Checking approval freshness at the moment of execution, not at the moment of request, lifted on-time radio delivery from 77.3 to 89.9 percent in a live radio emulation
AspireAI community · arXiv
LLM-based agentic control of radio access networks looks like a good fit for automation, but action-level safeguards alone are not enough: gathering evidence consumes time and shared resources, and network state can change so that an approval is already invalid by the time it would be acted on. This work proposes an executor-side model and contract that governs intents from assessment through to release or termination, combining local checks with selective retrieval of coordination evidence to cut acquisition cost while keeping mandatory evidence requirements, version and lease rechecks that prevent release under invalidated approvals, and reevaluation that requires recorded progress within each transaction's retry budget. Building on that contract, the authors formulate evidence allocation to maximize the expected rate of timely valid releases under risk, quality-of-service, deadline and shared byte-budget constraints, with thresholds selected offline and interpolated using an online load-phase estimate, reporting better timely valid release rates than arrival-order fetching and an offline global threshold under identical quotas. The concrete field result is that an immediate scripted resend after a version-mismatch refusal raised on-time receiver delivery from 77.3 to 89.9 percent at modest additional command cost. The limitation is that the evidence comes from radio emulation workloads with stable daily load cycles, which is a friendlier setting than a busy network.
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.