Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features

arXiv:2608.06008 · cs.RO · Submitted 2026-08-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features".

Dev: Here's the gist of the paper. Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory.

Rosa: The single most informative sentence about the paper's subject and finding is: Adaptive-WAM, a quality-aware multi-exit planner built on a Wan2.2-5B backbone.

Dev: Pruning redundant reasoning tokens The paper introduces Adaptive-WAM to address the question of how much of a video diffusion model must be executed to make a reliable driving decision.

Rosa: This is motivated by the observation that planning performance is largely insensitive to tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers.

Paper summary: Rosa: So we're looking at this paper called "Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features". It deals with using those big video diffusion models that give us rich information about where things are and how they move in time for autonomous driving. But the existing world action models still often have to pay the cost of generating future videos, even though what we actually need when we're driving is just a simple path or trajectory.

Dev: Exactly, Rosa. The core question this paper asks is how much of that massive video diffusion model do you really need to run to make a safe driving decision? It’s motivated by the idea that planning performance doesn't change much if you test it with different levels of video noise, and strong trajectories can actually be decoded from the middle layers of the model.

Taro: So, they propose Adaptive-WAM as this quality aware world action model that changes how much of the video backbone gets executed based on how good the plan looks so far. It keeps a Wan two point two five billion parameter backbone and attaches a five step trajectory diffusion transformer to six intermediate blocks.

Rosa: And they use a small scorer, called DINOv2 Small, to predict sub scores for each trajectory that gets decoded. If the best trajectory found so far passes a quality threshold, they stop running things immediately. It’s a quality aware approach that routes the computation based on the plan's current quality.

Dev: That sounds like it cuts out all that iterative classifier free denoising loop and VAE decoding that you usually need for future video synthesis. They claim this method avoids generating any future video during deployment, which is a big deal for real time systems.

Taro: The training process involves adapting the Wan backbone using lo ra while optimizing the trajectory projections and heads fully. The actor objective they use includes terms related to both the visual and action losses, plus terms related to the trajectory itself. After imitation learning, they freeze the Wan and that quality scorer part of the model.

Rosa: Then they refine those trajectory heads using a diff grpo formulation, which is similar to what you see in other systems like ReCogDrive. They also fine tune that quality scorer on generated trajectories using evaluator provided component targets with equal weight soft label bce.

Dev: The performance gains they show come from this quality guided routing mechanism and the selective execution of backbone depth. They say the adaptive policy hits ninety point seven nine percent PDMS when using a threshold of ninety, which is a significant improvement over the fixed b15 planner by zero point one seven PDMS while cutting latency down to one hundred seventy milliseconds.

Taro: They also mention that this system transfers well across different datasets, reaching an average two ten percent eight eight m average l squared error and a zero point zero eight percent collision rate on nu scenes without needing any target domain fine tuning. That's interesting for deployment in the wild.

Rosa: But they also did some analysis showing that while block fifteen is the strongest fixed exit, later blocks still hold scene specific advantages because exits solve overlapping but not identical scene sets. The scorer diagnostic shows that only fifty one scenes, or zero point four two percent, select a trajectory at least fifty percent worse than an available near perfect trajectory.

Dev: That suggests the quality guided fallback is practical because passing that quality threshold isn't a formal safety certificate. They report that Adaptive-WAM gets ninety point eight PDMS on navsim one and eighty nine point nine epdms on navsim two.

Taro: The efficiency aspect is also significant because the adaptive planner averages an end to end planning latency of one hundred seventy milliseconds at a threshold of ninety, which is about ten percent lower than that fixed block fifteen planner and much less than the three hundred twenty milliseconds that a fixed full depth planner takes.

Rosa: They keep the full video generation runtime separate, saying it takes about thirteen point two two seconds when using cached text context. This comparison only holds for the planning path because Adaptive-WAM only executes one conditional pass up to the selected depth and never decodes a future video during that planning step.

Dev: The scorer backbone evaluation showed that Wan features don't offer any meaningful advantage over dino small while substantially increasing online cost. They found that the best wan exit only improves the diagnostic score by zero point zero three points over DINO small, which is why they use DINO small as the quality verifier because a wan based scorer would add a large world model forward at every attempted exit.

Taro: The conclusion they draw is that intermediate video diT features are robust planning representations well before full image synthesis completes. They conclude that Adaptive-WAM allocates world model computation according to the quality of the current plan.

