pith. machine review for the scientific record. sign in

arxiv: 2605.14278 · v1 · submitted 2026-05-14 · 💻 cs.CV

Recognition: unknown

KVPO: ODE-Native GRPO for Autoregressive Video Alignment via KV Semantic Exploration

Ruicheng Zhang , Kaixi Cong , Jun Zhou , Zhizhou Zhong , Zunnan Xu , Shuiyang Mao , Wei Liu , Xiu Li
Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:35 UTC · model grok-4.3

classification 💻 cs.CV
keywords KVPOautoregressive video generationKV cache explorationODE dynamicspolicy optimizationvideo alignmentflow matchingGRPO
0
0 comments X

The pith

KVPO aligns streaming autoregressive video generators by routing semantic variations through the KV cache and modeling policies via velocity energy in ODE space.

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

The paper proposes KVPO as an online Group Relative Policy Optimization framework for aligning autoregressive video generators with human preferences. It shifts the source of exploration from stochastic noise to stochastic routing of entries in the historical key-value cache, producing semantically diverse generation branches that stay on the model's learned data manifold. For the policy update, KVPO defines a velocity-field surrogate based on Trajectory Velocity Energy that produces a reward-weighted contrastive loss matching the underlying ODE dynamics of distilled models. This setup targets high-level semantic storyline progression rather than low-level appearance changes, addressing mismatches in prior noise-based and SDE-based methods. Experiments show gains in visual quality, motion quality, and text-video alignment for both short single-prompt and long multi-prompt videos.

Core claim

KVPO introduces a causal-semantic exploration paradigm that relocates the source of variation from stochastic noise to the historical KV cache, constructing semantically diverse generation branches that remain strictly on the data manifold, and a velocity-field surrogate policy based on Trajectory Velocity Energy that yields a reward-weighted contrastive objective fully consistent with the native ODE formulation.

What carries the argument

Causal-semantic exploration by stochastic routing of historical KV cache entries, paired with a Trajectory Velocity Energy (TVE) velocity-field surrogate policy that quantifies branch likelihood in flow-matching velocity space.

If this is right

  • Improved visual quality in the aligned video outputs
  • Higher motion quality across generated sequences
  • Stronger text-video alignment on both short and long horizons
  • Consistent performance gains on multiple distilled autoregressive video generators

Where Pith is reading between the lines

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

  • The KV-routing idea could transfer to aligning autoregressive models in text or audio domains where semantic coherence over long sequences matters.
  • Velocity-field surrogates may serve as a general template for ODE-native reinforcement learning in other flow-matching generative systems.
  • Combining KVPO with varied reward signals beyond human preferences could be tested directly on existing video generators.

Load-bearing premise

Stochastically routing historical KV entries produces semantically diverse generation branches that remain strictly on the data manifold without introducing artifacts or deviating from the model's learned distribution.

What would settle it

Generating multiple branches via KV routing and finding that they introduce visible artifacts or fall outside the original model's output distribution would falsify the exploration paradigm.

Figures

Figures reproduced from arXiv: 2605.14278 by Jun Zhou, Kaixi Cong, Ruicheng Zhang, Shuiyang Mao, Wei Liu, Xiu Li, Zhizhou Zhong, Zunnan Xu.

