Using LiDAR geometry to pick image tokens before the vision backbone ran cut end-to-end FLOPs by up to 18 times and still beat the dense baseline
AspireAI community · arXiv
Vision-centric autonomous driving models encode thousands of image tokens per frame, most of which depict sky, buildings or empty road and dilute the planner's attention. The authors show that randomly discarding 75% of tokens already outperforms the dense model, because the planner can concentrate on the remaining driving-critical content. Geometry-Aware Token Selection (GATS) uses LiDAR and radar point projections to select a small fraction of image tokens before the vision backbone runs, and because the selector operates on measured 3D geometry rather than visual features, the backbone encodes only the selected patches, cutting end-to-end FLOPs by up to 18x. Each kept token receives a 3D anchor from its LiDAR point and feeds a self-supervised latent world model that predicts future features and 3D displacements without box annotations. At a 25% budget GATS reaches 0.520 m 3-second L2 displacement error on nuScenes and 88.9 EPDMS on NAVSIM, beating the dense baseline of 0.573 m and 85.9 at 6.6x fewer total FLOPs; at an extreme 10% budget it still beats the dense baseline on both benchmarks. Across 739 intent-critical scenarios it covers 95.8% of hazardous vehicles with a 0.09% collision rate, 73.5% lower than the dense model's 0.34%. The limitation is that the method depends on LiDAR or radar being available, so it does not transfer to camera-only stacks without a substitute geometry source.
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.