Rosa: So what this means for us is that we can use these large generative world models for planning without having to run them all the way through to generate a full video or even synthesize a future frame, just by intelligently selecting which parts of the model are necessary based on how confident we are in our current plan.

Dev: It's about making that computational cost manageable. The paper's final contribution is that they systematically diagnose how video noise level and diT depth affect driving decisions. They also introduce a multi exit world action architecture where a learned trajectory quality controller dynamically selects the required diT depth without completing video generation.

Taro: And they evaluated planning, transfer, trajectory scoring, and efficiency on navsim one and two obtaining state of the art planning performance among compared world model methods while improving over every fixed singletrajectory exit at lower average cost than the strongest one.

Rosa: So basically, we get better performance with less computation by being smart about when to use the heavy video models. The paper's final conclusion is that intermediate video diT features are robust planning representations well before full image synthesis completes, and Adaptive-WAM allocates world model computation according to the quality of the current plan.

Dev: And they show strong zero shot transfer results on nu scenes too, with an average two ten point three five m average l squared error and a zero point zero eight percent average collision rate. It seems like a very practical path from large generative world models to efficient driving policies.

Conclusion: Rosa: So we've been talking about Adaptive-WAM, which is this new way to use those huge video diffusion models for planning by only running what you need based on how good the plan is so far.

Dev: Yeah, that's the main idea. It’s about making sure you don't waste all that massive computation just trying to predict a future video when you only need a short path right now.

Taro: The authors, they built this quality aware world action model using a Wan two point two five billion parameter backbone and some trajectory diffusion transformer attached to six intermediate blocks.

Rosa: And they use this small scorer called DINOv2 Small to check the quality of each possible plan at each step, and if the plan passes a certain threshold, they just stop. It’s a quality guided early-exit planner.

Dev: That makes sense for real time control because you don't want to wait for a full video synthesis loop when you're trying to make an immediate driving decision.

Taro: They show that this adaptive approach gets ninety point seven nine percent PDMS on NAVSIM and eighty nine point nine EPDMS on NAVSIM two, which is good performance.

Rosa: But the really interesting part for me is how it handles the world misbehaving. It adapts its computation based on the plan's quality, so if things get tricky, it can adjust what it runs next.

Dev: And that’s where I see the safety case. If you have a strong trajectory already decoded in those intermediate blocks, you don't need to run the whole massive model to find it.

Taro: It also handles transfer well across different datasets without needing any extra fine tuning, which is a big deal for real world deployment outside of just one test track.

Rosa: So, basically, this work shows we can use these complex video models for planning by being smart about when and how deep we need to look into the data.

Dev: It's about making that computational cost manageable in a way that fits real time constraints. And it opens up a new path from these big generative world models to policies that actually run on cars without needing full video generation every time.

Sining Ang, Yuguang Yang, Yan Wang

Institute for AI Industry Research (AIR) at Tsinghua University · Department of Automation, University of Science and Technology of China · School of Electronic Information Engineering, Beihang University

cs.RO

Submitted: 2026-08-06

Updated: 2026-10-30

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 100/100

The gist: Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though…

Key concepts

Adaptive-WAM
This is a quality-aware multi-exit planner built on a Wan2.2 backbone. It dynamically selects the appropriate depth of video diffusion features to use for planning, terminating inference early once a trajectory meets a predefined quality threshold.
Video Diffusion Models
These large models provide rich spatiotemporal information useful for autonomous driving. The paper uses intermediate layers from these models as robust representations for planning, suggesting that planning can be done effectively before the entire video generation process is complete.
Quality-Guided Routing
This mechanism involves using a lightweight scorer to predict planning sub-scores for different decoded trajectories. This score guides the system to select only the most promising trajectory, avoiding unnecessary computation by skipping deeper or lower-quality planning steps.

Terminology

Summary

Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory. p. 0

Where the paper says it

cn, wangyan@air. cn Abstract Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory.

The single most informative sentence about the paper's subject and finding is: Adaptive-WAM, a quality-aware multi-exit planner built on a Wan2.2-5B backbone. p. 0
Where the paper says it

Based on this observation, we introduce AdaptiveWAM, a quality-aware multi-exit planner built on a Wan2. 2- 5B backbone.

Pruning redundant reasoning tokens

The paper introduces Adaptive-WAM to address the question of how much of a video diffusion model must be executed to make a reliable driving decision. This is motivated by the observation that planning performance is largely insensitive to tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. p. 0

Where the paper says it

Through a controlled study of video denoising timesteps and Diffusion Transformer (DiT) depth, we find that planning performance is largely insensitive to the tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. Based on this observation, we introduce AdaptiveWAM, a quality-aware multi-exit planner built on a Wan2.

