acceptodds
Under review as a conference paper at ICLR 2027

Video Models Reason Early: Exploiting Plan Commitment for Maze Solving

Abstract

Video diffusion models exhibit reasoning capabilities beyond content creation 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. Motivated by these findings, we introduce **Ch**aining with **Ea**rly **P**lanning, or ChEaP, an inference-time strategy that only spends compute on seeds with promising early plans and chains them together to tackle complex mazes. ChEaP improves accuracy from 7% to 67% on long-horizon mazes and by overall on hard tasks across Frozen Lake and VR-Bench using Wan2.2-14B and HunyuanVideo-1.5.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video generation, reasoning, inference-time scaling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_URJxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the denoising process of video models through maze example. It finds that trajectories extracted from early predictions are already close to those in the final generated videos. Based on this observation, it proposes Early Planning Beam Search (EPBS) to screen seeds before completing generation, and Chaining with Early Planning (ChEaP) to extend solutions across multiple video segments. Experiments show improvements in sampling efficiency and task success rates.

Critical strengths
  1. The paper is well organized, with a clear connection between its empirical observations and proposed methods.

  2. The experiments are extensive, covering trajectory analysis, hyperparameter selection, efficiency, and transfer to other tasks.

  3. The proposed screening strategy is simple and requires no model retraining. The efficiency analysis also considers intermediate decoding overhead.

Critical weaknesses
  1. Section 3.2 / Figure 2: The analysis relies on extracting a single trajectory from each intermediate prediction. If an early prediction contains multiple motion branches—a phenomenon I have frequently observed in maze-solving videos generated by video models—it is unclear how the tracker selects or combines these branches.

  2. Section 5 / Figure 6: The correlation between path length and failure does not establish whether failures are due to insufficient video duration or increased planning difficulty. A direct test with longer videos on the same mazes is missing. Furthermore, ChEaP uses approximately 630 NFEs versus 400 for the main baselines and additionally uses maze-layout information, so the comparison does not fully isolate the benefit of chaining.

  3. Section 6: The irregular-maze and LIBERO experiments extend the evaluation beyond grid mazes, but still rely on explicit goals and task-specific visual scoring. How can the approach be extended to a broader range of tasks, and how would suitable verifiers be designed for those tasks?

Questions for authors
  1. How does the trajectory extractor handle multiple motion branches in early predictions, and have you checked its reliability in such cases?

  2. Does generating longer videos improve success on the same long-path mazes?

  3. How can the approach be extended to a broader range of tasks, and what would be required to design suitable verifiers?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer URJx

Dear Reviewer URJx,

Thank you for appreciating the organization of our paper alongside our extensive experiments and efficient method. Your concerns are about (1) multiple motion branches, (2) the effect of video duration on long-path mazes, and (3) generalization of verifier design. We address each below.

W.1/ Q.1: Multiple motion branches in early predictions

Early predictions sometimes contain multiple motion branches. The Section 3.2 extractor maps the centroid of moving pixels in each frame to a grid cell. With two moving regions, both contribute to the centroid, which can fall between branches.

To test whether averaging explains the convergence in Fig. 2, we added a branch-preserving measure (Appendix E.2, Table 7). We remove isolated specks smaller than 5% of a grid cell or 50 pixels, whichever is larger. In each frame, we measure the fraction of moving pixels within one cell of the final video’s motion in the corresponding frame. We then average this fraction over frames. A competing branch that moves away from the final route lowers the score in proportion to its share of the predicted motion.

On the same 747 trajectories used in Fig. 2, we obtain the following results across all four maze sizes.

Share of moving pixels within one cell of the final motionStep 1Step 3Step 5Step 8Step 10
Same seed’s final video38.4%81.4%88.6%95.9%97.5%
Independent seed’s final video on the same maze36.0%58.6%53.5%52.5%52.0%

At step 1, the motion overlaps the two final videos to a similar extent. By step 5 of 40, 88.6% of the predicted motion lies near the same seed’s final motion, compared with 53.5% for an independent seed. The early convergence we report in Fig. 2 therefore remains visible when all branches contribute to the measurement, and is not caused by potential averaging. We also include example frames so readers can see how the extractor handles a prediction with multiple branches in Appendix E.2, Figure 12 (L1312).

W.2/Q.2: Longer videos and chaining controls

We directly tested whether longer videos improve success on the same long-path mazes. We evaluated Wan2.2-14B on all 64 Frozen Lake mazes requiring 10–13 moves, using four matched seeds per maze and increasing generation length from 81 to 121 frames. We also report ChEaP’s result on the same 64 mazes.