Figure 2
Figure 2. Figure 2: Overview of the KVPO training pipeline. Starting from a shared initial noise, the model first performs causal-semantic exploration via stochastic KV routing within a perturbed window to produce diverse candidate branches (a). These branches are then replayed under the unperturbed deployment-time context, where the Trajectory Velocity Energy of each branch is computed and converted into Gibbs-form surrogate… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison between LongLive and LongLive trained with KVPO. KVPO yields more faithful prompt grounding, cleaner object interactions, and smoother temporal evolution. Alignment (TA). The VQ reward is computed as the average HPSv3 score [14], while MQ and TA rewards are obtained via the official VideoAlign configuration [11]. For long-video generation, rewards are computed per segment and average… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison between MemFlow and MemFlow trained with KVPO. KVPO yields clearer semantic transitions, richer details, and stronger cross-segment consistency [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human study on long-video settings. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance improvements on LongLive. Figures 3 and 4 visually confirm these improvements. Relative to the baselines, the optimized mod￾els exhibit more accurate prompt grounding, cleaner object bound￾aries, more plausible motion conti￾8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the effectiveness of causal-semantic exploration. continuation. Second, this intervention operates on the model’s native conditioning state rather than introducing external stochastic perturbations into the latent trajectory. Third, because preferences for long-horizon video depend heavily on temporal coherence, subject consistency, and semantic progression, varying historical KV compositi… view at source ↗
Figure 8
Figure 8. Figure 8: GPU memory usage and compute utilization of KVPO during LongLive training. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative results on short-video generation. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative results on LongLive. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional qualitative results on LongLive. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional qualitative results on LongLive. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Additional qualitative results on LongLive. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional qualitative results on LongLive. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional qualitative results on MemFlow. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Additional qualitative results on MemFlow. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Additional qualitative results on MemFlow. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional qualitative results on MemFlow. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional qualitative results on MemFlow. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Additional qualitative results on MemFlow. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
read the original abstract

Aligning streaming autoregressive (AR) video generators with human preferences is challenging. Existing reinforcement learning methods predominantly rely on noise-based exploration and SDE-based surrogate policies that are mismatched to the deterministic ODE dynamics of distilled AR models, and tend to perturb low-level appearance rather than the high-level semantic storyline progression critical for long-horizon coherence. To address these limitations, we present KVPO, an ODE-native online Group Relative Policy Optimization (GRPO) framework for aligning streaming video generators. For diversity exploration, KVPO introduces a causal-semantic exploration paradigm that relocates the source of variation from stochastic noise to the historical KV cache. By stochastically routing historical KV entries, it constructs semantically diverse generation branches that remain strictly on the data manifold. For policy modeling, KVPO introduces a velocity-field surrogate policy based on Trajectory Velocity Energy (TVE), which quantifies branch likelihood in flow-matching velocity space and yields a reward-weighted contrastive objective fully consistent with the native ODE formulation. Experiments on multiple distilled AR video generators demonstrate consistent gains in visual quality, motion quality, and text-video alignment across both single-prompt short-video and multi-prompt long-video settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. KVPO is an ODE-native online Group Relative Policy Optimization (GRPO) framework for aligning streaming autoregressive video generators with human preferences. It replaces noise-based SDE exploration with causal-semantic exploration via stochastic routing of historical KV cache entries to generate semantically diverse branches that remain on the data manifold, and introduces a Trajectory Velocity Energy (TVE) velocity-field surrogate policy that produces a reward-weighted contrastive objective consistent with native ODE/flow-matching dynamics. The abstract claims consistent gains in visual quality, motion quality, and text-video alignment on distilled AR video generators for both short single-prompt and long multi-prompt settings.

Significance. If the on-manifold invariance under KV routing and the independent grounding of TVE hold, the method could provide a semantically focused alternative to noise-based exploration, improving long-horizon coherence in video alignment without low-level perturbations. The ODE-native formulation directly addresses the mismatch between existing RL methods and distilled deterministic models, which is a relevant direction for autoregressive video generation.

major comments (3)
  1. [Abstract] Abstract: the claim of 'consistent gains in visual quality, motion quality, and text-video alignment' is stated without any quantitative results, baselines, error bars, or experimental details, so the magnitude and reliability of the reported improvements cannot be assessed.
  2. [Exploration paradigm (implied §3)] The central claim that stochastically routing historical KV entries constructs branches that remain strictly on the data manifold and preserve velocity-field consistency lacks any derivation or check; without this, the TVE surrogate and reward-weighted contrastive objective rest on an unverified invariance, and the claimed advantage over noise-based SDE exploration is not established.
  3. [Policy modeling (implied §4)] TVE definition and usage: Trajectory Velocity Energy is presented as quantifying branch likelihood in flow-matching velocity space to yield a consistent objective, but without the full equations it is unclear whether TVE has independent grounding or reduces to a fitted quantity, creating potential circularity in the policy modeling.
minor comments (1)
  1. The manuscript would benefit from including the explicit mathematical definition of TVE and the KV routing procedure in the main text rather than relying on high-level description.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'consistent gains in visual quality, motion quality, and text-video alignment' is stated without any quantitative results, baselines, error bars, or experimental details, so the magnitude and reliability of the reported improvements cannot be assessed.

