On 650 enterprise procedures, pre-generating a whole workflow instead of calling the model each step stayed less accurate for every model tested, but the gap shrank by 79 percent and three fixes closed 10 to 12 points of it
AgentOrch compares two ways to run an enterprise workflow agent. Step-by-step calls the model after every action so it can observe and adapt, at a real inference cost. Global orchestration writes the complete executable workflow up front and runs it without further LLM calls, which is much cheaper but less accurate. The benchmark holds tools, environment and evaluation fixed across 650 enterprise standard operating procedures and 3,220 runtime variants so the comparison is about orchestration style rather than task difficulty.
Step-by-step stayed more accurate for every model evaluated, but the gap narrowed by about 79 percent from early models to the strongest recent one. Three routes narrow it further: incremental workflow construction adds an average of 9.98 percentage points without any runtime execution feedback; revising complete workflows using execution feedback improves every model, including 6.13 points on held-out runtime variants; and targeted training on complete workflows adds an average of 12.18 points across all five trained models.
The limitation is that a gap still remains in every configuration tested, and the workloads are enterprise SOPs rather than open-ended tasks, so the numbers do not transfer directly to general agentic use.
https://arxiv.org/abs/2610.33375
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.