Fixed pauses cost GUI testing agents 42.3 percent of their time, and 62.9 percent of that wait was already wasted
AspireAI community · arXiv preprint
Agent evaluations tend to blame model latency. Meituan's industrial GUI-testing data says a large share of the clock is something else: the fixed pause after every screenshot.
In the sampled interaction intervals, waiting accounted for 42.3% of elapsed time, close to the 54.8% spent on model calls, and 62.9% of that waiting happened after the interface had already finished loading - roughly 58 seconds per click sitting on a finished screen. WaitLens attacks that with a small vision-language model fine-tuned on GUI loading trajectories annotated with their true end times, using temporal resampling so one app's loading pattern does not have to generalize to every other app. It reaches 86.08% accuracy on the keep-waiting decision against 54.13% for the backbone VLM, and replaying Meituan's own pipeline it cuts mean extra waiting after the interface is ready by 94.98% against a 20-second fixed wait.
The practical takeaway is that a screen that has finished loading can be detected, so a test agent does not need to sleep for a worst-case guess. Caveat: the evaluation replays recordings from one industrial pipeline, and the comparison is against a fixed 20-second wait rather than other adaptive-waiting schemes.
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.