    Authors: We agree that the abstract would be strengthened by including specific quantitative support. In the revised manuscript we will add concise numerical results (e.g., relative improvements in the primary metrics together with references to the corresponding tables) while preserving brevity. The full experimental results, including all baselines, error bars, and statistical details, are already reported in Section 5. revision: yes

  2. Referee: [Exploration paradigm (implied §3)] The central claim that stochastically routing historical KV entries constructs branches that remain strictly on the data manifold and preserve velocity-field consistency lacks any derivation or check; without this, the TVE surrogate and reward-weighted contrastive objective rest on an unverified invariance, and the claimed advantage over noise-based SDE exploration is not established.

    Authors: We acknowledge that a formal derivation of the on-manifold invariance under KV routing is not explicitly provided. Section 3 presents the conceptual argument based on the causal structure of the KV cache, but we agree a rigorous check is needed. In the revision we will insert a new subsection containing a mathematical derivation showing that stochastic KV routing preserves both manifold membership and velocity-field consistency, together with a direct comparison to noise-based SDE perturbations. revision: yes

  3. Referee: [Policy modeling (implied §4)] TVE definition and usage: Trajectory Velocity Energy is presented as quantifying branch likelihood in flow-matching velocity space to yield a consistent objective, but without the full equations it is unclear whether TVE has independent grounding or reduces to a fitted quantity, creating potential circularity in the policy modeling.

    Authors: The definition and derivation of TVE appear in Section 4, where it is obtained directly from the flow-matching velocity field along each trajectory and is therefore independent of the reward model. To remove any ambiguity we will expand the equations in the revision, include an explicit statement of its grounding in the ODE dynamics, and add a short paragraph clarifying that the contrastive objective follows from this definition without circularity. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; claims presented as novel constructions without reduction to inputs by definition

full rationale

The provided abstract and description introduce KVPO's causal-semantic exploration (via stochastic KV routing) and TVE surrogate policy as new mechanisms that achieve on-manifold branches and ODE-consistent objectives. No equations, self-definitions, fitted parameters renamed as predictions, or self-citation chains are exhibited that would make any central result equivalent to its inputs by construction. The on-manifold guarantee and consistency are asserted as consequences of the proposed paradigm rather than tautologically defined into the inputs. The derivation chain therefore remains self-contained against external benchmarks, with no load-bearing steps reducing to prior fitted quantities or author-specific ansatzes.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The framework assumes the KV routing stays on-manifold and that TVE provides a valid surrogate without additional fitted parameters or external validation detailed in the abstract.

axioms (2)
  • domain assumption Stochastically routing historical KV entries constructs semantically diverse branches that remain strictly on the data manifold
    Central to the causal-semantic exploration paradigm described in the abstract.
  • domain assumption The velocity-field surrogate policy based on TVE is fully consistent with the native ODE formulation
    Invoked to justify the reward-weighted contrastive objective.
invented entities (1)
  • Trajectory Velocity Energy (TVE) no independent evidence
    purpose: Quantifies branch likelihood in flow-matching velocity space for the surrogate policy
    Newly introduced metric without independent evidence or external benchmarks mentioned.

pith-pipeline@v0.9.0 · 5526 in / 1455 out tokens · 30775 ms · 2026-05-15T02:35:53.482864+00:00 · methodology

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

33 extracted references · 33 canonical work pages · 11 internal anchors

  1. [1]

    arXiv preprint arXiv:2511.16955 (2025)

    He, D., Feng, G., Ge, X., Niu, Y ., Zhang, Y ., Ma, B., Song, G., Liu, Y ., Li, H.: Neighbor grpo: Contrastive ode policy optimization aligns flow models. arXiv preprint arXiv:2511.16955 (2025)

    work page arXiv 2025

  2. [2]

    arXiv preprint arXiv:2603.17461 (2026)

