager- 13B and Aether-5B, WorldDynCache achieves 4.92 times and 2.15 times speedups, respectively, while attaining the best gen- eration quality among the compared caching methods across WorldScore, PSNR, SSIM, and LPIPS.

Figures

Figures reproduced from arXiv: 2608.01845 by Junyi Wu, Leyang Chen, Shaoqiu Zhang, Yulun Zhang.

Figure 1
Figure 1. Figure 1: Visualization comparison on two representative [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Latent-transition safety diagnostics for World [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of WorldDynCache. The latent dynamics [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Latent-transition risk diagnostics. Panel (a) com [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative world-generation comparisons. Left: Aether-5B verdant-landscape sample 091. Right: Voyager-13B stylized [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Speed–quality tradeoffs for HunyuanVoyager-13B [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] 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

48 extracted references · 6 linked inside Pith

  1. [1]

    Akbari, A.; Zhang, C.; Akbari, A.; Zhao, L.; Chen, Y.; Chen, W.; Zhang, X.; Yuan, G.; and Wang, Y. 2026. Flash-WAM: Modality-Aware Distillation for World Action Models. arXiv preprint arXiv:2606.05254

  2. [2]

    Blattmann, A.; Dockhorn, T.; Kulal, S.; Mendelevitch, D.; Kilian, M.; Lorenz, D.; Levi, Y.; English, Z.; Voleti, V.; Letts, A.; Jampani, V.; and Rombach, R. 2023. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets. arXiv preprint arXiv:2311.15127

  3. [3]

    L.; Brunton, B

    Brunton, S. L.; Brunton, B. W.; Proctor, J. L.; and Kutz, J. N. 2016. Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control. PLOS ONE

  4. [4]

    Duan, H.; Yu, H.-X.; Chen, S.; Fei-Fei, L.; and Wu, J. 2025. WorldScore: A Unified Evaluation Benchmark for World Generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision

  5. [5]

    Feng, L.; Zheng, S.; Liu, J.; Lin, Y.; Zhou, Q.; Cai, P.; Wang, X.; Chen, J.; Zou, C.; Ma, Y.; and Zhang, L. 2026 a . HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration. In International Conference on Learning Representations

  6. [6]

    Feng, W.; Fan, G.; Qin, H.; Wu, M.; Li, Y.; Li, X.; An, Z.; Huang, L.; Wang, D.; Liao, L.; Magno, M.; Xu, Y.; and Yang, C. 2026 b . WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching. In International Conference on Machine Learning

  7. [7]

    Ha, D.; and Schmidhuber, J. 2018. World Models. arXiv preprint arXiv:1803.10122

  8. [8]

    Hafner, D.; Lillicrap, T.; Ba, J.; and Norouzi, M. 2020. Dream to Control: Learning Behaviors by Latent Imagination. In International Conference on Learning Representations

  9. [9]

    Hansen, N.; Su, H.; and Wang, X. 2024. TD-MPC2 : Scalable, Robust World Models for Continuous Control. In International Conference on Learning Representations

  10. [10]

    He, H.; Xu, Y.; Guo, Y.; Wetzstein, G.; Dai, B.; Li, H.; and Yang, C. 2025. CameraCtrl : Enabling Camera Control for Video Diffusion Models. In International Conference on Learning Representations

  11. [11]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems

  12. [12]

    Ho, J.; Salimans, T.; Gritsenko, A.; Chan, W.; Norouzi, M.; and Fleet, D. J. 2022. Video Diffusion Models. In Advances in Neural Information Processing Systems

  13. [13]

    Huang, T.; Zheng, W.; Wang, T.; Liu, Y.; Wang, Z.; Wu, J.; Jiang, J.; Li, H.; Lau, R. W. H.; Zuo, W.; and Guo, C. 2025. Voyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation. ACM Transactions on Graphics

  14. [14]

    Karras, T.; Aittala, M.; Aila, T.; and Laine, S. 2022. Elucidating the Design Space of Diffusion-Based Generative Models. In Advances in Neural Information Processing Systems

  15. [15]

    Koopman, B. O. 1931. Hamiltonian Systems and Transformation in Hilbert Space. Proceedings of the National Academy of Sciences

  16. [16]

    Li, Y.; and van der Schaar, M. 2024. On Error Propagation of Diffusion Models. In International Conference on Learning Representations

  17. [17]

