Most agent benchmarks measure whether the task got finished. DMAIC-Bench measures something else: whether the agent stopped when the evidence said it should. It evaluates self-restraint, the ability to act on intermediate evidence by continuing, redirecting or stopping, with 97 seeded tasks across 15 industrial cases in four sectors, built around the explicit statistical prerequisites of Lean Six Sigma.
The design uses paired counterfactual variants that require opposite actions at the same workflow position, so indiscriminate stopping is distinguishable from warranted restraint. Execution traces are scored by a private scorer and final claims by independent expert review.
The gap is stark. Across 13 models from six providers, execution-level routing accuracy averaged 92 percent on matched GO decisions and 19 percent on NO-GO decisions, counting any forbidden downstream computation as a route error. Execution traces show agents reporting correct diagnostics and then performing the analyses those diagnostics disallow. Separately, blinded expert review of 11 audited agents estimated that reports act on invalid evidence at 47 percent of reached NO-GO decisions, with an interval from 31 to 61 percent. Explicit stopping permission helped some models without consistently restoring correct decisions at both forks.
Caveat: the tasks are statistical-prerequisite reasoning scenarios, not general-purpose agent work.
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.