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. Project page: https://kevin-thu.github.io/ViDecide.
Figures & tables
Figure 2: Method overview. (a) Instead of jointly generating a fixed-length clip, the model generates state transitions causally and ends with a visual EOS frame. (b) The decision-focused schedule moves training to high noise levels, where the next state is hidden and must be inferred from the rules and the history. (c) From the state hd before a failed decision, CFT resamples a successful continuation and trains the noise ϵd− that produced the failed decision to generate the successful decision Xd+ , shifting the decision distribution toward solvable moves.
Wan2.2-5B fine-tuned
DiT-B trained from scratch
Method
FlowFree
Sokoban
VBVR-MultiStep
RoboTwin-Plan
FlowFree
Sokoban
Zero-shot video generators
Seedance 2.5 †
0.0
5.0
20.0
6.7
–
–
Veo 3.1 †
0.0
5.0
0.0
0.0
–
–
Wan2.2-5B
0.0
0.0
0.0
0.0
–
–
VBVR-Wan2.2-A14B
0.0
0.0
11.3
0.0
–
–
Table 1: Generation paradigms (SSR, %). VBVR-MultiStep and RoboTwin-Plan report the success rate over all tasks; per-task results are in Tabs. 2 and 3 . Zero-shot generators are judged with a relaxed criterion that accepts non-optimal solutions and minor errors after manual review. † Evaluated on stratified random subsets of 20–30 test samples per benchmark due to API cost.
Figure 3: Qualitative comparison of generation paradigms on one test puzzle per benchmark. The first four columns show the same frame index t in every row, and the last column shows each model’s EOS frame (red border). Orange borders mark rule violations.
Method
Hanoi
Ordinal
Sliding
Snake
Sokoban
TSP
Fixed-length generation
Bidirectional
8.0
2.0
9.0
5.0
39.0
71.0
HDR [ 38 ]
13.0
27.0
12.0
1.0
19.0
72.0
S2PD [ 21 ]
21.0
94.0
52.0
70.0
64.0
86.0
Causal generation with adaptive termination
Ours ( K=8 )
33.0
94.0
47.0
37.0
51.0
85.0
Table 2: Per-task results on VBVR-MultiStep (SSR, %).
Blocks World
Tower of Hanoi
Color Sorting
Method
Row
Triangle
Row
Triangle
Buffer
No buffer
Fixed-length generation
Bidirectional
66.0
30.0
98.0
66.0
55.0
5.0
HDR [ 38 ]
86.0
56.7
90.0
72.0
75.0
35.0
S2PD [ 21 ]
79.0
46.3
92.0
84.0
80.0
54.0
Causal generation with adaptive termination
Table 3: Per-task results on RoboTwin-Plan (SSR, %).
Benchmark
s=1
s=5
s=10
Wan2.2-5B fine-tuned
FlowFree
46.8
73.5
77.9
Sokoban
68.4
79.6
83.0
VBVR-MultiStep
44.0
70.3
74.8
RoboTwin-Plan
2.6
33.3
73.9
DiT-B trained from scratch
Table 4: Training schedule (SSR, %). A larger shift s in Eq. ( 4 ) moves more training to high noise levels; s=1 applies no shift, and s=5 is the default of Wan2.2.
Wan2.2-5B fine-tuned
DiT-B trained from scratch
FlowFree
Sokoban
VBVR-MultiStep
RoboTwin-Plan
FlowFree
Sokoban
Method
Test
Hard
Test
Hard
Test
Test
Test
Hard
Test
Hard
S2PD [ 21 ]
68.1
41.2
75.7
38.2
64.5
62.9
41.2
25.6
56.4
11.6
SFT model (ours)
77.9
44.8
83.0
50.0
74.8
73.9
58.6
24.4
73.8
32.8
+ Continued SFT
75.5
48.6
85.5
59.6
74.2
74.3
55.8
34.4
74.1
42.4
Sample-level
Table 5: Comparison of post-training methods (SSR, %). The SFT model is our best result in Tab. 1 , and all methods with + start from it under the same training budget. S2PD is the strongest baseline in Tab. 1 .
Figure 4: Failure analysis of generation paradigms on Wan2.2-5B, with each test video categorized by its first error (%).
