A trajectory-quality scorer let a driving planner stop early inside a video diffusion backbone, keeping most of the planning accuracy while skipping the denoising loop
AspireAI community · arXiv
Large video diffusion models give rich spatiotemporal priors for autonomous driving, but world-action models inherit the cost of iterative future-video generation even though deployment needs only an ego trajectory. The authors ask how much of a video diffusion model must actually run to produce a reliable driving decision, and find that planning performance is largely insensitive to the tested video-noise levels, while strong trajectories can already be decoded from intermediate layers. Adaptive-WAM is a quality-aware multi-exit planner on a Wan2.2-5B backbone: trajectory diffusion heads attach to selected DiT blocks, and a lightweight trajectory-quality scorer terminates inference once the best trajectory decoded so far meets 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. On NAVSIM the adaptive single-trajectory planner reaches 90.8 PDMS, a fixed-exit variant reaches 92.6 PDMS with 64 proposals, and it obtains 89.9 EPDMS on NAVSIM v2, the best reported among the compared front-view video world-model planners. Without target-domain fine-tuning it transfers to nuScenes with 0.88 m average L2 error and a 0.08% collision rate, and on an A100 adaptive routing moves PDMS from 90.62 to 90.79 while cutting cost. The limitation is that the gain over the dense baseline is small on the headline metric, so the value is mostly in latency and cost rather than in planning quality.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.