A runtime rule that no running policy ever waits on the reasoning model lifted success to 95.9 percent on 60 LIBERO-Pro tasks while cutting control steps by a quarter
AspireAI community · arXiv
Foundation models reason well at a high level but are too slow and costly to serve as fine-grained controllers, while fast policies such as vision-language-action models and whole-body controllers are efficient reflexes that struggle to generalize. Pairing a deliberative brain with a reactive cerebellum is natural for embodied systems, but methods are built and scored on incompatible stacks and their asynchronous coordination is difficult: an agent that calls the VLA inside the agentic loop can stall fast execution, while a VLA that calls the agent on errors tends to re-trigger fallbacks repeatedly or miss silent failures entirely. REA (Routed Embodied Agent) is a unified runtime where a deliberative System 2, an embodied router System 1 and reactive System 0 policies cooperate under a single rule, that no running policy waits on the brain. Instead of querying the brain at each failure, REA keeps a route memory of each policy's measured success, adds a boundary-triggered judge gate that consults the brain on policy execution, and uses a shared-budget escalation ladder for bounded recovery. Evaluated on 3,000 test episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success against 94.7% for the best single-agent baseline in its own policy pool, while reducing control steps by 25% and control time by 17%. The limitation is that the policy pool is the runtime's own, so the gain is over a baseline drawn from the same set rather than over the current state of the art on LIBERO-Pro as a whole.
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.