ViDecide: Teaching Video Diffusion Models to Reason and Plan Step by Step

Kaiwen Zhang, Yiming Gao, Zeqi Xiao, Xintao Wang, Pengfei Wan, Kun Gai, Xingang PanarXiv 2026 · arXiv

TL;DR

ViDecide is a decision-centric framework that transforms video diffusion models from mere "pixel-mimics" into reliable reasoners and planners. By treating video generation as a causal chain of decisions and introducing a novel Corrective Fine-tuning (CFT) method, it boosts the strict success rate of planning tasks from a dismal 23.3% to 83.8%, enabling models to solve complex puzzles and robot manipulation tasks without breaking physical or logical rules.

What Problem Does This Paper Address?

While state-of-the-art video models can generate stunningly realistic visuals, they are notoriously bad at "thinking." When asked to solve a multi-step puzzle like Sokoban or plan a robot's movements, they often hallucinate: blocks might vanish, the "snake" in a game might grow without eating, or a robot arm might simply dissolve through an object.

The authors argue this stems from a fundamental mismatch in how we train these models. Standard video diffusion treats a solution as a fixed-length clip generated all at once. This ignores the causal order of decisions (where step B depends on step A) and focuses the model's training on low-noise levels where it's just refining pixels, rather than high-noise levels where it needs to actually "decide" what happens next.

Figure 1. Left: Seedance 2.5 [3] hallucinates physically impossible moves to reach the target arrangement, and the reasoning-tuned VBVR-Pro-RLVR [53] breaks the rules of Snake without reaching the goal. Our model solves both puzzles with valid state transitions and stops by itself at an EOS frame (red border). Right: Standard SFT is progressively improved with causal generation, the decision-focused schedule, and Corrective Fine-tuning (CFT). Our designs raise the strict success rate on all multi-step reasoning and planning benchmarks, lifting the average from 23.3% to 83.8%.

Methods

ViDecide introduces three core innovations to bridge the gap between "seeing" and "reasoning":

  1. Causal Diffusion with Adaptive Termination: Instead of generating a whole clip, the model generates one segment at a time, conditioned on the entire history. It learns a visual "End-of-Sequence" (EOS) frame (a red border) to decide when the task is complete, allowing for flexible solution lengths.
  2. Decision-Focused Training Schedule: The authors observed that reasoning happens at high noise levels (where the future is still "hidden"), while low noise is just for visual cleanup. They shifted the training distribution to spend more time at high noise levels, forcing the model to infer transitions from rules and history.
  3. Corrective Fine-tuning (CFT): This is a targeted post-training strategy. The model rolls out solutions; if it fails, the system finds the exact "divergence step" where it went wrong. It then takes the original noise that led to failure and trains the model to map that same noise to a successful continuation. This provides explicit, in-distribution supervision for critical errors without the instability of traditional policy-gradient reinforcement learning.
Figure 2. Method overview showing causal generation, the shifted training schedule, and the CFT noise-mapping process.

Datasets

The framework was evaluated across three diverse benchmarks:

  • VideoRLVR: Abstract rule-based games including Maze, FlowFree, and Sokoban.
  • VBVR-MultiStep: Six procedural planning tasks like Tower of Hanoi and Snake Routing.
  • RoboTwin-Plan: A newly introduced benchmark for multi-step robot planning using a dual-arm Aloha robot in the RoboTwin 2.0 simulator. It includes tasks like Blocks World and Color Sorting, requiring continuous arm motion and physical consistency.

Key Insights and Ideas

  • Reasoning vs. Rendering: There is a clear divide in the diffusion process. Reasoning (deciding the next state) occurs early in denoising (high noise), while rendering (adding textures) occurs late. Standard training schedules over-emphasize the latter, which is why models look good but act "dumb."
  • Noise as a Decision Space: CFT treats the initial noise as a "latent decision." By mapping failure-prone noise to successful outcomes, it effectively reshapes the model's decision distribution.
  • Causality is King: Moving from bidirectional (all-at-once) to causal (step-by-step) generation is the single biggest factor in reducing rule violations.

Key Results

  • Massive Success Rate Gains: On average, ViDecide raised the Strict Success Rate (SSR)—which requires every transition to be valid and the goal to be reached—from 23.3% to 83.8%.
  • Outperforming RL: ViDecide significantly outperformed prior diffusion RL methods (like Flow-GRPO), which often struggle with credit assignment in high-dimensional video space.
  • Generalization: The models didn't just memorize; they generalized to larger boards (e.g., 9x9 FlowFree) and unseen robot layouts, suggesting they learned the underlying rules of the "world."
  • Efficiency: Despite being autoregressive, the models are relatively fast, taking ~4-11 seconds per video with 5 denoising steps.

Limitations and Future Work

  • Task-Specific Verifiers: CFT currently relies on automated rule verifiers to identify failures and successes. While these can be built "agentically" using LLMs for games, they are harder to define for open-ended real-world tasks.
  • Specialist Models: The current experiments mostly train one specialist model per benchmark. Scaling to a single generalist "reasoning" model across many domains is a natural next step.
  • Video-to-Action Gap: While the model plans perfectly in "video space," executing these plans on a real robot requires an inverse dynamics model to translate pixels into motor commands. The authors suggest integrating ViDecide with existing world-action models to close this loop.

Summary updated

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. Project page: https://kevin-thu.github.io/ViDecide.
BibTeX
@misc{zhang2026videcide,
  author = {Kaiwen Zhang and Yiming Gao and Zeqi Xiao and Xintao Wang and Pengfei Wan and Kun Gai and Xingang Pan},
  title = {{ViDecide: Teaching Video Diffusion Models to Reason and Plan Step by Step}},
  year = {2026},
  eprint = {2611.26389},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26389},
}
Figure