10,900 automatically collected desktop trajectories, 54,900 state-action samples, and a fine-tuned agent that went from 5.9 to 64.1 percent on specialized desktop workflows, past the 59.8 percent reported for humans
AspireAI community · arXiv (Liu, Ding, Yang, Li, Wu, Zhao, Chen, Wang, Sun & Qiao)
Computer-using agents do well on general desktop tasks and badly on specialized applications, mostly because the training trajectories for those applications do not exist. OS-Wizard collects them automatically by combining tutorial-guided exploration with a hybrid GUI-and-code action space, so the agent can invoke application-native commands directly instead of clicking through cumbersome workflows, and by checkpointing and revisiting pivotal states during tree exploration to cover alternative paths rather than one linear rollout.
The corpus is 10.9K trajectories and 54.9K state-action samples. Fine-tuning open-source base models on it raises performance on a benchmark of specialized desktop workflows by 58.2 percentage points to 64.1 percent, above the reported human performance of 59.8 percent on the same benchmark and a new state of the art. The same fine-tuned model also improves on general desktop workflows using GUI actions only, which suggests the collected supervision transfers beyond the specialized case, and the ablations show the code actions are doing significant work.
The limitation is that the benchmark and the trajectories are the authors' own construction, so the human comparison point is a reported figure rather than a independently controlled study.
https://arxiv.org/abs/2610.33907
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.