Figure 5: Test SSR during post-training on Wan2.2-5B (%), evaluated on the held-out test sets every half epoch. All methods start from the same SFT model at epoch 0.
Figure 6: CFT shifts the decision distribution toward successful decisions. We probe the 149 points in the Sokoban test sets where a failed rollout and a successful continuation diverge. (a) Probability of the successful decision over 32 fresh noises per point. (b) Percentage of divergence points where the noise of the failed rollout now yields the successful decision.
Figure 7: Out-of-distribution results. Our models generalize to a larger puzzle (a) and a new layout (b).
FlowFree
Sokoban
Setting
Test
Hard
Test
Hard
CFT, round 1
84.4
52.0
90.8
78.2
+ same updates, round-1 data
84.5
53.0
90.7
76.4
+ round 2
86.2
57.6
92.4
80.2
Table 6: Multi-round CFT (SSR, %) on Wan2.2-5B.
Sokoban
FlowFree
RoboTwin-Plan
Method
4 boxes
5 boxes
9 × 9
Blocks 2 × 2
Hanoi 2 × 2
Bidirectional
0.0
2.0
0.0
14.0
16.0
Ours ( K=8 )
23.0
10.0
3.0
36.0
18.0
Ours ( K=1 )
51.0
37.0
32.0
4.0
2.0
+ CFT (ours)
74.0
54.0
38.0
32.0
14.0
Table 7: Out-of-distribution generalization (SSR, %) on Wan2.2-5B. CFT starts from the better causal model of each benchmark.
Figure S1 : Test examples of the three benchmarks. For each task, we show the first frame, the middle frame, and the last frame of the reference video. Every video ends with EOS frames that repeat the final state with a red border.
Benchmark
Task
Train
Test
Setting
Frames
VideoRLVR
Maze
9,500
1,000
5–15 passages per side
77 / 105
FlowFree
9,500
1,000
5×5 – 8×8 boards, 3–5 colors
101 / 173
Sokoban
9,500
1,000
1–3 boxes, 2–51 moves
69 / 253
VBVR-MultiStep
Tower of Hanoi
5,000
100
3–5 blocks, 1–11 moves
97 / 193
Ordinal Number Sequence
5,000
100
3 numbers
89 / 177
Sliding Puzzle
5,000
100
3×3 – 5×5 boards, 7–15 moves
69 / 105
Table S1: Benchmark statistics. The setting describes the test puzzles, and the frame counts are the median and maximum over the training videos. VideoRLVR holds out 500 of its 10K training puzzles per game for validation.
Sokoban
VBVR-MultiStep
Method
End state
Strict
Strict
+ Shortest
Bidirectional
98.2
50.8
24.3
22.3
Ours ( K=8 )
84.5
75.0
65.2
57.8
Ours ( K=1 )
83.5
83.0
85.7
74.8
Table S2 : Strict and relaxed success (%) of the generation paradigms on Wan2.2-5B, scored on the same videos. “End state” only checks the final frame, and “strict” additionally checks every transition. On VBVR-MultiStep, “shortest” additionally requires the optimal number of moves where the prompt asks for it.
Wan2.2-5B
DiT-B
VideoRLVR
VBVR-MultiStep
RoboTwin-Plan
VideoRLVR
Updates
47,500
37,500
16,700
11,000
Batch
2
4
6
32
Epochs
10
5
10
37
Peak learning rate
1×10−5
1×10−5
1×10−5
1×10−4
Warmup / schedule
500 / cosine
500 / cosine
500 / cosine
500 / constant
Table S3: SFT settings. All paradigms of a benchmark share these settings, and DiT-B is trained on FlowFree and Sokoban.
Wan2.2-5B
DiT-B
FlowFree
Sokoban
VBVR-MultiStep
RoboTwin-Plan
FlowFree
Sokoban
Start model
K=1
K=1
K=1
K=8
K=1
K=1
Corrections
8,483 (2,257)
6,962 (2,747)
10,937 (3,932)
3,430 (1,258)
9,633 (3,128)
5,791 (1,802)
Updates
7,600
6,200
9,800
3,200
1,000
1,000
Batch
3
3
3
3
32
32
Peak learning rate
1×10−6
1×10−6
1×10−6
1×10−6
5×10−6
5×10−6
Table S4: Post-training settings. “Corrections” is the number of successful continuations used for post-training, and the number in parentheses counts the training puzzles, or RoboTwin-Plan episodes, from which they are collected.
Figure S2 : SSR by puzzle size on Wan2.2-5B. FlowFree is grouped by board size and Sokoban by the number of boxes. The shaded sizes are not seen in training (Sec. A ).
FlowFree
FlowFree-Hard
Sokoban
Sokoban-Hard
Method
k=1
k=2
k=4
k=1
k=2
k=4
k=1
k=2
k=4
k=1
k=2
k=4
SFT model (ours)
77.9
90.8
96.1
44.8
59.4
72.2
83.0
88.7
92.0
50.0
64.2
73.8
+ Continued SFT
75.5
89.6
94.6
48.6
60.4
73.4
85.5
90.3
93.2
59.6
73.0
82.2
+ Sample-level RFT
83.2
92.9
96.6
50.2
63.0
76.6
90.1
93.1
94.8
61.6
75.2
83.4
+ Decision-level RFT
80.1
91.7
96.6
52.0
67.2
80.0
89.8
94.0
96.2
73.8
83.4
88.0
+ CFT (ours)
84.4
93.2
97.4
52.0
68.6
79.8
90.8
94.5
96.8
78.2
85.0
88.6
Table S5: pass@ k after post-training (SSR, %) on Wan2.2-5B. A puzzle counts as solved if at least one of k samples, drawn with k fixed seeds, is strictly successful (on VBVR-MultiStep, strictly successful and shortest, as in Tab. 5 ).
Method
Hanoi
Ordinal
Sliding
Snake
Sokoban
TSP
All
SFT model (ours)
37.0
97.0
61.0
86.0
75.0
93.0
74.8
+ Continued SFT
32.0
97.0
62.0
85.0
75.0
94.0
74.2
Sample-level
+ Flow-GRPO
30.0
97.0
59.0
87.0
70.0
96.0
73.2
+ DiffusionNFT
32.0
96.0
57.0
85.0
72.0
94.0
72.7
+ RFT
54.0
97.0
69.0
87.0
79.0
93.0
79.8
Table S6 : Post-training on VBVR-MultiStep, per task (SSR, %, strict and shortest). The last column is the overall SSR of Tab. 5 .
Wan2.2-5B
DiT-B
Method
FlowFree
Sokoban
FlowFree
Sokoban
Bidirectional
1.1
50.8
0.0
8.8
HDR [ 38 ]
11.6
45.9
0.4
10.3
S2PD [ 21 ]
68.1
75.7
41.2
56.4
Ours, diffusion forcing
69.9
81.8
59.6
58.1
Ours, teacher forcing
77.9
83.0
58.6
73.8
Table S7 : Teacher forcing and diffusion forcing (SSR, %). Both causal models generate with adaptive termination and are trained with the same data, schedule, and number of updates.
FlowFree
Sokoban
VBVR-MultiStep
RoboTwin-Plan
Method
Steps
SSR
Time
SSR
Time
SSR
Time
SSR
Time
Bidirectional
2
0.0
2.5
28.5
2.5
10.8
2.3
11.7
2.2
5
1.0
3.1
41.0
3.1
23.3
2.9
52.5
2.7
10
1.5
4.1
42.5
4.0
25.8
3.8
65.8
3.5
20
4.0
6.0
44.0
6.0
25.8
5.7
69.2
5.1
HDR
5–50
9.0
188.2
34.0
184.2
27.5
180.4
67.5
140.8
Table S8 : Number of sampling steps per segment (SSR, %, and seconds per video), on stratified subsets of 120–200 test samples per benchmark. Ours is frame-wise generation, and chunk-wise generation on RoboTwin-Plan. HDR and S2PD use their default schedules, with 5–50 steps per level for HDR and 10 steps for S2PD. Time is measured on the same hardware at batch size 1.
Sampling shift
Benchmark
Training
1
5
10
FlowFree
s=1
34.5
46.0
46.5
s=5
67.0
71.5
69.0
s=10
73.5
73.0
70.5
Sokoban
s=1
55.0
57.0
59.0
s=5
74.0
73.0
74.0
Table S9 : Training shift s versus sampling shift (SSR, %, five steps), on the subsets of Tab. S8 .
Figure S3 : Two divergence points of the probe in Fig. 6 , at which the failed rollout (a) reaches a dead end and (b) hallucinates. Left: the last frame of the shared history, the successful move, and the failing move, shown in (b) two moves later, where the failure becomes visible. Right: the share of 32 fresh noises with which each model generates the successful decision, the failing decision, or another move.
Successful decision
Safe move
Method
Fresh
Failure noise
Fresh
Failure noise
SFT model (start)
43.8
1.3
81.0
63.8
+ Continued SFT
49.3
20.1
84.2
73.8
+ Sample-level RFT
55.7
36.2
87.2
83.2
+ Decision-level RFT
65.9
48.3
95.0
90.6
+ CFT (ours)
67.9
57.0
95.0
93.3
Table S10 : Decisions at the 149 divergence points (%). “Fresh” averages over 32 fresh noises per point, and “failure noise” uses the noise of the failed rollout. A move is safe if it is legal and the puzzle remains solvable after it.
Wan2.2-5B
DiT-B
Noise
Timesteps
Test
Hard
Test
Hard
Fresh
pDF(t)
80.1
52.0
59.0
28.0
Fresh
q(t)
80.8
48.6
62.1
27.0
Failed rollout
pDF(t)
83.6
53.6
63.4
25.4
Failed rollout
q(t)
84.4
52.0
65.6
27.2
Table S11 : Noise mapping and timestep distribution on FlowFree (SSR, %). Fresh noise with pDF(t) is decision-level RFT, and the noise of the failed rollout with q(t) is CFT.
P(t=1)=0
P(t=1)=0.5
P(t=1)=1
Game
Test
Hard
Test
Hard
Test
Hard
FlowFree
63.4
25.4
65.6
27.2
61.3
23.6
Sokoban
79.1
46.1
79.6
48.3
79.2
47.9
Table S12 : Corrective timestep distribution q(t) on DiT-B (SSR, %). P(t=1) is the probability of placing a corrective sample at t=1 ; the remaining samples follow pDF(t) . CFT uses 0.5 .
Figure S4 : Self-generated versus solver corrections on Wan2.2-5B (test SSR, %), evaluated every half epoch. Both start from the same SFT model at epoch 0 and train for two epochs over their own corrections.
Figure S5 : Failure analysis on the larger puzzles (%, 100 puzzles each), with each video categorized by its first error as in Fig. 4 . CFT starts from the frame-wise model.
FlowFree
Sokoban
Method
9 × 9
4 boxes
5 boxes
SFT model (ours)
32.0
51.0
37.0
+ Continued SFT
38.0
58.0
33.0
Sample-level
+ Flow-GRPO
30.0
45.0
22.0
+ DiffusionNFT
33.0
45.0
28.0
Table S13 : Post-training methods on the larger puzzles (SSR, %, 100 puzzles each).
Test set
Method
Success
Rule
Unfinished
Visual
Original
Ours ( K=8 )
73.9
7.7
7.7
10.7
+ Continued SFT
74.3
8.9
8.6
8.3
+ Sample-level RFT
77.9
7.0
3.1
11.9
+ Decision-level RFT
70.7
6.1
5.4
17.7
+ CFT (ours)
78.0
4.9
4.1
13.0
New sets
Ours ( K=8 )
26.7
38.0
14.7
20.7
Table S14 : First errors on RoboTwin-Plan (% of episodes) of the chunk-wise start model and the post-training methods of Tab. 5 . Rule: a task rule is broken; unfinished: the video ends before the goal; visual: a visual defect of the arm or the blocks. New sets: the 150 episodes of Tab. S15 .
Blocks World
Hanoi
Longer
Method
2 × 2 (50)
2 × 2 (50)
plans (50)
Bidirectional
14.0
16.0
0.0
Ours ( K=8 )
36.0
18.0
26.0
Ours ( K=1 )
4.0
2.0
22.0
+ CFT (ours)
32.0
14.0
18.0
Table S15 : RoboTwin-Plan beyond the training distribution (SSR, %). The first two sets apply trained rules to new layouts; the last one requires longer plans. CFT starts from the chunk-wise model.
FlowFree
Sokoban
Method
Test
Hard
Test
Hard
Qwen3.8-27B
68.0
42.0
36.0
40.0
Claude Opus 5.5
100.0
100.0
100.0
96.0
ViDecide (video)
82.0
66.0
90.0
82.0
Table S16 : Comparison with frontier multimodal language models (SSR, %) on 50 random puzzles of each game test set. Language models answer with moves as text, which the same verifiers replay.
Figure S6 : Comparison with HDR and S2PD on one test puzzle per benchmark. Orange borders mark rule violations.
Figure S7 : More comparisons of the generation paradigms on two tasks not included in Fig. 3 . Orange borders mark rule violations, and the last column shows the final frame of each video.
Figure S8 : Qualitative results of (a) the decision-focused schedule on RoboTwin-Plan and (b) post-training on VBVR-MultiStep. Orange borders mark rule violations, and the last column shows the final frame of each video.
Figure S9 : Long-horizon solutions. The longest strictly successful videos of our models, with frames evenly spaced from the first frame to the EOS frame. The longest expert videos last 15.8 s on Sokoban, 11.6 s on Blocks World, 10.8 s on FlowFree, and 12.1 s on Tower of Hanoi.
Benchmark
Task
Prompt
VideoRLVR
Maze
Solve the maze by drawing one continuous ⟨ color ⟩ path from the ⟨ color ⟩ start to the ⟨ color ⟩ goal without crossing a wall.
FlowFree
Fill the grid one cell at a time. Start at any colored endpoint and visit every remaining endpoint of that color. Then switch at an endpoint of a new color. Moves are horizontal or vertical, cannot cross or overlap, and every cell must be filled. Keep a single continuous walk across color changes: the next endpoint must be adjacent to the previous filled cell.
Sokoban
Solve the Sokoban puzzle. Move the agent and push every box onto a target. Boxes cannot be pulled and walls cannot be crossed.
VBVR-MultiStep
Tower of Hanoi
[Scenario] The image shows colored blocks arranged in stacks. There is a CURRENT state on the left and a TARGET state on the right. [Rules] 1. Only the topmost block of any stack can be moved at a time. 2. A move consists of lifting a block, moving it horizontally, and lowering it onto another stack or empty position. [Task] Generate a video solving the puzzle by transforming the CURRENT state to exactly match the TARGET state. Animate the blocks moving one by one, using the minimum total number of moves required.
Ordinal Number Sequence
[Scenario] The image shows a grid with a green start point, a red end point, and several yellow cells marked with sequential numbers (1, 2, 3…). An orange circular agent is positioned at the start. [Rules] 1. The agent can only move orthogonally (up, down, left, right). 2. The agent MUST visit the numbered yellow cells in strictly increasing numerical order (1, then 2, then 3, etc.). 3. The agent must reach the red end point only after all numbered cells have been visited in order. [Task] Generate a video tracking the orange agent as it visits the numbers sequentially and heads to the final goal. The animation must reflect the shortest valid path for each sequential segment.
Sliding Puzzle
[Scenario] The image displays a sliding tile puzzle grid. The tiles are numbered, and there is exactly one blank space. [Rules] 1. Only tiles orthogonally adjacent to the blank space can be moved. 2. Only slide one tile per move into the blank space. 3. The goal state requires tiles to be arranged in sequential order (top to bottom, left to right), with the blank space at the bottom-right corner. [Task] Generate a video showing the completion of this sliding puzzle. Animate the exact sequence of ⟨ n ⟩ moves required to reach the goal state without any extra moves. Keep the camera view fixed.
Table S17 : Task prompts used in training and testing. Angle brackets mark parts that change with the puzzle, e.g . , the colors of the maze theme, the number of moves of a Sliding Puzzle, or the number and shape of the TSP rewards. The two layouts of each RoboTwin-Plan task share one prompt.