    Liu, F.; Zhang, S.; Wang, X.; Wei, Y.; Qiu, H.; Zhao, Y.; Zhang, Y.; Ye, Q.; and Wan, F. 2025 a . Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  18. [18]

    Liu, J.; Zou, C.; Lyu, Y.; Chen, J.; and Zhang, L. 2025 b . From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers. In Proceedings of the IEEE/CVF International Conference on Computer Vision

  19. [19]

    Lu, C.; Zhou, Y.; Bao, F.; Chen, J.; Li, C.; and Zhu, J. 2022. DPM-Solver : A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. In Advances in Neural Information Processing Systems

  20. [20]

    N.; and Brunton, S

    Lusch, B.; Kutz, J. N.; and Brunton, S. L. 2018. Deep Learning for Universal Linear Embeddings of Nonlinear Dynamics. Nature Communications

  21. [21]

    Ma, X.; Fang, G.; and Wang, X. 2024. DeepCache: Accelerating Diffusion Models for Free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  22. [22]

    Peebles, W.; and Xie, S. 2023. Scalable Diffusion Models with Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision

  23. [23]

    Ren, X.; Shen, T.; Huang, J.; Ling, H.; Lu, Y.; Nimier-David, M.; M \"u ller, T.; Keller, A.; Fidler, S.; and Gao, J. 2025. GEN3C : 3D -Informed World-Consistent Video Generation with Precise Camera Control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  24. [24]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  25. [25]

    Salimans, T.; and Ho, J. 2022. Progressive Distillation for Fast Sampling of Diffusion Models. In International Conference on Learning Representations

  26. [26]

    Song, J.; Meng, C.; and Ermon, S. 2021. Denoising Diffusion Implicit Models. In International Conference on Learning Representations

  27. [27]

    Song, Q.; Wang, X.; Zhou, D.; Lin, J.; Chen, C.; and Ma, Y. 2025. HERO : Hierarchical Extrapolation and Refresh for Efficient World Models. arXiv preprint arXiv:2508.17588

  28. [28]

    Team Wan ; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen, D.; et al. 2025. Wan: Open and Advanced Large-Scale Video Generative Models. arXiv preprint arXiv:2503.20314

  29. [29]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing

  30. [30]

    O.; Kevrekidis, I

    Williams, M. O.; Kevrekidis, I. G.; and Rowley, C. W. 2015. A Data-Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition. Journal of Nonlinear Science

  31. [31]

    Wimbauer, F.; Wu, B.; Schoenfeld, E.; Dai, X.; Hou, J.; He, Z.; Sanakoyeu, A.; Zhang, P.; Tsai, S.; Kohler, J.; Rupprecht, C.; Cremers, D.; Vajda, P.; and Wang, J. 2024. Cache Me if You Can: Accelerating Diffusion Models through Block Caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  32. [32]

    Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; Yin, D.; Zhang, Y.; Wang, W.; Cheng, Y.; Xu, B.; Gu, X.; Dong, Y.; and Tang, J. 2025. CogVideoX : Text-to-Video Diffusion Models with An Expert Transformer. In International Conference on Learning Representations

  33. [33]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  34. [34]

    Zhou, X.; Liang, D.; Chen, K.; Feng, T.; Chen, X.; Lin, H.; Ding, Y.; Tan, F.; Zhao, H.; and Bai, X. 2025. Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching. arXiv preprint arXiv:2507.02860

  35. [35]

    Zhu, H.; Wang, Y.; Zhou, J.; Chang, W.; Zhou, Y.; Li, Z.; Chen, J.; Shen, C.; Pang, J.; and He, T. 2025. Aether: Geometric-Aware Unified World Modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision

  36. [36]

    Zou, C.; Liu, X.; Liu, T.; Huang, S.; and Zhang, L. 2025. Accelerating Diffusion Transformers with Token-wise Feature Caching. In International Conference on Learning Representations

  37. [37]

    Zou, C.; Zhang, E.; Guo, R.; Xu, H.; He, C.; Hu, X.; and Zhang, L. 2024. Rethinking Token-wise Feature Caching: Accelerating Diffusion Transformers with Dual Feature Caching. arXiv preprint arXiv:2412.18911

  38. [38]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  39. [39]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  40. [40]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

  41. [41]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  42. [42]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

  43. [43]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  44. [44]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  45. [45]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  46. [46]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  47. [47]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

  48. [48]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet