A benchmark of 18 open VLA models on Jetson and Ryzen AI edge hardware found that how a policy generates actions moves latency far more than model size does
AspireAI community · arXiv (cs.RO, cs.AI)
Vision-language-action models get benchmarked on task success, which says nothing about whether they can actually run inside a robot's control loop. VLA-HWBench measures 18 open VLA models on NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395 and breaks inference into seven stages, and its central finding is uncomfortable for anyone shopping on parameter count: action-generation strategy dominates 14 of 19 Jetson configurations, and parameter count explains at most 15 percent of latency variance.
The measured spreads are large. Latency varies by 82x on Jetson and 157x on AMD across the configurations tested, and the paper attributes the gap to whether a model emits long action chunks in a few flow-matching steps or decodes token by token - only the former keeps up with control. On the software side, kernel fusion with 4-bit weights helps most, running pi_0.5 and GR00T N1.7 2.6 to 2.9x faster than PyTorch, but that path exists for just those two models and only on the Jetson. Optimization also pushes time into the stages it shortens least: CPU preprocessing grows from 12 percent to 39 percent of the GR00T N1.7 call, and after optimization memory bandwidth binds only the token-light stages while software limits everything else.
The authors' recommendation is concrete - few action-generation iterations, broader kernel-fusion and low-bit weight support, and edge hardware with high memory bandwidth and a fast host CPU. The caveat is scope: two edge platforms, eighteen open models, and the authors' own harness, with the huge latency multiples coming from different decoding strategies on the same hardware rather than from differences in model capability.
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.