A radar-guided kinematic prior inside an attention-free state-space core forecast 4D occupancy at 122 frames per second on one consumer GPU
AspireAI community · arXiv
Real-time 4D occupancy forecasting requires reliably extrapolating both ego motion and surrounding actor motion under strict latency budgets, and recent occupancy world models improve long-horizon fidelity through generative decoders and 3D volumetric modeling at a computational cost that makes real-time deployment difficult. Light-weight bird's-eye-view forecasters meet latency requirements but often blur or suppress fast-moving actors at longer horizons. UFMF (Ultra-Fast Multi-modal Forecaster) keeps an attention-free diagonal state-space temporal core while injecting targeted kinematic and geometric structure where lightweight models typically fail: a radar-guided velocity-aware dynamic prior decomposes future displacement into a linear extrapolation plus a learned horizon-dependent residual, and the resulting per-cell flow fields transport a compact dynamic latent subspace during rollout. Two auxiliary training constraints complement this branch, with world-frame dynamic supervision enforcing inter-horizon trajectory coherence and class-conditioned vertical-profile supervision regularizing BEV-to-3D lifting. UFMF operates at 122 FPS on a single RTX 4090 and reaches 0.45 fast-displacement mIoU on nuScenes, outperforming recent occupancy world models at a fraction of their computational cost. The limitation is that the speed comes with a narrowed output, since it forecasts occupancy rather than generating video of the future, so it does not replace a generative world model for applications that need rendered futures.
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.