Paired tests of a frozen acceptance rule separated web-agent verifier failures from agent failures, 32 of 33 correct deliveries against 1 of 11
AspireAI community · arXiv
A reusable verifier for web agents has to follow requirements that change and still support subsequent execution. AgentProofBench connects paired tests of a frozen acceptance rule to execution from identical saved openings, spanning 97 tasks, 79 families and four web services. With DeepSeek Flash and Pro it measures two linked things: whether complete-panel and decision-replay tests retain the source programme, and whether deployment-state panel correctness predicts whether the two saved openings are actually delivered.
Across 48 families, complete-panel and decision-replay tests retain 82 of 96 Flash and 79 of 96 Pro source programmes. Across 22 execution families, panel correctness separates cleanly: 32 of 33 assignments with error-free panels delivered both openings correctly, against 1 of 11 with panel errors, and every reference-defined programme-gate regression and withheld delivery fell into that latter group. Neutral continuation attained the highest correct delivery under the generic feedback tested, giving a concrete baseline for measuring what a gate adds.
The honest limit: the resource covers 97 tasks over four web services, so the separation is demonstrated on a narrow slice of agent behaviour.
https://arxiv.org/abs/2610.34634
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.