Agent skills were scored on nine dimensions without ever executing them, and that execution-free quality estimate improved both task success and execution efficiency downstream
AspireAI community · arXiv
Agent skills package instructions, scripts and resources for specialized tasks and are increasingly distributed through public web repositories, but existing skill benchmarks focus on downstream task execution, which misses specification quality, behavioral robustness, semantic alignment and responsibility boundaries and does not scale because it needs task-specific tools, environments and repeated runs. SkillCT broadens skill assessment to nine dimensions across static analysis of the specification, dynamic checks of whether induced behavior stays stable across backbone models and request variations, and human-assisted checks of behavior-chain consistency and task granularity. Applied to a public skill corpus it becomes a reference benchmark for analyzing quality weaknesses, and the authors add an execution-free baseline that predicts the multidimensional quality profile of an unseen skill by retrieving and leveraging similar reference skills without executing it, reducing token consumption and evaluation time while staying consistent with full assessment. SkillCT-guided optimization then improves downstream task success and execution efficiency. The limitation is that the execution-free baseline predicts quality from similar skills rather than measuring the skill, so it is a proxy for unseen skills.
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.