GUI agents kept solving CAPTCHAs correctly and still failed the task: removing the binding step dropped end-to-end success from 82.22 to 71.56 percent
AspireAI community · arXiv (Song, Peng, Su, Wu, Song, Liu, Zhang & Liu)
CAPTCHAs and similar anti-automation checks judge behavior - interaction trajectories and timing - not just answers. SolveBind names the gap: an agent can solve a verification task correctly and still fail end to end, because the correct answer is never converted into an accepted interaction. Its hierarchical solving tree decomposes verification into interdependent Perceive, Ground, Solve, Act and Verify capabilities, while a runtime supervisor grounds action commitments in tool-derived evidence, revises them as evidence arrives, and mediates execution under explicit recovery budgets.
Across nine CAPTCHA task families and nine backbone models, SolveBind achieves the highest success rate on eight backbones and beats the strongest baseline by 13.3 to 34.5 percentage points. It is also best on all six AndroidWorld backbones. In a paired GPT-5.5 analysis, MobileAgent produces correct answers but fails dynamic verification on 13 of 90 tasks, and SolveBind succeeds on seven of those.
The mechanism is not free, and the paper prices it: removing the binding step drops success from 82.22% to 71.56% while raising latency and token use, and removing recovery caps lifts success only from 82.22% to 83.56% while multiplying latency by 3.7 and token use by 2.4. The limitation is scope: the benchmark is built around protected GUI environments, so the numbers do not transfer directly to ordinary desktop automation.
https://arxiv.org/abs/2610.32223
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.