MethodMazes solved
Wan2.2, 81 frames4/64 (6.3%)
Wan2.2, 121 frames14/64 (21.9%)
ChEaP42/64 (65.6%)

Longer videos do improve success on these mazes. Increasing the duration raises the number of mazes solved from 4 to 14, although 50 of 64 remain unsolved at 121 frames. ChEaP solves 42 of the same 64 mazes by chaining shorter generations. We agree that path length alone cannot distinguish insufficient video duration from increased planning difficulty. The longer-video experiment shows that duration contributes to success, while leaving open how much of the remaining difficulty comes from planning the route or executing it within a single video. We revise Section 5 to clarify this interpretation (L430).

We also added a separate comparison to address your concern about ChEaP’s access to the maze layout and its higher NFE budget (Appendix C, Table 5). On 145 Frozen Lake mazes, we gave both non-chained baselines the same obstacle information and matched their budgets to ChEaP’s mean cost at each grid size, rounded up to the next 40 NFEs. This gave budgets of 800, 840, and 640 NFEs on 6×6, 8×8, and 10×10 mazes, respectively. Under these settings, ChEaP solved 48 mazes, obstacle-aware EPBS solved 23, and obstacle-aware best-of-N solved 12.

Finally, we kept the chaining and pivot-selection rules fixed and generated every candidate in full. This increased the number solved from 48 to 51, while raising the mean cost from 731 to 1,842 NFEs per maze. Full generation thus required 2.52x as many NFEs per maze for three additional solves.

W.3/Q.3: Generalization to broader tasks and verifier design.

EPBS requires a verifier that can rank early decoded predictions from different seeds according to task-relevant progress. We chose simple verifiers that measure progress directly. For regular and irregular mazes, we use spatial distance to the goal. For LIBERO, we use a SAM 2-based object-relation score. These verifiers are inexpensive enough to apply across many seeds before deciding which generations to complete.

For broader tasks when such relations are difficult to specify by hand, a learned or VLM-based verifier conditioned on the task instruction could provide a more flexible score. These verifiers would need to reliably evaluate intermediate predictions and the computational cost must remain low to preserve EPBS’s efficiency gains. We use lightweight verifiers for this reason, while we acknowledge that tasks without readily verifiable progress remain challenging.

Thank you for the detailed response and the additional experiments.

Multiple branches. Table 7 resolves my concern about Fig. 2. My question on whether the endpoint-based score prefers wall-crossing or discontinuous trajectories remains unanswered; a brief note on how EPBS handles such cases would help.

Video duration and controls. Good job. Please make sure to state the revised interpretation clearly in Section 5.

Verifier design. I truly understand the limitation and I suggest explicitly discussing it in the limitations section, ideally with any preliminary evidence on whether a VLM-based verifier can reliably score noisy early decoded predictions.

Overall, the response addresses most of my concerns, and I maintain my positive assessment.

Reviewer_PZZWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies when video diffusion models settle on the trajectories they eventually generate. On maze tasks, routes extracted from intermediate clean-sample predictions already resemble the final route fairly early, and re-noising these predictions produces less route diversity than sampling fresh seeds. The authors use this observation to build EPBS, which briefly probes many seeds and finishes only the promising ones. They then introduce ChEaP, which chains valid partial generations to tackle longer mazes.

