FOR-Prompting: From Objection to Revision via an Asymmetric Prompting Protocol

The paper introduces FOR-Prompting, a model-agnostic, asymmetric prompting protocol that enhances reasoning and iterative refinement across diverse tasks by structuring interactions between a Defender, a Questioner, and an optional Host, enabling even small models to achieve performance comparable to or better than standard baselines without requiring training or access to model internals.

The Gist

Imagine you are trying to solve a tricky puzzle, but you've never seen one like it before. You take a guess, write down your answer, and feel pretty good about it. But then, a friend leans over and says, "Hey, have you considered what happens if the wind blows the pieces away?" or "Did you check if that piece actually fits there?"

They don't give you the answer. They don't even move the pieces for you. They just ask questions that make you pause, rethink, and fix your own work.

That is the core idea behind FOR-Prompting, a new method for teaching Artificial Intelligence (AI) how to think better.

Here is a simple breakdown of how it works, using some everyday analogies:

The Problem: The "Overconfident Student"

Usually, when we ask an AI a question, it acts like a student who is too eager to please. It reads the question, thinks for a second, and immediately writes down an answer. Sometimes it's right, but often it misses hidden details, makes silly mistakes, or forgets to check its own work.

Other methods try to fix this by making the AI talk to itself (like "Chain of Thought"), which is like a student muttering to themselves while solving a math problem. It helps, but the student is still the only one in the room. They might miss their own blind spots.

The Solution: The "Defender, The Debater, and The Host"

FOR-Prompting changes the game by introducing a team of three distinct roles. Think of it like a courtroom drama or a sports team practice:

1. The Defender (The Player):

   • Role: This is the AI that tries to solve the problem. It proposes an answer.
   • Analogy: Imagine a soccer player trying to score a goal. They kick the ball and say, "Goal!"

2. The Debater (The Coach/Referee):

   • Role: This is the special part of FOR-Prompting. The Debater never gives the answer. Instead, they act like a strict coach or a curious friend who only asks tough questions.
   • Analogy: The coach doesn't kick the ball for the player. Instead, they shout: "Wait! Did you check if the goalie is actually there?" or "What if it starts raining?" or "Are you sure that piece fits?"
   • The Magic: Because the Debater only asks questions, the Defender is forced to look at their own work, find the holes, and fix it themselves. This is called "From Objection to Revision."

3. The Host (The Announcer):

   • Role: This role is optional. They listen to the whole conversation between the Defender and the Debater and write down the final, polished answer.
   • Analogy: The announcer who summarizes the game and tells the crowd the final score, making sure everyone understands what happened.

Why is this a big deal?

1. It works even with "small" brains.
The paper tested this on very small, cheap AI models (like a 1-year-old child's brain compared to a genius adult). Usually, these small models make lots of mistakes. But when you put a small model in the "Defender" seat and a slightly smarter model in the "Debater" seat (asking the questions), the small model suddenly gets much smarter!

• Analogy: It's like a young apprentice who can't build a perfect chair alone. But if a master carpenter stands by and just asks, "Is that leg level?" and "Did you measure twice?" the apprentice ends up building a perfect chair. The apprentice does the work; the master just asks the right questions.

2. It saves money and time.
Because the "Debater" (the questioner) doesn't need to be a super-smart, expensive AI, you can use a cheap, small AI to ask the questions and a powerful AI to do the heavy lifting. This is like hiring a cheap intern to ask the questions so you don't have to pay a senior engineer to do the whole job.

3. It handles real-life messiness.
The paper tested this on things like planning a vacation.

• Scenario: You ask an AI to plan a trip to Rio.
• Standard AI: Gives you a nice list of places.
• FOR-Prompting: The Debater asks, "What if it rains on Day 3?" or "What if the train tickets are sold out?" or "Is it safe to walk there at night?"
• Result: The Defender revises the plan to include backup hotels, indoor museums, and safety tips. The final plan is much more realistic and useful.

The Bottom Line

FOR-Prompting is a new way to talk to AI. Instead of just asking for an answer, it sets up a system where the AI is challenged by questions rather than given answers.

It turns the AI from a "know-it-all" who rushes to finish, into a "careful thinker" who double-checks their work because someone is constantly asking, "Are you sure?"

It's a simple, powerful trick that makes AI smarter, more reliable, and much better at handling complex, real-world problems without needing expensive training or human intervention.

Technical Summary

Here is a detailed technical summary of the paper "FOR-Prompting: From Objection to Revision via an Asymmetric Prompting Protocol."

1. Problem Statement

Current Large Language Model (LLM) reasoning protocols, such as Chain-of-Thought (CoT), Tree of Thought (ToT), and Self-Consistency, primarily organize internal deliberation within a single agent. While effective, they lack an explicit mechanism for external questioning that elicits self-revision without providing direct solutions.

Existing multi-agent approaches (e.g., debates, committees, or reviewer-reviser workflows) often introduce external thinking or competing solutions into the reasoning trace. This creates several issues:

• Confounded Evaluation: It becomes difficult to distinguish whether improvements stem from the solver's revised reasoning or the external agent's injected answers.
• Diluted Provenance: Multiple authors shaping the content obscure the single line of accountability.
• Bias: Methods like "Self-Ask" rely on a single model to generate and answer its own follow-ups, potentially leading to self-confirmation bias or overthinking.