The authors propose Adaptive-WAM as a quality-aware, layer-adaptive world-action model. p. 0
Where the paper says it

This evidence motivates allocating computation according to the quality of the current plan rather than committing to one readout depth. Based on this observation, we propose Adaptive-WAM, a quality-aware, layer-adaptive world-action model.

This architecture retains a Wan2. p. 0
Where the paper says it

Through a controlled study of video denoising timesteps and Diffusion Transformer (DiT) depth, we find that planning performance is largely insensitive to the tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. Based on this observation, we introduce AdaptiveWAM, a quality-aware multi-exit planner built on a Wan2.

2-TI2V-5B backbone and attaches a ReCogDrive-style five-step trajectory DiT to six intermediate blocks. p. 0
Where the paper says it

2025) backbone and attach a ReCogDrive-style five-step trajectory DiT (Li et al. 2025e) to six intermediate blocks.

The lightweight DINOv2-Small scorer predicts NAVSIM planning sub-scores for each decoded trajectory, terminating inference once the best trajectory decoded so far satisfies a quality threshold. p. 0
Where the paper says it

A lightweight DINOv2-Small (Oquab et al. 2024) scorer predicts NAVSIM planning sub-scores for each decoded trajectory.

How the training signal is built

The training process involves several stages to adapt the model and train the components. During imitation learning, the Wan backbone is adapted using LoRA while trajectory projections and heads are optimized in full. p. 4

Where the paper says it

During imitation learning, the Wan backbone is adapted using LoRA (Hu et al. 2022a), while the trajectory projections and heads are optimized in full.

The actor objective is defined as Lactor = λvidLvid + Xl∈E λlLltraj. p. 4
Where the paper says it

2022a), while the trajectory projections and heads are optimized in full. The actor objective is Lactor = λvidLvid + X ℓ∈E λℓL ℓ traj, (8) where Lvid is the native Wan video-diffusion objective.

After imitation learning, the Wan and quality scorer are frozen. p. 5
Where the paper says it

Planner-only reinforcement learning. After imitation learning, we freeze the Wan backbone and quality scorer and refine each trajectory head using the DiffGRPO (Li et al.

The trajectory heads are then refined using the DiffGRPO formulation adopted by ReCogDrive. p. 2
Where the paper says it

We use the same trajectory-space formulation at every exit, providing a controlled action-decoding interface for studying intermediate video-DiT representations and adaptive backbone depth. After imitation learning, planner-only DiffGRPO refines the exit heads while leaving the video backbone fixed.

The quality scorer is fine-tuned on generated trajectories using evaluator-provided component targets with equal-weight soft-label BCE. p. 5
Where the paper says it

The scorer, including its DINOv2-Small encoder, is fine-tuned on generated trajectories using evaluator-provided component targets. All six components use equal-weight soft-label BCE-with-logits: Lscore = X i X m∈R BCElogit ai,m, roracle i,m , (9) where each oracle component is used directly as a target in [0, 1], without binarization or component-specific regression losses.

Where the gains come from

The performance gains stem from the quality-guided routing mechanism and the selective execution of backbone depth. The adaptive policy reaches 90. p. 1

Where the paper says it

Second, adaptive exits improve the performance– computation frontier. At the selected threshold, the adaptive policy reaches 90.

79 PDMS when using a threshold η = 90, improving on the fixed B15 planner by 0. p. 6
Where the paper says it

08 Table 4: nuScenes comparison using horizon-averaged L2 error and collision rate; FT denotes target-domain fine-tuning. Policy PDMS ↑ Exit by B15 Lat.

17 PDMS while reducing latency to 170 ms. p. 6
Where the paper says it

At η = 90, the adaptive policy improves on fixed B15 by 0. 17 PDMS while reducing average end-to-end planning latency by approximately 10%, from 190 to 170 ms.

The system avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis. p. 0
Where the paper says it

Trajectory diffusion heads are attached to selected DiT blocks, and a lightweight trajectory-quality scorer terminates inference once the best trajectory decoded so far satisfies a quality threshold; otherwise, computation continues from the cached hidden state to a deeper exit. The deployed planner therefore avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis, while dynamically allocating backbone depth according to trajectory quality.