Critical strengths
  1. The early-commitment observation is useful because the paper turns it into an actual sampling improvement. The authors measure intermediate-to-final route similarity, check whether re-noising changes the route, and use the results to decide where to spend compute. This makes the analysis more convincing than a few examples of recognizable early predictions.
  2. EPBS is simple and gives meaningful improvements over best-of-N without retraining. I also appreciate that the authors account for the extra VAE decoding cost and report wall-clock comparisons. Those costs could easily eat up the benefit of early screening, so showing that some gains survive this accounting matters.
  3. The experiments give a useful picture of where the approach helps and where it runs out of room. The second video model, the breakdown by maze size, and the tests beyond regular grids add substance. The paper also reports limited gains on harder HunyuanVideo cases and failures on very long paths, which helps put the positive results in perspective.
Critical weaknesses
  1. I could not reconcile the path-length breakdown in Fig. 6b with the stated evaluation set. If the four bins partition the same 302 Frozen Lake mazes, with one binary top-2 outcome per maze and method, the displayed percentages do not admit integer group sizes summing to 302 under ordinary rounding. Fig. 6a, by contrast, is consistent with 76, 78, 77, and 71 mazes across the four grid sizes. There may be a straightforward explanation, such as a different subset or an averaging procedure, but I could not find it. This matters because the 9.1% to 72.7% improvement on paths of length 10–13 is a central piece of evidence for chaining. I would want to see the group sizes and success counts before relying on that result.
  2. ChEaP changes more than the generation horizon, so it is hard to tell how much of its gain comes from chaining itself. It gets access to the obstacle layout for rejecting invalid candidates and uses about 630 NFEs on Wan, compared with the 400-NFE EPBS and best-of-N rows. The paper is upfront about both differences, but the main comparison still leaves their effects mixed together. The privileged-verifier analysis in Table 1 is helpful, though it does not replace an end-to-end comparison under the same information and compute budget.
  3. The success checks could use more validation. Appendix A.4 does not clearly explain how teleportation, wall crossings between tracked positions, or goal drift are rejected. On LIBERO, the 400-NFE gain is only two tasks out of 116, and the judge disagrees with humans on 4 of 55 audited videos.
Questions for authors
  1. Could you provide the number of mazes and successful outcomes for each method in every Fig. 6b bin? Is this exactly the same evaluation set and configuration as Fig. 6a, or does it use a subset, repeated trials, or another averaging rule? This would resolve my main concern about the reported numbers.
  2. Do you have results comparing ChEaP with non-chained EPBS or best-of-N using the same obstacle-aware filtering and a matched compute budget? A full-generation sampling method with the same chaining and pivot-selection rules would also help establish how much early screening contributes within ChEaP. These comparisons would affect my assessment of the method’s contribution.
3 comments
AuthorsOpenReview ↗

Reply to Reviewer PZZW

Dear Reviewer PZZW,

Thank you for the kind words about our paper, particularly for recognizing our concrete validation of early plan commitment and attention to wall-clock-adjusted NFE accounting. Your concerns are (1) consistency of our path-length breakdown, (2) apples-to-apples comparison of ChEaP with budgets and verifiers, and (3) extra success validation.

W.1/Q.1: Figure 6(b) path-length breakdown

Your numerical check of Fig. 6b was correct. Fig. 6b was mistakenly computed on a narrower subset that omitted the 10×10 mazes. We apologize for this mismatch. The corrected breakdown below uses the same 302-maze evaluation set and the corresponding configurations in Figure 6a. Each maze contributes one binary top-2 outcome per method.

Optimal path lengthMazesBest-of-NEPBSChEaP
≤ 5 moves7668 (89.5%)71 (93.4%)75 (98.7%)
6–97451 (68.9%)64 (86.5%)71 (95.9%)
10–13646 (9.4%)17 (26.6%)42 (65.6%)
≥ 14880 (0.0%)3 (3.4%)8 (9.1%)
All302125 (41.4%)155 (51.3%)196 (64.9%)

We have updated Fig. 6b to reflect this correction (L394). The largest improvement remains on mazes with optimal paths of 10–13 moves. Including the 10×10 mazes changes the best-of-N and ChEaP comparison from the originally reported 9.1% versus 72.7% to 9.4% versus 65.6%.

W2/Q.2: Equal compute budget and verifier comparisons between EPBS and ChEaP

To isolate the benefit of chaining, we give non-chained baselines the same obstacle information as ChEaP and match their NFE budgets to its mean cost at each maze size.

On the 145 Frozen Lake mazes (6×6–10×10) where chaining runs, ChEaP averages 796, 815, and 601 NFEs by grid size. We round the baselines’ budgets up to 800, 840, and 640 NFEs, respectively. Obstacle-aware EPBS uses ChEaP’s lake-penalized progress score; obstacle-aware best-of-N rejects full videos entering a lake using ChEaP’s pivot-validity check.

Mazes solved (of 145)ChEaPEPBS (obstacle-aware)Best-of-N (obstacle-aware)
6×6–10×10482312

With the same obstacle information and comparable compute, ChEaP solves 48 mazes versus 23 for non-chained EPBS and 12 for best-of-N.

We also tested your suggestion of keeping the chaining and pivot-selection rules fixed while fully generating every candidate. Full generation solves 51 mazes at an average of 1,842 NFEs per maze, compared with 48 mazes at 731 NFEs for ChEaP. Fully generating every candidate therefore requires 2.52× as many NFEs per maze for three additional solves. With early screening, ChEaP solves 94% as many mazes using 60% less compute on average. We have added both comparisons to Section 5 and Appendix C (L422, L1079, Table 5)

W3: Success check validation

For maze evaluation, we agree that the original Appendix A.4 did not explain the success checks clearly enough. We now describe our framewise SAM 2 tracking, how we detect forbidden-cell crossings, and how we handle large jumps and goal drift (L830-861). On Frozen Lake and VR-Bench, we track the agent in every frame. Entering a lake or crossing a wall before reaching the goal is an automatic failure. We also check cells skipped between frames, so a jump along a row or column that crosses a lake fails just as walking into it would. A diagonal jump fails if both corner cells are forbidden. We count the first valid goal visit when the agent enters the goal cell or touches the gift from an adjacent cell.

