Giving the same tools to an agent through MCP and through a command line produced different behaviour in 94.3 percent of 2,004 episodes, and a fix for it recovered up to 9 points of score
AspireAI community · arXiv (Shi, Pang & Li)
MICA renders MCP servers and command-line interfaces from one shared semantic action contract and routes both through the same runtime, which removes the usual confound: in most agent comparisons the interface and the capability change together, so nobody can tell which one moved the result. With capabilities genuinely matched, the interface still matters. Across six models and three interactive environments, task performance and interaction cost differed substantially, and the divergence was not just cosmetic - it showed up in which semantic actions were selected, where runs failed, and how they recovered. Paired trajectories found structured divergence in 94.3 percent of 2,004 episodes, with early differences propagating into later turns.
The paper's intervention, DART, decouples action realization from the validated semantic commitment and repairs the interface-specific part, which improved benchmark score in all six settings by up to 9.0 points and cut MCP token cost in all three environments by up to 27.1k tokens per task.
The limitation is that the underlying capabilities are synthetic contract-matched actions, not real third-party MCP servers, so the magnitude should not be read as a measurement of any particular production integration.
https://arxiv.org/abs/2610.34298
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.