    He, D., Feng, G., Ge, X., Zhang, Y ., Ma, B., Song, G., Liu, Y ., Li, H.: Ar-copo: Align autore- gressive video generation with contrastive policy optimization. arXiv preprint arXiv:2603.17461 (2026)

    work page arXiv 2026

  3. [3]

    arXiv preprint arXiv:2603.21299 (2026)

    Hu, B., Qi, Z., Huang, G., Xu, Z., Zhang, R., Ye, C., Zhou, J., Li, X., Wang, J.: Identity- consistent video generation under large facial-angle variations. arXiv preprint arXiv:2603.21299 (2026)

    work page arXiv 2026

  4. [4]

    LoRA: Low-Rank Adaptation of Large Language Models

    Hu, J.E., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Chen, W.: Lora: Low- rank adaptation of large language models. ArXivabs/2106.09685(2021), https://api. semanticscholar.org/CorpusID:235458009

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

    Huang, X., Li, Z., He, G., Zhou, M., Shechtman, E.: Self forcing: Bridging the train-test gap in autoregressive video diffusion. arXiv preprint arXiv:2506.08009 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Ji, S., Chen, X., Yang, S., Tao, X., Wan, P., Zhao, H.: Memflow: Flowing adaptive memory for consistent and efficient long video narratives (2025),https://arxiv.org/abs/2512.14699

    work page arXiv 2025

  7. [7]

    Ji, Y ., Zhong, Z., Zhang, J., Yang, Q., Jin, X., Qin, Y ., Luo, W., Mao, S., Liu, W., Li, H.: Forcing-kv: Hybrid kv cache compression for efficient autoregressive video diffusion models (2026),https://arxiv.org/abs/2605.09681

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PqvMRDCJT9t

    Lipman, Y ., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PqvMRDCJT9t

    work page 2023

  9. [9]

    arXiv preprint arXiv:2602.11146 (2026)

    Liu, G., Yang, B., Zhi, Y ., Zhong, Z., Ke, L., Deng, D., Gao, H., Huang, Y ., Zhang, K., Fu, H., et al.: Beyond vlm-based rewards: Diffusion-native latent reward modeling. arXiv preprint arXiv:2602.11146 (2026)

    work page arXiv 2026

  10. [10]

    Flow-GRPO: Training Flow Matching Models via Online RL

    Liu, J., Liu, G., Liang, J., Li, Y ., Liu, J., Wang, X., Wan, P., Zhang, D., Ouyang, W.: Flow-grpo: Training flow matching models via online rl. arXiv preprint arXiv:2505.05470 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    Improving Video Generation with Human Feedback

    Liu, J., Liu, G., Liang, J., Yuan, Z., Liu, X., Zheng, M., Wu, X., Wang, Q., Xia, M., Wang, X., et al.: Improving video generation with human feedback. arXiv preprint arXiv:2501.13918 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation.arXiv preprint arXiv:2512.04678,

    Lu, Y ., Zeng, Y ., Li, H., Ouyang, H., Wang, Q., Cheng, K.L., Zhu, J., Cao, H., Zhang, Z., Zhu, X., et al.: Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation. arXiv preprint arXiv:2512.04678 (2025)

    work page arXiv 2025

  13. [13]

    arXiv preprint arXiv:2509.24200 (2025)

    Luo, J., Lin, J., Zhang, Z., Wu, B., Fang, M., Chen, L., Tang, H.: Univid: The open-source unified video model. arXiv preprint arXiv:2509.24200 (2025)

    work page arXiv 2025

  14. [14]

    Ma, Y ., Wu, X., Sun, K., Li, H.: Hpsv3: Towards wide-spectrum human preference score (2025), https://arxiv.org/abs/2508.03789

    work page arXiv 2025

  15. [15]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    work page 2023

  16. [16]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z., Wang, P., Zhu, Q., Yang, R., Xu, J., Li, M., Zhang, Y .W., Li, Y .K., Gao, Y ., Ma, D., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    In: International Confer- ence on Machine Learning (2023) 10

    Song, Y ., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: International Confer- ence on Machine Learning (2023) 10

    work page 2023

  18. [18]

    In: International Conference on Learning Representations (2021)

    Song, Y ., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021)

    work page 2021

  19. [19]

    In: Thirty-eighth Conference on Neural Information Processing Systems (2024),https://openreview.net/forum?id=pYNl76onJL

