ViDecideTeaching Video Diffusion Models to Reason and Plan Step by Step

Kaiwen Zhang1, Yiming Gao2, Zeqi Xiao1, Xintao Wang3, Pengfei Wan2, Kun Gai2, Xingang Pan1

1S-Lab, Nanyang Technological University2KlingAI Research3CUHK

Strict success rate of each benchmark from standard fine-tuning to the full method
Standard SFT uses the default timestep shift of Wan2.2 (s = 5).

Existing video models break the physical or logical task rules, even when the final frame looks right.
ViDecide solves the same test puzzles with valid state transitions and stops at its own EOS frame (left),
raising the average strict success rate from 23.3% with standard fine-tuning to 83.8% (right).

Abstract

Current leading video generation models can synthesize highly realistic videos, while still frequently violating physical and logical constraints in multi-step reasoning and planning, even after training with large-scale reasoning data or existing diffusion RL methods. We attribute these failures to the conventional video training paradigm, which treats a solution as a fixed-length clip rather than a chain of decisions. We present ViDecide, a decision-centric framework that trains video diffusion models to predict state transitions causally and decide when to stop, with a decision-focused training schedule. We further propose Corrective Fine-tuning (CFT), which corrects failed decisions by mapping their noise to the model’s own successful continuations, achieving stable improvement without policy-gradient optimization. Across abstract games, planning tasks, and a new benchmark of multi-step robot planning, ViDecide achieves an average strict success rate of 83.8%, far above 23.3% for standard fine-tuning and 67.8% for the strongest prior method, and generalizes to larger puzzles and unseen layouts. Our framework therefore unlocks greater potential for video generation models as reliable reasoners and planners in the visual world.

Method

0:00

Results

VideoRLVR

Rule-based games, 1,000 test puzzles per game

VBVR-MultiStep

Six planning tasks, 100 test puzzles each

RoboTwin-Plan new

Dual-arm block planning, 700 test episodes

Strict success rate (SSR)

The share of test videos that

reach the goalandfollow the task rules at every state transition

A rule verifier replays every transition of the video, so a final frame that looks solved does not count by itself. On VBVR-MultiStep the solution must also be the shortest.

Design 1 · Generation paradigm

Causal generation with adaptive termination

Bidirectional generation denoises all frames of a fixed-length clip together. ViDecide generates the video causally in segments of K latent frames, frame-wise (K = 1) or chunk-wise (K = 8), and stops at its own EOS frame. With the same data and backbone, this raises the average strict success rate by 48.2 points.

Strict success rate (%) of the generation paradigms.
Causal generation is the best on every benchmark: frame-wise on the games,
chunk-wise on RoboTwin-Plan, where each state lasts longer.
MethodWan2.2-5B fine-tunedDiT-B from scratch
FlowFreeSokobanVBVR-MultiStepRoboTwin-PlanFlowFreeSokoban
Zero-shot video generators†
Seedance 2.5‡0.05.020.06.7––
Veo 3.1‡0.05.00.00.0––
Wan2.2-5B0.00.00.00.0––
VBVR-Wan2.2-A14B0.00.011.30.0––
VBVR-Pro-Wan2.2-5B-RLVR0.00.07.70.0––
Fixed-length generation
Bidirectional1.150.822.342.60.08.8
HDR11.645.924.063.90.410.3
S2PD68.175.764.562.941.256.4
Causal generation with adaptive termination
Ours (K = 8)28.575.057.873.95.818.3
Ours (K = 1)77.983.074.870.158.673.8

†Judged with a relaxed criterion that accepts non-optimal solutions and minor errors after manual review.
‡Evaluated on 20–30 test samples per benchmark due to API cost.

Videos: comparison with the prior methods HDR and S2PD
Videos: zero-shot video generators on the games

Design 2 · Training schedule

Decision-focused training schedule

The decision-focused schedule shifts training toward high noise levels, concentrating it on the decisions themselves. On RoboTwin-Plan, strict success rises from 33.3% with the default s = 5 of Wan2.2 to 73.9% with s = 10, likely because robot videos contain much more visual detail.

Strict success rate (%) for each shift s.
s = 1 applies no shift;
s = 5 is the Wan2.2 default.
Benchmarks = 1s = 5s = 10
Wan2.2-5B
FlowFree46.873.577.9
Sokoban68.479.683.0
VBVR-MultiStep44.070.374.8
RoboTwin-Plan2.633.373.9
DiT-B
FlowFree0.051.558.6
Sokoban2.070.873.8

Design 3 · Post-training

Corrective Fine-tuning

All post-training methods start from the same SFT model and use the same training budget. CFT improves every benchmark, while Flow-GRPO and DiffusionNFT rarely improve the model, whether the reward is given to the whole video or to continuations branched at a failed decision.

Test puzzles that the SFT model fails

Diffusion RL and CFT from the same SFT model

Strict success rate (%) after post-training.
CFT is the best or tied best on 9 of the 10 test sets. Hard: held-out sets of 500 harder puzzles.
MethodWan2.2-5B fine-tunedDiT-B from scratch
FlowFreeSokobanVBVR-MultiStepRoboTwin-PlanFlowFreeSokoban
TestHardTestHardTestTestTestHardTestHard
S2PD68.141.275.738.264.562.941.225.656.411.6
SFT model (ours)77.944.883.050.074.873.958.624.473.832.8
+ Continued SFT75.548.685.559.674.274.355.834.474.142.4
Reward on the whole video
+ Flow-GRPO72.947.080.645.473.273.457.731.873.037.2
+ DiffusionNFT79.543.080.545.672.772.060.226.071.628.0
+ RFT83.250.290.161.679.877.960.730.877.943.2
Reward on continuations from a failed decision
+ Flow-GRPO73.045.881.047.074.375.057.031.074.139.0
+ DiffusionNFT73.843.278.640.067.263.361.425.072.828.6
+ RFT80.152.089.873.878.870.759.028.079.146.3
+ CFT (ours)84.452.090.878.282.278.065.627.279.648.3
Test strict success rate (%) during post-training, evaluated every half epoch.
CFT improves on every benchmark within half an epoch, by more than 20 points on Sokoban-Hard.
Test strict success during post-training

Generalization

Longer plans and new layouts

Every puzzle below lies outside the training range, with more boxes, a larger board, a longer plan or a layout never seen in training. Our models still solve many of them, which suggests that they learn transferable rules rather than only mimicking the training solutions.