We also re-evaluated the results under a stricter criterion that treats jumps of three or more cells and gift movement exceeding 30% of its bounding-box size as failures. For best-of-N and EPBS, we retained the original verifier-based candidate selection and recomputed success under the stricter criterion. For ChEaP, we additionally checked every segment of each successful chain and its concatenated trajectory against the original maze. Every segment passed the stricter checks, and every complete trajectory reached the original goal without entering a lake. The stricter evaluation leaves all Wan2.2 Frozen Lake results in Figs. 6a and 6b unchanged.

For LIBERO, the Gemini 3 Flash judge agreed with human labels on 51/55 audited videos (92.7% agreement), with three false passes and one false fail. We agree that the two-task advantage at 400 NFEs is modest given these disagreements, and we have revised the text to reflect this. The clearest improvement is at the lower budget of 120 NFEs, where EPBS solves 97/116 tasks (83.6%) under the judge, compared with 86/116 (74.1%) for best-of-N. This gives EPBS an 11-task advantage at the same compute budget. EPBS also reaches 87.1% at 160 NFEs, compared with 85.3% for best-of-N at 400 NFEs, achieving a comparable judged success rate with 40% of the compute budget. These results demonstrate that EPBS provides its greatest benefit in compute-constrained settings.

Reviewer_PZZWOpenReview ↗

Thanks for your reply.

Thanks for the author response. I went through the paper again and also looked at the other reviewers' comments. I think the rebuttal addressed my earlier concerns pretty well.

I see this as a nice exploration paper. Through a fairly detailed set of experiments, it gives good evidence for the early plan commitment phenomenon. The work is also pretty complete. It doesn't just show that coarse structures show up early in the diffusion process—it also measures when the task-related trajectories become fixed. The authors use DTW, different models, different maze sizes, and a bunch of additional analyses to rule out some simple alternative explanations.

The paper also turns this observation into a practical inference-time compute allocation strategy. It combines early screening with long-horizon chaining, and doesn't need any extra training. I think readers can get some useful insights from this work. For example, a video model's "plan" may be decided much earlier than the final video is done. This could also matter for understanding video reasoning, world models, and action models. The results also suggest that failures on long-horizon tasks may come from the limited video duration, not just from the difficulty of reasoning itself.

Overall, I think the paper gives useful insights and backs them up with a solid set of experiments. So I've decided to raise my initial score and now lean toward acceptance.

Thank you for taking the time to revisit our paper and for your encouraging assessment. We really appreciate your constructive feedback which has helped us strengthen this paper.

Reviewer_ymvnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the underlying reasoning dynamics of diffusion video models when solving visual reasoning tasks. To facilitate this investigation, the authors use 2D maze-solving as a controlled testbed. A major finding is that video diffusion models tend to commit to a high-level motion plan within the first few denoising steps. Building on this finding, they propose ChEaP, an improved inference strategy that screens models’ early plans to allow for more exploration. Empirical results demonstrate the efficiency of ChEaP.

Critical strengths
  1. An important problem: understanding the underlying reasoning dynamics of video generation models may provide insights into how to improve model design.
  2. A controlled testbed: 2D maze-solving is a visual reasoning task that minimizes perceptual complexity, allowing for a focused study of visual reasoning.
  3. An improved inference strategy: building on the observed early commitment to plans, the authors propose an inference strategy that screens plans and improves inference efficiency.
Critical weaknesses
  1. The scope of the study is limited in terms of model paradigms and tasks. Specifically, the authors only consider diffusion video models, without experimenting with autoregressive video models, making the title’s claim that “video models reason early” overly broad. Additionally, 2D mazes represent a narrow set of visual reasoning tasks. Although irregular mazes and robot manipulation tasks are included, it is unclear how they complement the 2D maze experiments.
  2. The proposed ChEaP strategy is evaluated only on Wan and Hunyuan video models, so it is unclear whether it generalizes across video-model paradigms. I would encourage the authors to include autoregressive video models. Moreover, including proprietary video models would enable a more comprehensive comparison.
Questions for authors

A similar early-plan-commitment phenomenon was observed in previous work [1] (see Figure 3). It would be interesting if the authors could establish a connection to that work or clarify how their findings differ.

