Post-Reasoning: answer first, justify later — a free way to boost LLM quality

One prompt, one request:

YOUR QUESTION/TASK
+ "First give the final answer, then justify it."

→ Answer: [final conclusion]
→ Justification: [reasoning after the answer]

The reasoning comes after the answer — so you get the result right away, without waiting for the model to "think." Read the justification if you want to.

Task: You're preparing for an investor meeting. You need to decide quickly whether to meet Sarah Collins from Harbor Street Ventures, given that she specializes in e-commerce and you run a B2B SaaS for HR teams.

Prompt:

I'm the founder of a B2B SaaS product for HR automation 
(onboarding, employee ramp-up). Average contract value is $15,000/year, 
customers are companies with 200+ employees. I've been offered a meeting with 
an investor who specializes in e-commerce startups 
and marketplaces. Should I take it?

First give the final answer: yes or no. Then justify it.

Result: The model will put a clear "Yes" or "No" on the first line — no "on the one hand... on the other hand..." preamble. Then it'll lay out the arguments: why the decision is justified, which risks and opportunities were weighed. You see the conclusion immediately, and the reasoning backs it up.

The LLM's weakness in standard mode: the model generates text token by token and sort of "drifts" toward an answer — without any firm commitment to a final conclusion. With no explicit "state your conclusion" instruction, it can circle the topic and land on a weak or fuzzy verdict.

The LLM's strength: models follow structural instructions well. When the prompt has an explicit order ("first X, then Y"), the model sticks to it. An instruction on how to answer shapes the content of the answer — even before any reasoning gets generated.

How the method works: the phrase "answer first, justification after" creates a hidden commitment. The model generates the final conclusion already "knowing" an explanation will follow. That shifts the probability distribution for the very first token of the answer — making it more accurate. It's like a good speaker: if they know they'll be asked to explain their logic, they state their thesis more cleanly.

Control knobs: - Add "in one sentence" after the answer instruction → forces brevity and removes vagueness - Drop the justification requirement → you're back to standard mode, but the insight is already applied - Add "justify step by step" instead of just "justify" → you get a structured explanation as a bonus - Set the answer format ("yes/no," "recommend/don't recommend") → makes the commitment even firmer

{Question or task}

First give the final answer: {answer format — one sentence / 
yes or no / a recommendation}. Then justify it.

What to fill in: - {Question or task} — any request where you need a clear conclusion - {answer format} — the format you want for the final answer: "one sentence," "yes or no," "a score from 1-10"

Phrase variants for different tasks:

→ For analysis:
"First state the final conclusion in one sentence. 
Then explain the logic."

→ For decisions:
"First give a recommendation: what to do. 
Then — why exactly that."

→ For reviewing a text:
"First rule on it: does the text work or not. 
Then go through the specific reasons."

🚀 Quick start — paste into the chat:

Here's a Post-Reasoning template. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.

{Question or task}

First give the final answer: {answer format — one sentence / 
yes or no / a recommendation}. Then justify it.

The LLM will ask which answer format you need and what exactly to analyze — because without that it can't set up the right "answer → justification" structure for your context. It'll take the pattern from the template and adapt it to your task.

⚠️ Simple tasks: On questions an LLM already handles well (simple arithmetic, factual questions with an obvious answer), the effect is minimal — improvement is near zero or absent.

⚠️ Inconsistency: For some models and tasks the results got worse (12% of cases). This was especially noticeable on certain AMC 12 problems for specific models. The technique isn't universal — but it works most of the time.

⚠️ Hard tasks need more: On olympiad math, the prompt method gives a solid gain, but fine-tuning the model on this principle gives an even bigger one. The prompt version is the ceiling without final training, not the final solution.

⚠️ Knowledge tasks — smaller effect: If the task calls for plain factual knowledge rather than reasoning, the gain is smaller than on tasks with multi-step logic.

Researchers from Nanyang Technological University took 13 models — from small 4B ones to a big 70B, including open models (Llama, Gemma, Mistral, Qwen) and commercial ones (GPT-5.4-Mini, Claude Haiku, Gemini Flash) — and simply added one instruction to every request: "answer first, then justify." They compared this against a direct answer with no prompt engineering at all.

Interestingly, the biggest gains came from the small models — where baseline quality is lower. A small model with post-reasoning often beats its own no-post-reasoning self by 30–140%. The logic makes sense: a hint about output structure helps a weaker model more — it has less "safety margin" for building an answer on its own.

The real surprise: self-distillation turned out better than distillation from an expert model. When a model was trained on its own explanations (rather than GPT-4's), it performed better. The reason: the explanations match its "inner language," so there's no clash of styles. It's an important principle: explanations in someone else's style can hurt rather than help.

Prompt-based Post-Reasoning instruction (from Appendix A.6):

"What is 10 + 2? State the final answer immediately 
and justify your answer."

General instruction augmentation (δp):
"State the final result first, then justify."

Context: The authors used this formula as the base prompt template for all tests. That minimal addition alone produced gains in 88%+ of cases.

💡 Adaptation for working with text

The same principle helps with editing and evaluating content — when you need a clear verdict, not "gentle criticism."

Here's my post for LinkedIn:

{post text}

First give a verdict: should this post be published as is — 
yes or no. One word. 
Then explain the specific reasons.

The model won't water down the answer with reasoning — you'll see a clear position first, then the breakdown.

🔧 Technique: add a constraint to the answer format → a firmer commitment

Instead of just "answer first," add a specific format:

❌ Weak:
"Give an answer, then explain."

✅ Stronger:
"First — one sentence with the final conclusion, starting 
with 'I recommend' or 'I don't recommend'. Then — the justification."

The more specific the format of the first answer, the less the model wanders off and the more precise the final conclusion.

🔧 Technique: Post-Reasoning + Chain-of-Thought for hard tasks

The authors studied the method as an alternative to CoT, but the principles can be combined. For especially hard tasks — think first, then structure the conclusion again:

{Hard question}

Step 1: Break the task down step by step.
Step 2: Based on that breakdown, state the final answer 
in one sentence.
Step 3: Justify why exactly this conclusion follows from the steps above.

This is no longer pure Post-Reasoning, but it's the same idea: an explicit conclusion before the justification removes the fuzziness.

Paper: Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost (2025)

Authors: Richmond Sin Jing Xuan, Rishabh Bhardwaj, Soujanya Poria — Nanyang Technological University, Singapore

Contact: [email protected], [email protected], [email protected]