TinyAstral shows a robot agent's harness, not just its planner, is where the capability lives
AspireAI community · arXiv
Robot agents built on vision-language models depend on the planner and on the harness around it - the tools, feedback, memory and recovery logic. TinyAstral asks whether a large teacher's harness gains can be carried into a compact agent at lower inference cost, in two stages: first a frozen open-weight teacher has its harness improved by recursive self-improvement, where frontier LLM agents read the teacher's trajectories and propose harness edits and each edit is accepted only if it passes a paired statistical test; then the final harness is frozen, teacher trajectories with reasoning and failure recovery are collected, and a small student is distilled into that same harness. On RoboTwin the authors report harness evolution raises the teacher's success from 33.4% to 63.6% and the final 4B student reaches 62.8%; across four harness versions distilled 4B students score close to their corresponding teachers, later versions yield stronger students, and the harness gains also transfer to other frozen planners without retraining. The 4B student reportedly beats open 4B agents, the open 27B harness-RSI model SenseNova-RoboRSI and their own teacher placed in three competitors' harnesses, while using about one fifth of its teacher's accelerator time. The limitation is that the harness-RSI loop depends on frontier LLM agents to propose the edits, so the training-time cost is not removed, only shifted.
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.