Structured output is the arrangement of a model response into recognizable parts rather than an undifferentiated continuation of text. It can range from direct instruction–response turns to headings, lists, line breaks, emojis, reasoning, and a required final answer string. Format is distinct from tone: tone changes the manner of an answer, while format changes its structure. section 1.2 section 4.0
Instruction fine-tuning establishes structured output as an interface. A pretrained next-token model may continue a question with related text; training on instruction–response examples instead teaches it to answer the query directly. Chat templates then serialize system prompts, user queries, and assistant responses into token sequences with special tokens, while prompt and multiturn masking determine which response tokens receive training signal. section 4.0 section 4.4
Output format also determines how behavior is measured:
- Few-shot multiple-choice evaluation can require generating an answer letter such as A, B, C, or D, then score it with exact match; alternatively, conditional log-likelihood can compare candidate-token probabilities without sampling. section 16.1.1
- Exact-match evaluation may require rigid suffixes such as “The answer is:” or use regular expressions to find an answer in generated text. The choice makes formatting part of the measured result. section 16.1.1
- Reasoning-era prompts can request concise reasoning but require the response to end with an exact answer phrase; special formatting may separate thinking tokens from answer tokens. section 16.1.4
- A mismatch between evaluation prompting and post-training format can drive benchmark performance close to zero. Even small answer-format differences across training datasets can conflict and reduce performance. section 16.2 section 16.6
A constrained multiple-choice response
A reasoning prompt can demand concise reasoning while fixing the final field: “Therefore, the answer is (ANSWER_LETTER).” This makes the answer machine-checkable by exact match while allowing generated reasoning before it. section 16.1.4
Answer the following multiple-choice question by giving the correct answer
letter in parentheses.
Provide CONCISE reasoning for the answer, and make sure to finish the
response with "Therefore, the answer is (ANSWER_LETTER)" where
(ANSWER_LETTER) is one of (A), (B), (C), (D), (E), etc.
Question: {question}
(A) {choice_A}
(B) {choice_B}
(C) ...
Answer the above question and REMEMBER to finish your response with the
exact phrase "Therefore, the answer is (ANSWER_LETTER)" where
(ANSWER_LETTER) is one of (A), (B), (C), (D), (E), etc.
The prompt listing makes the split between free-form reasoning and an exact, checkable final answer concrete. section 16.1.4
Post-training changes output format alongside style and tone. RLHF is used with instruction fine-tuning and other methods to produce responses that are more helpful, warm, and engaging, but optimizing these properties can overproduce length or chattiness and harm performance elsewhere. In evaluation, format must therefore be controlled as a configuration rather than treated as a cosmetic detail. section 1.7 section 16.6
