Llm agents struggle to stop or redirect correctly in multi-step tasks
DMAIC-Bench: From Self-Reflection to Self-Restraint for Agents in Multi-Stage Analytical Tasks
Artificial IntelligenceComputation and LanguageMachine Learning
Summary
When large language models (LLMs) work on tasks that need many steps, they must decide at each step whether to move forward, change direction, or stop. The paper's authors point out that while LLMs can reflect on their work, they often fail to change their actions properly based on what they learn, which they call a lack of self-restraint. They created a new test called DMAIC-Bench with tasks from industry and science to see how well models make these decisions. Their results show models do well at continuing correctly but perform poorly in stopping or redirecting when they should.
What this means in practice
- •For industrial data science teams: Improve decision-making processes in multi-step industrial workflows by evaluating and refining agent stopping and redirecting actions using DMAIC-Bench.
- •For bioinformatics developers: Test and enhance evidence-driven decision points in bioinformatics pipelines, such as RNA-seq and GWAS analyses, using the same evaluation framework.
Authors
Shan Peng, Shuaiqi Cheng, Miao Peng, Chenyi Zi, Jia Li
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.