[1] Probing Visual Planning in Image Editing Models, https://arxiv.org/abs/2604.22868

2 comments
AuthorsOpenReview ↗

Reply to Reviewer ymvn

Dear Reviewer ymvn,

Thank you for recognizing the importance of understanding the underlying reasoning dynamics of video diffusion models and the value of our controlled testbed and inference-time scaling strategy. Your concerns are (1) the scope of our claims and tasks, (2) applicability to autoregressive and proprietary video models, and (3) our comparison with related work. We address each below.

W.1: Scope of the claims and tasks

Our work investigates when video diffusion models commit to motion plans during denoising, and how this property can be exploited through EPBS and ChEaP. Our title describes how "Video Models Reason Early," which our abstract immediately specifies as "Video diffusion models...". Our key finding is likewise explicitly introduced in Section 3 as "Early Plan Commitment in Video Diffusion Models." We agree that our findings should not be interpreted as establishing this phenomenon across all families of video generators, and clarify that our contribution focuses on video diffusion models.

The second half of our title, "Exploiting Plan Commitment for Maze Solving," establishes the experimental setting. We choose mazes because "each maze has a well-defined solution, and correctness can be checked in an automated manner" (L44-45). This allows us to "investigate the intermediate samples of the denoising process to understand how the solutions take shape" (L45-46). We adopt this setting over alternatives such as symmetry puzzles, which have less well-defined difficulty and intermediate reasoning states. We further extend our experiments to settings with progressively different task structures. Irregular mazes remove the discrete gridworld structure while retaining objectively verifiable navigation goals. LIBERO extends the approach to diverse object-manipulation tasks, requiring a different, object-relation-based verifier. These experiments demonstrate that verifier-guided early screening can be adapted beyond grid-based navigation, using different measures of task progress.

Together, these settings allow us to study early planning in a controlled environment and investigate the transfer of our inference-time scaling approach to more general tasks.

W.2: Autoregressive and proprietary models

Our primary experiments use Wan2.2-14B and HunyuanVideo-1.5, two recent open video diffusion models. Models from both families have also been included in recent video reasoning evaluations [1]. For proprietary models, our analysis requires access to intermediate clean-sample predictions during denoising. The proprietary video APIs available do not expose these intermediate predictions, preventing us from performing the same analysis. Nonetheless, we also investigated whether EPBS could be applied to an autoregressive diffusion-forcing video model, SkyReels-V2-DF-14B. On 20 Frozen Lake mazes (10 4×4 and 10 6×6), evaluated with four seeds per maze, SkyReels solved 0/80 generations. In comparison, Wan 2.2 solved 11/80 generations, with at least one successful generation on 7/20 distinct mazes. SkyReels frequently failed to preserve the task structure: the agent remained stationary or disappeared, or the model altered the maze itself, leaving no meaningful plan for EPBS to select or ChEaP to extend. We provide the setup, failure examples, and evaluation details in Appendix E.3.

This experiment illustrates an important prerequisite of EPBS and ChEaP, which is that the underlying generator must be capable of producing a nontrivial number of valid candidate solutions. These methods improve inference-time utilization of existing generation capabilities, but they do not provide a model with task-solving capabilities it does not already possess. We therefore view the SkyReels experiment as an initial investigation of applicability to autoregressive diffusion-forcing generation, rather than evidence that our approach is generally ineffective for autoregressive models. Establishing how EPBS and ChEaP can be adapted to stronger autoregressive video generators remains an interesting direction for future work.

Q1: Relation to Zhou et al. (2026)

Thank you for pointing us to this related work. Zhou et al. (2026) investigate when spatial paths emerge in an image-editing model. In contrast, we study temporally ordered motion trajectories across video frames in off-the-shelf video diffusion models. We quantify when these trajectories converge toward their final plans during denoising and exploit this property to guide inference-time computation through EPBS and ChEaP. We cite this concurrent work in our revised draft and clarify the distinction in Section 2 (L103-105).

References

[1] Wang, Maijunxian, et al. "A very big video reasoning suite." arXiv preprint arXiv:2602.20159 (2026).

I thank the authors for their rebuttal. My main concern remains the limited task diversity. Irregular mazes address this concern to some extent, but they do not sufficiently capture the range of possible maze geometries. The LIBERO experiments broaden the evaluation, but their connection to maze solving remains unclear. Thus, I maintain my original assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.