Multi-step reasoning means using what is known to decide what follows and what to check next. It matters when an answer depends on context, not just a familiar pattern: diagnosing a patient, interpreting a scene, or noticing that an automated system has made an obvious mistake. section 8.0 section 6.2.2 section 8.1
Working toward a diagnosis
A diagnostic process first checks for emergencies, such as blood loss. If none is present, it considers an obvious explanation, such as a reported broken arm, and confirms it with a decisive test. If the answer remains unclear, it looks for facts that distinguish likely diagnoses, favoring safer, cheaper, less invasive tests before more burdensome ones. Each step changes which question is worth asking next. section 8.0
Visual understanding offers another example. A child can connect a banana’s changing color with ripeness and infer that a banana-shaped piece of plastic will not ripen, peel, or be eaten. Recognizing the shape alone does not supply those relationships. section 6.2.2
These examples set a practical limit on automation. A model trained on data may work well within its domain yet struggle with common-sense or context-based reasoning; a system can also repeat an error a person readily notices. Use automated outputs as evidence to assess, not as a substitute for checking whether the conclusion fits the situation. section 7.3 section 8.1
