Two-Stage Reasoning–Answer Extraction

First prompt for a reasoning trace, then a second self-augmented prompt that appends an answer-format trigger and parses the final answer.

Zero-shot chain of thought is implemented as two feedforward passes:

Reasoning extraction — wrap the question as Q: … A: [trigger] (e.g. “Let’s think step by step”) and generate a continuation reasoning trace (often greedy for reproducibility).
Answer extraction — concatenate the first prompt, the generated reasoning, and a format trigger tailored to the output type (“Therefore, the answer (arabic numerals) is …”, “among A through E …”), then generate and parse (\hat{y}).

Few-shot CoT often embeds the answer format in exemplars and needs only one call; zero-shot CoT buys lower exemplar engineering cost at the price of two model calls plus cleansing/parsing. In a harness, treat the second stage as an explicit post-processing contract: keep format triggers versioned with the prompt catalog, run deterministic answer cleansing, and fail closed when parsing cannot recover a canonical answer (same discipline as action format enforcement for tool turns). Use stop sequences (e.g. "Q:") when needed to prevent the model from emitting another question–answer turn instead of finishing extraction.

Sources

Large Language Models are Zero-Shot Reasoners (Kojima et al.), §3.1