Across 1,229 remote-sensing agent trajectories, a tool that ran fine still returned wrong measurements at least 22.7 percent of the time, and the error reached the final answer at least 82 percent of the time
AspireAI community · arXiv (Liu, Yu, Wang, Feng, Yao, Zhou, Yang, Ren, Wan & Feng)
Tool-using agents are supposed to get sharper from precise measurements, but a tool can execute successfully and still be wrong. RSure-Agent studies that gap in remote sensing agents, where perception, measurement and raster-analysis tools return quantitative judgments about ground objects. Across 1,229 execution trajectories on three agent benchmarks, at least 88.1% of tasks depend on tool observations, and among those tasks at least 22.7% contain incorrect observations despite the tool executing successfully.
Those errors propagate to the final answer in at least 82.0% of affected tasks on each benchmark. The proposed fix is a verifiable observation protocol that requires tools to return process evidence the agent can check, combined with a task-tool reliability prior built from offline task feedback that summarizes how each tool configuration has performed across task types. With both, the agent decides whether to accept an observation, request more evidence, or reject it.
Evaluated on EarthBench, ThinkGeo, TerraLogic and CHOICE-420, this cuts the error-propagation rate by 21.3 to 25.9 percentage points against the same setup with verification and prior disabled. On CHOICE-420 it improves overall accuracy over direct answering by 5.71 points on average across 11 backbone models, and on EarthBench it reduces the tool-call ratio by 25.9% relative to Earth-Agent. The limitation is scope: this is Earth-observation work with raster tools, the reliability prior needs prior feedback per tool configuration, and only a demo is released - the code is still promised.
https://arxiv.org/abs/2610.04836
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.