Furthermore, the learned representation transfers across datasets, reaching 0. p. 1
Where the paper says it

Its average end-to-end planning latency is 170 ms, approximately 10% lower than the 190 ms fixed block-15 planner and 47% lower than the 320 ms fixed full-depth planner. Third, the learned representation transfers across datasets, reaching 0.

88 m average L2 error and a 0. p. 0
Where the paper says it

Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0. 88 m average L2 error and a 0.

08% collision rate on nuScenes without target-domain fine-tuning. p. 1
Where the paper says it

08% collision rate on nuScenes (Caesar et al. 2020) without target-domain fine-tuning.

The analysis of layer-wise planning quality shows that while block 15 is the strongest fixed exit, later blocks retain scene-specific advantages. Post-RL Jaccard overlap indicates that exits solve substantially overlapping but non-identical scene sets. p. 2

Where the paper says it

Post-RL high-quality set overlap B5 B9 B15 B18 B22 B30 B5 B9 B15 B18 B22 B30 1. 00 Figure 2: Post-RL Jaccard overlap of high-quality scene sets.

The scorer diagnostic shows that only 51 scenes (0. p. 6
Where the paper says it

Table 4 retains only averages in the main paper; horizon-wise results and supervision details are provided in Appendix L. Scorer Reliability and Adaptive Trade-off The scorer is evaluated on an offline diagnostic pool covering 12,146 scenes.

42%) select a trajectory at least 50% worse than an available near-perfect trajectory. p. 6
Where the paper says it

Only 51 scenes (0. 42%) select a trajectory at least 50% worse than an available near-perfect trajectory, and 69 scenes (0.

This suggests that the quality-guided fallback is practical, as passing the quality threshold is not a formal safety certificate. p. 4
Where the paper says it

Predicted Future Scene Images Decision Head / Trajectory Generator . Trajectory branch fits future trajectories Trajectory Branch Candidate Trajectory Adaptive expansion & acceleration Continue inference to next selected DiT layer 5 9 15 18 22 30 (Final) Core Idea • Preserve WAM architecture • Extract intermediate DiT features for decision making • Use discriminator-guided early exit for faster inference Loop Discriminator Trajectory Quality Scorer Score > Threshold?

The final results show that Adaptive-WAM achieves 90. p. 6
Where the paper says it

By coupling multi-depth trajectory heads with a tieaware trajectory-quality controller, Adaptive-WAM allocates world-model computation according to the quality of the current plan. It achieves strong NAVSIM and zero-shot nuScenes results while exposing a practical path from large generative world models to efficient driving policies.

p. 6
Where the paper says it

By coupling multi-depth trajectory heads with a tieaware trajectory-quality controller, Adaptive-WAM allocates world-model computation according to the quality of the current plan. It achieves strong NAVSIM and zero-shot nuScenes results while exposing a practical path from large generative world models to efficient driving policies.

p. 6
Where the paper says it

By coupling multi-depth trajectory heads with a tieaware trajectory-quality controller, Adaptive-WAM allocates world-model computation according to the quality of the current plan. It achieves strong NAVSIM and zero-shot nuScenes results while exposing a practical path from large generative world models to efficient driving policies.

8 PDMS on NAVSIM v1 and 89. p. 0
Where the paper says it

On NAVSIM, the adaptive single-trajectory planner achieves 90. 8 PDMS; a separate fixed-exit variant reaches 92.

p. 0
Where the paper says it

On NAVSIM, the adaptive single-trajectory planner achieves 90. 8 PDMS; a separate fixed-exit variant reaches 92.

p. 0
Where the paper says it

On NAVSIM, the adaptive single-trajectory planner achieves 90. 8 PDMS; a separate fixed-exit variant reaches 92.

9 EPDMS on NAVSIM v2. p. 0
Where the paper says it

It further obtains 89. 9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners.

p. 0
Where the paper says it

It further obtains 89. 9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners.

p. 0
Where the paper says it

It further obtains 89. 9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners.

The zero-shot transfer to nuScenes is also strong, with Adaptive-WAM obtaining 0. p. 5
Where the paper says it

Zero-shot transfer. Without nuScenes fine-tuning, Adaptive-WAM obtains 0.

p. 5
Where the paper says it

Zero-shot transfer. Without nuScenes fine-tuning, Adaptive-WAM obtains 0.

p. 5
Where the paper says it

Zero-shot transfer. Without nuScenes fine-tuning, Adaptive-WAM obtains 0.

35 m average L2 error and a 0. p. 0
Where the paper says it

Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0. 88 m average L2 error and a 0.

p. 0
Where the paper says it

Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0. 88 m average L2 error and a 0.

p. 0
Where the paper says it

Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0. 88 m average L2 error and a 0.

08% average collision rate. p. 0
Where the paper says it

88 m average L2 error and a 0. 08% collision rate.

p. 0
Where the paper says it

88 m average L2 error and a 0. 08% collision rate.

p. 0
Where the paper says it

88 m average L2 error and a 0. 08% collision rate.

The efficiency gains are significant, as the adaptive planner averages an end-to-end planning latency of 170 ms at η = 90. This is approximately 10% lower than the 190 ms fixed block-15 planner and substantially less than the 320 ms fixed full-depth planner. p. 1

Where the paper says it

62 for the strongest fixed single-trajectory exit. Its average end-to-end planning latency is 170 ms, approximately 10% lower than the 190 ms fixed block-15 planner and 47% lower than the 320 ms fixed full-depth planner.

The full video generation runtime takes approximately 13. p. 6
Where the paper says it

This comparison is kept separate from the 13. 22 s full video-generation runtime.

22 s when using cached text context. p. 5
Where the paper says it

The adaptive planner, fixed block-15 planner, and fixed full-depth planner average 170, 190, and 320 ms, respectively. For context, with cached text context, full 40-step classifier-free future-video generation takes 13.

This comparison is kept separate from the full video-generation pipeline because the adaptive planner only executes one conditional pass up to the selected depth and never decodes a future video. p. 6
Where the paper says it

the corresponding zero-shot setting. However, DriveVA executes the full Wan backbone and generates future images to support trajectory prediction, making its inference cost substantially higher and closer to a full Wan video-generation pipeline than to our single conditional early-exit planning pass.

The primary 320-to-170 ms claim compares two planning paths and is therefore kept separate from full video generation. p. 6
Where the paper says it

This comparison is kept separate from the 13. 22 s full video-generation runtime.

The scorer backbone evaluation shows that Wan features offer no meaningful advantage over DINO-Small while substantially increasing online cost. The best Wan exit improves the diagnostic score by only 0. p. 6

Where the paper says it

55 for ResNet-34/50. Wan features offer no meaningful scorer advantage while substantially increasing online cost: the best Wan exit improves the diagnostic score by only 0.

03 points over DINO-Small. p. 1
Where the paper says it

Across scorer-backbone ablations with Wan, ResNet, ViT, and DINO features, the best Wan variant improves the diagnostic score by only 0. 03 points over DINOv2-Small while incurring substantially higher inference cost.

This confirms that DINO-Small is used as the quality verifier because a Wan-based scorer would add a large world-model forward at every attempted exit. p. 0
Where the paper says it

Because trajectory rewards are highly saturated and frequently tied, the scorer predicts metric components and acts primarily as an exit-quality verifier rather than imposing a strict total ordering over trajectories. We therefore evaluate both tie-aware selection and consequential large-gap errors.

The paper concludes that intermediate video-DiT features are robust planning representations well before full image synthesis completes. p. 6
Where the paper says it

22 s full video-generation runtime. 6 Conclusion In summary, intermediate video-DiT features are robust planning representations well before full image synthesis completes.

The deployment-comparable video path includes denoising and both VAE operations but not online T5 encoding or file saving. p. 5
Where the paper says it

Hardware latency is measured at batch size one on a single A100 80GB. End-to-end planning latency includes VAE image encoding, the conditional Wan forward, every attempted trajectory head, and scorer evaluation.

This demonstrates a practical path from large generative world models to efficient driving policies. p. 6
Where the paper says it

By coupling multi-depth trajectory heads with a tieaware trajectory-quality controller, Adaptive-WAM allocates world-model computation according to the quality of the current plan. It achieves strong NAVSIM and zero-shot nuScenes results while exposing a practical path from large generative world models to efficient driving policies.

The paper's contributions include systematically diagnosing how video-noise level and DiT depth affect driving. The authors conclude that Adaptive-WAM allocates world-model computation according to the quality of the current plan. p. 6
Where the paper says it

6 Conclusion In summary, intermediate video-DiT features are robust planning representations well before full image synthesis completes. By coupling multi-depth trajectory heads with a tieaware trajectory-quality controller, Adaptive-WAM allocates world-model computation according to the quality of the current plan.

The final results show that Adaptive-WAM achieves strong NAVSIM and zero-shot nuScenes results while improving over every fixed singletrajectory exit at lower average cost than the strongest one. p. 1
Where the paper says it

• We introduce a multi-exit world-action architecture whose learned trajectory-quality controller dynamically selects the required DiT depth without completing video generation. • We evaluate planning, transfer, trajectory scoring, and efficiency on NAVSIM v1/v2 and nuScenes, obtaining stateof-the-art planning performance among compared worldmodel methods while improving over every fixed singletrajectory exit at lower average cost than the strongest one.

The paper's contributions are: we systematically diagnose how video-noise level and DiT depth affect driving, we introduce a multi-exit world-action architecture whose learned trajectory-quality controller dynamically selects the required DiT depth without completing video generation, and we evaluate planning, transfer, trajectory scoring, and efficiency on NAVSIM v1/v2 obtaining stateof-the-art planning performance among compared worldmodel methods while improving over every fixed singletrajectory exit at lower average cost than the strongest one. The paper's final conclusion is that intermediate video-DiT features are robust planning representations well before full image synthesis completes. The paper's final conclusion is that Adaptive-WAM allocates world-model computation according to the quality of the current plan. The final results show that Adaptive-WAM achieves 90.8 PDMS on NAVSIM v1 and 89.9 EPDMS on NAVSIM v2. The zero-shot transfer to nuScenes is also strong, with Adaptive-WAM obtaining 0.35 m average L2 error and a 0.08% average collision rate.

The paper's final conclusion is that intermediate video-DiT features are robust planning representations well before full image synthesis completes. The paper's final conclusion is that Adaptive-WAM allocates world-model computation according to the quality of the current plan. The final results show that Adaptive-WAM achieves 90.8 PDMS on NAVSIM v1 and 89.9 EPDMS on NAVSIM v2. The zero-shot transfer to nuScenes is also strong, with Adaptive-WAM obtaining 0.35 m average L2 error and a 0.08% average collision rate.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive WAM System Improvement: The system can dynamically allocate backbone depth according to trajectory quality, avoiding unnecessary computation by terminating inference once the best trajectory decoded so far satisfies a quality threshold. This avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis.

  2. Efficient Planning Latency Reduction: The adaptive routing mechanism improves the performance–computation frontier, achieving an average end-to-end planning latency of 170 ms compared to a fixed block-15 planner's 190 ms and a fixed full-depth planner's 320 ms.

  3. Robust Representation Transfer: The model demonstrates strong zero-shot transfer capabilities, reaching 0.88 m average L2 error and a 0.08% collision rate on nuScenes without target-domain fine-tuning, indicating the learned representation is highly generalizable across driving datasets.

  4. Quality Verification Mechanism: A lightweight DINOv2-Small scorer predicts planning sub-scores for each decoded trajectory, acting as an exit-quality verifier rather than imposing a strict total ordering over trajectories. This allows the system to select the highest scoring accumulated trajectory from multiple exits.

  5. Training Optimization via Gradient Isolation: During imitation learning, the quality scorer is trained such that nabla theta Wan, theta G Lscore = 0, ensuring that the actor and scorer updates alternate on the same training stream without affecting trajectory generation objectives.

Abstract

Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory. We ask a more basic question: how much of a video diffusion model must be executed to make a reliable driving decision? Through a controlled study of video denoising timesteps and Diffusion Transformer (DiT) depth, we find that planning performance is largely insensitive to the tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. Based on this observation, we introduce Adaptive-WAM, a quality-aware multi-exit planner built on a Wan2.2-5B backbone. Trajectory diffusion heads are attached to selected DiT blocks, and a lightweight trajectory-quality scorer terminates inference once the best trajectory decoded so far satisfies a quality threshold; otherwise, computation continues from the cached hidden state to a deeper exit. The deployed planner therefore avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis, while dynamically allocating backbone depth according to trajectory quality. On NAVSIM, the adaptive single-trajectory planner achieves 90.8 PDMS; a separate fixed-exit variant reaches 92.6 PDMS with 64 proposals. It further obtains 89.9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners. Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0.88 m average L2 error and a 0.08% collision rate. On an A100, adaptive routing improves PDMS from 90.62 to 90.79 while averaging 170 ms end-to-end planning latency, approximately 10% below the 190 ms fixed block-15 planner and 47% below the 320 ms fixed full-depth planner. Code will be released.

Sources

Related papers