Saltar al contenido principal

Cadena de pensamiento (CoT)

Definición

El prompting de cadena de pensamiento (CoT) pide al modelo que genere pasos intermedios de razonamiento before the final answer. This often improves accuracy on math, logic, and multi-step tasks.

Es one of the simplest razonamiento patterns: sin herramientas ni búsqueda, solo prompting. Úselo cuando the task benefits from explicit steps (por ej. arithmetic, deduction) and you want to avoid fine-tuning. For exploring multiple solution paths, see tree of thoughts; for tool-using agents, see ReAct.

Cómo funciona

You give the model a question (or task) and ask it to reason paso a paso. The model produce Step1, Step2, … (intermediate razonamiento) and then the answer. Zero-shot CoT: add “Let’s think paso a paso” (or similar) to the prompt. Few-shot CoT: include example (question, steps, answer) triples so the model mimics the format. The model generates the sequence in one pass; you can optionally parse the steps and verify or score them. Quality depends on prompt engineering and model capability.

Casos de uso

Chain-of-thought es más útil cuando la tarea se beneficia de pasos intermedios explícitos (matemáticas, lógica, código).

  • Math and arithmetic where intermediate steps improve accuracy
  • Logic puzzles and multi-step deduction
  • Code or diseño razonamiento where showing steps aids debugging

Documentación externa

Ver también