Video diffusion models generate or edit temporally coherent video by iteratively denoising latent sequences conditioned on text, images, motion, or other signals.
Video diffusion models exhibit emergent reasoning capabilities like solving mazes and puzzles, yet little is understood about how they reason during generation. We take a first step towards understanding this and study the internal planning dynamics of video models using 2D maze solving as a controlled testbed. Our investigations reveal two findings. Our first finding is early plan commitment: video diffusion models commit to a high-level motion plan within the first few denoising steps, after which further denoising alters visual details but not the underlying trajectory. Our second finding is that path length, not obstacle density, is the dominant predictor of maze difficulty, with a sharp failure threshold at 12 steps. This means video models can only reason over long mazes by chaining together multiple sequential generations. To demonstrate the practical benefits of our findings, we introduce Chaining with Early Planning, or ChEaP, which only spends compute on seeds with promising early plans and chains them together to tackle complex mazes. This improves accuracy from 7% to 67% on long-horizon mazes and by 2.5x overall on hard tasks in Frozen Lake and VR-Bench across Wan2.2-14B and HunyuanVideo-1.5. Our analysis reveals that current video models possess deeper reasoning capabilities than previously recognized, which can be elicited more reliably with better inference-time scaling.
Video diffusion models have made rapid progress in perceptual realism and temporal coherence, but they remain primarily optimized for plausible generation rather than verifiable reasoning. This limitation is especially pronounced in tasks where generated videos must satisfy explicit spatial, temporal, or logical constraints. Inspired by the role of reinforcement learning with verifiable rewards (RLVR) in reasoning-oriented language models, we introduce VideoRLVR, a practical recipe for optimizing video diffusion models with rule-based feedback. VideoRLVR formulates video reasoning as the generation of verifiable visual trajectories and consists of an SDE-GRPO optimization backbone, dense decomposed rewards, and an Early-Step Focus strategy for efficient training. The Early-Step Focus strategy restricts policy optimization to the early denoising phase, reducing training latency by about 40% while preserving performance. We evaluate VideoRLVR on Maze, FlowFree, and Sokoban, three procedurally generated domains with objective success criteria. Across these tasks, VideoRLVR consistently improves over supervised fine-tuning baselines, with dense decomposed rewards proving especially important in low-success-rate settings. Our RL-optimized model also outperforms the evaluated proprietary and open-source video generation models on these verifiable reasoning benchmarks and out-of-domain benchmarks. These results suggest that verifiable RL can move video models beyond perceptual imitation toward more reliable rule-consistent visual reasoning.
Tinghui Zhu, Sheng Zhang, James Y. Huang +5
University of California, Davis · Microsoft Research · University of Southern California +1
Video generation models have demonstrated emerging zero-shot capabilities for visual reasoning, perception, and other vision tasks. However, diffusion-based video generation is inherently stochastic, while many downstream vision tasks are deterministic. Motivated by the effectiveness of self-consistency in chain-of-thought reasoning for large language models, we investigate whether self-consistency can similarly improve diffusion-based video reasoning. We first introduce a training-free test-time scaling method that samples multiple video generations and aggregates their predictions through self-consistency. Specifically, we aggregate extracted paths, locations, or masks from multiple rollouts into a consensus prediction. To reduce the inference overhead of multi-rollout generation, we read out predictions early in the denoising trajectory, which preserves consensus quality while reducing denoising steps by more than half. We further propose Rejection Fine-Tuning (RFT) to distill consensus predictions into the video generation model. The resulting model internalizes the benefit of multi-sample consensus and requires only a single generation at inference time, while substantially outperforming the original model. Experiments on three tasks, including maze solving, visual search, and referring segmentation, show that both our self-consistency inference and consensus distillation dramatically improve video-based perception and reasoning, without requiring ground-truth videos or task-specific verification. For visual search, self-consistency raises task accuracy from 48.4% for a single generation to 99.0%. The distilled model retains much of the consensus benefit with a single rollout. For 4-by-4 maze solving, consensus-based training improves the single-generation strict success rate from 72.0% to 84.0% with the same inference latency.
Zhenghao Ni, Weimin Qiu, Meng Tang
University of Toronto · University of California, Merced