Organizations: Zhejiang University · KnowledgeXLab at Shanghai AI Lab · National University of Singapore · Chinese Academy of Sciences · Tencent Youtu Lab · Nanyang Technological University · Wuhan University
Long-horizon action-conditioned video generation aims to synthesize temporally coherent videos that follow complex action instructions over extended horizons, requiring procedural ordering, persistent action execution, and scene consistency beyond conventional TI2V's short-term fidelity. Existing single-shot video generation models typically operate in an open-loop manner, leading to incomplete action execution, hallucinated motions, and temporal drift. To address this, we propose SPIRAL, a closed-loop framework that performs sequential planning and iterative reflection for action-conditioned long-horizon video generation. Specifically, SPIRAL instantiates a think-act-reflect process: a PlanAgent decomposes high-level goals into sub-actions, which condition a VideoGenerator to synthesize each segment alongside a memory context, while a CriticAgent evaluates intermediate video segments to provide corrective feedback for iterative refinement. This closed-loop design further supports self-evolution by utilizing PlanAgent-proposed actions and CriticAgent-derived rewards for GRPO-based post-training to enhance the video generator's long-horizon consistency. Moreover, we introduce ActVideoGen-Dataset for task-specific training, and establish ActVideoGen-Bench as a dedicated evaluation suite for measuring action quality and temporal coherence. Experiments across multiple TI2V backbones alongside the self-evolving strategy show consistent gains on ActVideoGen-Bench and VBench, demonstrating the effectiveness of SPIRAL.
Figures & tables
Figure 2: S P I R A L Overview. (a) Closed-Loop Framework : The PlanAgent decomposes abstract goals into atomic plans for action-conditioned video generation, while the CriticAgent evaluates action-video alignments and triggers dual-level (inner/outer) feedback for refinement (Fig. 11 ). (b) Self-Evolving via GRPO : Guided by the PlanAgent, the VideoGenerator produces video rollouts, using CriticAgent rewards for policy optimization (Fig. 12 ).
Figure 3: Overview of ActVideoGen-Dataset. (a) A structured data example comprising goal, CoT, and step-wise video-action-critic tuples. (b-f) Detailed dataset distributions across diversity, temporal scale, and action complexity.
Figure 3Figure 4Figure 5Table 6
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
SFT
DPO
Description
Optimization & Training Strategy
Base model
Qwen3-VL-8B-Thinking
Backbone vision-language model
Learning rate
1.0×10−4
5.0×10−6
Peak learning rate for AdamW
LR scheduler
Cosine
Cosine
Decay schedule type
Warmup ratio
0.1
0.1
Ratio of warmup steps
Max epochs
3.0
3.0
Number of training epochs
Appendix
Table 7: Hyperparameter settings for PlanAgent training . Parameters are differentiated by two training phases: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
Parameter
SFT
RM
Description
Optimization & Training Strategy
Base model
Qwen3-VL-8B-Instruct
Backbone vision-language model
Learning rate
1.0×10−4
5.0×10−6
Peak learning rate for AdamW
LR scheduler
Cosine
Cosine
Decay schedule type
Warmup ratio
0.1
0.1
Ratio of warmup steps
Max epochs
3.0
1.0
Number of training epochs
Appendix
Table 8: Hyperparameter settings for CriticAgent training . Parameters are differentiated by the two training phases: Supervised Fine-Tuning (SFT) and Pairwise Reward Modeling (RM).
Figure 11: Closed-Loop Think-Act-Reflect Pipeline. PlanAgent decomposes the goal and reference context into atomic action plans, VideoGenerator executes each step into video segments, and CriticAgent verifies action-video alignment. Successful steps are stored in ContextMemory, while critic feedback triggers local refinement or global replanning for closed-loop inference-time correction .
Figure 12: Training-Time Self-Evolving via GRPO. PlanAgent samples action steps, VideoGenerator produces group rollouts, and CriticAgent assigns rewards for GRPO-based policy optimization. A curriculum gradually increases task complexity from atomic actions to long-horizon procedures, enabling VideoGenerator to internalize planning and verification signals for progressive-evolution .