Strict success rate (%) on puzzles outside the training range.
CFT starts from the better causal model of each benchmark.
MethodSokobanFlowFreeRoboTwin-Plan
4 boxes5 boxes9×9Blocks World
2×2
Tower of Hanoi
2×2
Bidirectional0.02.00.014.016.0
Ours (K = 8)23.010.03.036.018.0
Ours (K = 1)51.037.032.04.02.0
+ CFT (ours)74.054.038.032.014.0
Strict success rate (%) by puzzle size.
Frame-wise generation with CFT stays on top at every size.
Strict success by puzzle size

Analysis

Generation paradigm

Where joint generation fails

Each test video is categorized by its first error. Most failures of joint generation break a rule, and frame-wise causal generation removes most of them.

Failure analysis of generation paradigms

97% → 2%

FlowFree videos that break a rule,
bidirectional → frame-wise

98.2% vs 50.8%

bidirectional Sokoban videos,
reaching the goal vs. following every rule

Post-training

Why CFT works

At 149 points where the SFT model failed on the Sokoban test sets,
CFT shifts the decision distribution toward successful decisions,
and the noise that caused the failure now often produces the successful decision.

Decisions at the divergence points

44% → 68%

successful decisions over fresh noises,
SFT model → CFT

57%

noises that caused a failure,
now producing the successful decision

Self-generated corrections also work far better than corrections drawn by a solver, which cover more puzzles but lie outside the model’s own distribution.

84.4% vs 48.2%

FlowFree strict success,
self-generated vs. solver corrections

2.8×

as many FlowFree puzzles covered by solver corrections

Self-generated versus solver corrections

More results

Strict and end-state success of the same videos
Checking every transition separates the paradigms.
Success (%) of the same videos when only the final frame is checked and when every transition is checked;
on VBVR-MultiStep, strict success and strict success with the shortest solution.
Pass@k after post-training
Pass@k after post-training (%).
A puzzle counts as solved if one of k samples is strictly successful. CFT is the best or tied best in 15 of the 18 settings.
MethodFlowFreeFlowFree-HardSokobanSokoban-Hard
k=1k=2k=4k=1k=2k=4k=1k=2k=4k=1k=2k=4
SFT model (ours)77.990.896.144.859.472.283.088.792.050.064.273.8
+ Continued SFT75.589.694.648.660.473.485.590.393.259.673.082.2
+ Sample-level RFT83.292.996.650.263.076.690.193.194.861.675.283.4
+ Decision-level RFT80.191.796.652.067.280.089.894.096.273.883.488.0
+ CFT (ours)84.493.297.452.068.679.890.894.596.878.285.088.6
MethodVBVR-MultiStepRoboTwin-Plan
k=1k=2k=4k=1k=2k=4
SFT model (ours)74.882.787.873.984.389.7
+ Continued SFT74.281.587.274.383.691.1
+ Sample-level RFT79.886.090.577.986.390.9
+ Decision-level RFT78.886.890.770.780.987.6
+ CFT (ours)82.287.591.378.085.990.7
Multi-round CFT
A second round of CFT on corrections of the updated model adds 1.6 to 5.6 points of strict success (%),
while the same number of extra updates on the round-1 data does not.
First errors on the larger puzzles
Each video categorized by its first error (%, 100 puzzles each).
On Sokoban with 4 and 5 boxes, CFT removes most of the rule violations of the frame-wise model.
Failure analysis on the larger puzzles
Two divergence points in detail
At these divergence points the failed rollout reaches a dead end (a) or hallucinates (b).
Right: the decisions of each model over 32 noises.
Two divergence points
Results per task on VBVR-MultiStep and RoboTwin-Plan
VBVR-MultiStep (strict and shortest, %).
MethodTower of HanoiOrdinal Number
Sequence
Sliding PuzzleSnake RoutingSokobanTSP Reward
Collection
Bidirectional8.02.09.05.039.071.0
HDR13.027.012.01.019.072.0
S2PD21.094.052.070.064.086.0
Ours (K = 1)37.097.061.086.075.093.0
+ CFT (ours)61.098.068.090.083.093.0
RoboTwin-Plan (strict success, %).
MethodBlocks WorldTower of HanoiColor Sorting
RowTriangleRowTriangleBufferNo buffer
Bidirectional66.030.098.066.055.05.0
HDR86.056.790.072.075.035.0
S2PD79.046.392.084.080.054.0
Ours (K = 8)85.064.098.074.090.064.0
Ours (K = 1)75.050.710088.096.074.0
Random test puzzles, 12 per benchmark with every outcome

Citation

@article{zhang2026videcide,
  title   = {ViDecide: Teaching Video Diffusion Models to Reason and Plan Step by Step},
  author  = {Zhang, Kaiwen and Gao, Yiming and Xiao, Zeqi and Wang, Xintao and Wan, Pengfei and Gai, Kun and Pan, Xingang},
  journal = {arXiv preprint arXiv:2611.26389},
  year    = {2026}
}