    Wang, W., Yang, Y .: Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. In: Thirty-eighth Conference on Neural Information Processing Systems (2024),https://openreview.net/forum?id=pYNl76onJL

    work page 2024

  20. [20]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, Z., Yu, Z., Zhou, Z., Zhou, J., Jin, X., Hong, F.T., Ji, X., Zhu, J., Cai, C., Tang, S., et al.: Hunyuanportrait: Implicit condition control for enhanced portrait animation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 15909–15919 (2025)

    work page 2025

  21. [21]

    DanceGRPO: Unleashing GRPO on Visual Generation

    Xue, Z., Wu, J., Gao, Y ., Kong, F., Zhu, L., Chen, M., Liu, Z., Liu, W., Guo, Q., Huang, W., et al.: Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    LongLive: Real-time Interactive Long Video Generation

    Yang, S., Huang, W., Chu, R., Xiao, Y ., Zhao, Y ., Wang, X., Li, M., Xie, E., Chen, Y ., Lu, Y ., Chen, S.H.Y .: Longlive: Real-time interactive long video generation. arXiv preprint arXiv:2509.22622 (2025)

    work page internal anchor Pith review arXiv 2025

  24. [24]

    In: Conference on Computer Vision and Pattern Recognition (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: Conference on Computer Vision and Pattern Recognition (2024)

    work page 2024

  25. [25]

    In: CVPR (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024)

    work page 2024

  26. [26]

    In: CVPR (2025)

    Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: CVPR (2025)

    work page 2025

  27. [27]

    RoboStereo: Dual-Tower 4D Embodied World Models for Unified Policy Optimization

    Zhang, R., Chen, G., Xu, Z., Liu, Z., Zhong, Z., Zhang, M., Zhou, J., Li, X.: Ro- bostereo: Dual-tower 4d embodied world models for unified policy optimization. arXiv preprint arXiv:2603.12639 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    arXiv preprint arXiv:2512.06628 (2025)

    Zhang, R., Zhang, M., Zhou, J., Guo, Z., Liu, X., Xu, Z., Zhong, Z., Yan, P., Luo, H., Li, X.: Mind-v: Hierarchical video generation for long-horizon robotic manipulation with rl-based physical alignment. arXiv preprint arXiv:2512.06628 (2025)

    work page arXiv 2025

  29. [29]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, R., Zhou, J., Xu, Z., Liu, Z., Huang, J., Zhang, M., Sun, Y ., Li, X.: Zo3t: Zero-shot 3d-aware trajectory-guided image-to-video generation via test-time training. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 12708–12716 (2026)

    work page 2026

  30. [30]

    Astrolabe: Steering forward-process reinforcement learning for distilled autoregressive video models.arXiv preprint arXiv:2603.17051,

    Zhang, S., Xue, Z., Fu, S., Huang, J., Kong, X., Ma, Y ., Huang, H., Duan, N., Rao, A.: Astrolabe: Steering forward-process reinforcement learning for distilled autoregressive video models. arXiv preprint arXiv:2603.17051 (2026)

    work page arXiv 2026

  31. [31]

    Advances in Neural Information Processing Systems36, 49842–49869 (2023)

    Zhao, W., Bai, L., Rao, Y ., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems36, 49842–49869 (2023)

    work page 2023

  32. [32]

    VBench-2.0: Advancing Video Generation Benchmark Suite for Intrinsic Faithfulness

    Zheng, D., Huang, Z., Liu, H., Zou, K., He, Y ., Zhang, F., Zhang, Y ., He, J., Zheng, W.S., Qiao, Y ., Liu, Z.: VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Manifold-aware exploration for reinforcement learning in video generation, 2026

    Zheng, M., Kong, W., Wu, Y ., Jiang, D., Ma, Y ., He, X., Lin, B., Gong, K., Zhong, Z., Bo, L., et al.: Manifold-aware exploration for reinforcement learning in video generation. arXiv preprint arXiv:2603.21872 (2026) 11 Appendix A KVPO Training Pipeline We summarize the full training procedure ofKVPO. At a high level, each iteration consists of three tig...

    work page arXiv 2026