Figure 13: Detailed annotation example from ActVideoGen-Dataset. A procedural task is represented by a global goal and CoT planning rationale, then decomposed into temporally grounded action-video pairs. Each step contains an action instruction, an object-tool tuple, and pre-/post-conditions, providing structured supervision for long-horizon action planning and video generation.
Figure 14: Case studies of our multi-agent evaluator. The evaluator grounds step-wise judgments in action decompositions and visual evidence, providing multi-dimensional assessments of action quality. Successful executions satisfy post-conditions, while failed actions reveal implausible motions.
Table 13
Setting
Model
Inference Cost
Action Quality
Avg. Time (s)
Memory (GB)
Action Completeness
Action Smoothness
Object Interaction
Physical Fidelity
Overall Quality
Single-Shot Baseline
LongLive-2B [ 14 ]
8
28
0.273
0.646
0.776
0.808
0.626
PlanAgent (Ours)
Qwen3-VL-8B [ 24 ]
4
19
-
-
-
-
-
VideoGenerator (Per-step)
LongLive-2B [ 14 ]
8
28
-
-
-
-
-
CriticAgent (Per-step)
Qwen3-VL-8B [ 24 ]
7
22
-
-
-
-
-
SPIRAL (T=3, Retry=1)
Plan+Gen+Critic
49
69
0.721 +44.8%
0.902 +25.6%
0.860 +8.4%
0.924 +11.6%
0.852 +22.6%
Appendix
Table 11: Inference Cost and Generation Quality Analysis. We report average inference time, memory usage, and action quality. Experiments are conducted on the same NVIDIA A100 GPUs.
Figure 15: End-to-end pipeline of SPIRAL. Given a user goal, PlanAgent first decomposes the task into step-wise actions with explicit pre-/post-conditions. At each step, VideoGenerator synthesizes a video segment, while CriticAgent evaluates action-video alignment and returns judgment with feedback. Verified segments are sequentially accumulated to compose the final long-horizon video.
Figure 16: Closed-loop feedback refinement. For the soccer procedure, CriticAgent detects that the Step 2 first generation bypasses the defender but loses ball control, violating the planned post-condition. SPIRAL uses this feedback to locally refine the action instruction, regenerate a successful crossover while preserving possession, and continue to final shooting step without propagating error.
Figure 17: Procedural action video generation results. We visualize two procedural generation modes: third-person human kinematics , where actions control full-body motion (e.g., sports behaviors); and egocentric behavior generation , where actions are executed from a first-person perspective.
Figure 18: Ultra-long procedural generation results. We visualize two eight-step kitchen procedures: tomato preparation and storage, and making tomato and cucumber salad. SPIRAL maintains coherent scene context, object states, and action ordering across extended multi-step manipulation.
Figure 19: Single-shot open-loop vs. closed-loop generation on long-horizon procedural tasks. The open-loop baseline often fails to complete the full procedure, exhibiting missing actions, incorrect ordering, and implausible transitions. In contrast, SPIRAL leverages explicit planning and verification, generating complete and long-chain action sequences with coherent ordering and procedural fidelity.
Figure 20: Single-shot open-loop vs. closed-loop generation on long-horizon procedural tasks. The open-loop baseline often fails to complete the full procedure, exhibiting missing actions, incorrect ordering, and implausible transitions. In contrast, SPIRAL leverages explicit planning and verification, generating complete and long-chain action sequences with coherent ordering and procedural fidelity.
Figure 21: Effect of GRPO-based self-evolving. Compared with the model without GRPO, GRPO-based self-evolving improves action completion, temporal coherence, and physical plausibility across long-horizon procedural generation by leveraging critic-derived reward signals during post-training.
