A recipe that teaches a 27B model to control a robot through a simple tool harness hit 96.7 percent success on real-world tasks with no real robot data in training
AspireAI community · arXiv (Benjamin Pikus, Daniel Saunders, Mircea Mironenco et al.)
A robot policy needs to understand what a task is asking for in a scene and how to move the robot to accomplish it. ZAP splits those two skills: a pretrained 27B parameter model bootstraps world understanding and reasoning, then supervised fine-tuning and GRPO on simple simulated tasks teach the movement. Without any real-world robot data during training, ZAP reaches 96.7% success on real-world tasks, matching GPT-6 Astra in the paper's harness and outperforming the MolmoAct2 VLA (33.3%), while running 3x faster and 17x cheaper than GPT-6 Astra. It generalises to objects, goals, strategies and language absent from training, including seating a test tube in a rack and choosing a different hole in the rack when prompted, handing a test tube between arms mid-air, drawing an X on a tic-tac-toe board and playing chess. It also transfers to an unseen embodiment, reaching 55.9% on RoboLab-120 with a single-arm Franka, 4.9 percentage points ahead of the leading VLA/WAM baseline. The limitation: the real-world evaluation is the authors' own harness and task set, and the 27B bootstrap model is far larger than the small vision-language model it teaches, so the compute accounting differs from a native small-policy approach.
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.