The paper posits that the most effective way to improve an answer is often not another answer, but a well-posed question that surfaces gaps, unstated assumptions, or overlooked constraints.

2. Methodology: FOR-Prompting

The authors introduce FOR-Prompting (From Objection to Revision Prompting), an asymmetric, multi-agent protocol designed to automate a Human-In-The-Loop (HITL) questioning process.

Core Roles

The protocol coordinates three distinct roles with strict divisions of responsibility:

1. Defender: Proposes an initial answer and iteratively revises it based on external challenges. The Defender is the sole author of the final solution.
2. Debater (Questioner): Raises question-style objections, clarifying questions, constraint checks, or counterexamples. Crucially, the Debater is forbidden from supplying direct fixes or alternative solutions. Their sole function is to apply "interrogative pressure."
3. Host (Optional): Synthesizes the full interaction history (Defender's revisions and Debater's questions) to produce the final output. This role can be omitted, returning the Defender's final revision directly.

Protocol Flow

1. Initialization: The Defender generates an initial response ($A_0$) to the input question $Q$.
2. Iteration Loop (for $r = 1$ to $N$):
   • The Debater analyzes $A_{r-1}$ and generates a set of questions/objections ($O_r$) without offering solutions.
   • The Defender receives $Q$ and the history of objections ($O_1, \dots, O_r$) and generates a revised answer ($A_r$).
3. Finalization: The Host (if present) synthesizes the final answer $A^*$ from the complete dialogue history.

Key Design Principles

• Asymmetry: The roles are not symmetric; the Debater only questions, and the Defender only reasons/revises.
• Question-Only Intervention: External input is restricted to questions, preventing the "answer replacement" problem found in other multi-agent systems.
• Model Agnostic: The protocol works with any LLM and does not require model internals access or fine-tuning.

3. Key Contributions

1. Conceptual Novelty: Formalizes questioning (rather than answer substitution) as the exclusive form of external intervention. This preserves a single accountable reasoning chain while leveraging external pressure.
2. Protocol Design: A lightweight, role-based interaction loop that enforces "objections as questions only," enabling systematic study of questioning as a reasoning mechanism.
3. Empirical Validation: Extensive testing on both commercial (GPT-4o) and open-source (LLaMA-3.2-1B) models, demonstrating efficacy in structured math, error correction, and open-ended planning.

4. Experimental Results

A. Structured Math Tasks (GSM8K)

• Large Models (GPT-4o): FOR-Prompting achieved 0.94 accuracy, matching Chain-of-Thought (CoT) and Self-Ask (0.94) and approaching Self-Consistency (0.95). It consistently outperformed the single-prompt baseline (0.92).
• Small Models (LLaMA-3.2-1B):
  • Single-prompt baseline: 7%.
  • CoT: 23%.
  • FOR-Prompting (Host-False): 23% (matching CoT).
  • FOR-Prompting (Host-True): 19% (slightly lower due to synthesis errors in small models).
  • Insight: Even with severe capacity constraints, the external questioning mechanism doubled the accuracy of the baseline, proving that small models can effectively act as "Questioners" while stronger models act as "Defenders."

B. Cross-Model Role Swapping

• Strong Defender / Weak Debater: GPT-4o (Defender) + LLaMA-1B (Debater) achieved 93.2% accuracy.
• Weak Defender / Strong Debater: LLaMA-1B (Defender) + GPT-4o (Debater) dropped to 21% accuracy.
• Conclusion: Performance is primarily determined by the Defender's capability. The Debater role does not require high model capacity to be effective, offering a cost-efficient strategy for hybrid pipelines.

C. Error Correction & Open-Ended Tasks

• Error Correction: In a "count the 'r's in 'strarrtrabbbery'" task, a single prompt failed (4 'r's). FOR-Prompting corrected this to the right answer (5 'r's) through iterative questioning.
• Itinerary Planning: In a 5-day Rio de Janeiro itinerary task, FOR-Prompting generated more robust, safety-conscious, and actionable plans than baselines.
• Human Preference Study: In a blind study with 77 participants regarding a 7-day Barcelona itinerary, 74% preferred the FOR-Prompting output over outputs from strong frontier LLMs (GPT-4.1 and another top-tier model), citing better completeness, realistic pacing, and contingency planning.

5. Significance and Implications

• Scalability for Low-Resource Settings: FOR-Prompting allows small, on-device models to achieve reasoning performance comparable to larger models by offloading the "questioning" role to a lightweight agent or even a smaller model, while the "reasoning" is handled by a capable model. This reduces token costs and enables privacy-preserving local inference.
• Interpretability and Accountability: By keeping the reasoning chain within a single agent (the Defender) and using questions as the only external input, the protocol maintains a clear audit trail of how an answer was refined, avoiding the "black box" nature of multi-agent consensus or debate.
• Automation of HITL: It automates the human practice of refining outputs through questioning rather than direct editing, making iterative refinement scalable and reproducible without human intervention.
• Applicability: The method is particularly valuable for tasks requiring progressive elaboration, such as complex planning, strategy design, and multi-stage tasks where new constraints emerge dynamically.

In summary, FOR-Prompting demonstrates that structured external questioning is a powerful, low-cost mechanism for improving LLM reasoning, offering a practical alternative to complex multi-agent debate or expensive model scaling.