Spatially consistent long-horizon video generation aims to maintain temporal and spatial consistency along predefined camera trajectories. Existing methods mostly entangle memory modeling with video generation, leading to inconsistent content during scene revisits and diminished generative capacity when exploring novel regions, even when trained on extensive annotated data. To address these limitations, we propose a decoupled framework that separates memory conditioning from generation. Our approach significantly reduces training costs while simultaneously enhancing spatial consistency and preserving the generative capacity for novel scene exploration. Specifically, we employ a lightweight, independent memory branch to learn precise spatial consistency from historical observations. We first introduce a hybrid memory representation to capture complementary temporal and spatial cues from generated frames, then leverage a per-frame cross-attention mechanism to ensure each frame is conditioned exclusively on the most spatially relevant historical information, which is injected into the generative model to ensure spatial consistency. When generating new scenes, a camera-aware gating mechanism is proposed to mediate the interaction between memory and generation modules, enabling memory conditioning only when meaningful historical references exist. Compared with the existing method, our method is highly data-efficient, yet the experiments demonstrate that our approach achieves state-of-the-art performance in terms of both visual quality and spatial consistency. Project page can be found at https://iguoyanjun.github.io/memorize-when-needed-project-page/.
Yanjun Guo, Zhengqiang Zhang, Pengfei Wang +3
The Hong Kong Polytechnic University · OPPO Research Institute
World models, game simulators, and long-take video creation require coherent scene evolution and sustained dynamics over extended durations. Autoregressive (AR) video diffusion provides a natural framework for long-horizon generation, yet extended rollouts often become near-static or lose visual quality. We hypothesize that these failures reflect the limited guidance provided by short-video supervision on how ongoing scene dynamics develops over longer durations. This motivates us to introduce LongTake, a two-stage training pipeline built around Long-Horizon Teacher Forcing (TF) on curated real long videos. Long-Horizon TF trains the AR model to predict later frames conditioned on long ground-truth video prefixes, extending direct supervision beyond the short training horizon. This supervision is designed to help the model sustain dynamics and preserve visual quality during long-horizon generation. Our central finding is that this training stage strengthens direct initialization for distribution matching distillation (DMD) under student self-rollout, without the intermediate few-step distillation stage used in standard pipelines. Under the same five-second DMD training setup, our initialization yields substantially higher dynamic degree than short horizon TF initialization on 30-second rollouts at comparable aesthetic quality, and surpasses the evaluated baselines in both measures. Hybrid DMD further reuses this teacher to extend supervision to later frames of the self-rollout while retaining bidirectional joint supervision over the initial window. On long-horizon self-rollouts, LongTake lies on the Pareto front of dynamic degree and aesthetic quality, and Hybrid DMD attains the highest dynamic degree among evaluated methods at both 30s and 60s.
Autoregressive (AR) video diffusion models enable low-latency streaming generation by synthesizing videos chunk by chunk with cached visual context, but this chunk-wise formulation makes temporal instruction following ambiguous. A single global prompt does not specify which sub-event should be realized in each chunk, while naively switching to step-wise prompts often leads to delayed reactions, blended step semantics, and error propagation across prompt transitions. These failures are difficult to address with supervised fine-tuning or distillation alone: SFT suffers from exposure bias, while rollout-based distillation still optimizes low-level denoising or teacher-distribution matching rather than directly enforcing action ordering and prompt-transition correctness. We address these challenges with TempAct, a planner--executor reinforcement learning framework that jointly optimizes temporal decomposition and step-conditioned execution for temporally plausible AR video generation. TempAct uses an LLM planner to explore span-aware step prompts that are executable by the video model, and trains an AR diffusion executor to follow these prompts under its own generated histories. Its key mechanism is hierarchical group exploration: candidate plans form planning groups, and each plan induces an execution group of multiple continuations from a shared visual context, enabling plan-level credit assignment for long-horizon temporal outcomes and executor-level credit assignment for prompt-switch behavior. We further design hierarchical rewards that combine plan-quality and full-video temporal feedback for the planner with local transition-level step-following rewards, aesthetic regularization, and KL constraints for the executor. Experiments on Self-Forcing and LongLive show that TempAct improves temporal consistency while preserving overall visual quality.
Jing Wang, Xiangxin Zhou, Jiajun Liang +5
Shenzhen Campus of Sun Yat-Sen University · Tencent Hunyuan · Tsinghua University +2