An agent that could not reliably pick the right enterprise API did not need a better model, it needed the process graph - and multi-step task completion went from 10.3 to 38.7 percent
AspireAI community · arXiv
*Retrieval-augmented service discovery treats each API endpoint as an independent candidate, so an agent can pick the right booking function and still fail by inventing a fare code, dropping a required reference or booking the hotel before the flight.* This work folds OData metadata and BPMN reference processes into a typed property graph, and a hybrid retriever uses dense embeddings to find candidate endpoints and then follows process edges to retrieve the endpoints that come next. On 42 enterprise services with process-chain ground truth, queries where the wording has little lexical overlap with the downstream entities improve Recall@3 by 5.7 points across embedding models and cut-offs, and crucially the method shows no gain on OpenAPI services that have no process layer - which is the evidence that the gain comes from the process structure rather than from the extra text. Run as an agent on 81 multi-step queries, it completes 38.7 percent of tasks against 10.3 percent for a chunk-based baseline. The elicitation probe is the sharpest number: simply supplying valid code lists removes invented codes but still leaves invalid keys in 46 percent of write runs, while full typed context cuts invalid calls to 3 percent, 3 of 101 runs, with no invented codes at all. The limitation is the small service count and that the process layer has to exist somewhere for the retriever to follow.
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.