Skip to main content

Adversarial Self-Critique: Improving Through Opposition

Good ideas are forged in the crucible of debate. By prompting an LLM to have a structured argument with itself, we can simulate this powerful process, leading to outputs that are not just generated, but battle-tested.

Introduction​

In the last article, we taught the model to be skeptical of its own ideas. Now, we will take this concept and give it a formal structure. Adversarial Self-Critique is a sophisticated technique where you prompt the LLM to adopt two or more distinct, opposing personas and have them critique each other's work in a structured debate.

This goes beyond simple self-skepticism. Instead of just asking the model to find flaws in its own work, we are creating a multi-agent system within a single prompt. We set up a "proposer" who generates ideas and a "critic" whose sole job is to find fault with those ideas. This adversarial dynamic forces a much deeper and more rigorous exploration of the problem space.

The Core Idea: A Debate in a Box​

The power of adversarial critique lies in the cognitive separation of roles. It's psychologically easier to critique someone else's work than your own. By assigning the roles of "proposer" and "critic" to different personas, we free the LLM from the cognitive bias of having to defend its own initial thoughts.

The typical workflow for Adversarial Self-Critique is as follows:

  1. Proposal Generation: The "proposer" persona is prompted to generate an initial plan, idea, or piece of text.
  2. Adversarial Critique: The "critic" persona is prompted to review the proposal and provide a detailed, harsh, but fair critique. It should identify specific weaknesses, logical fallacies, or potential risks.
  3. Revision and Rebuttal: The "proposer" persona is given the critique and prompted to revise its original proposal to address the points raised by the critic. It can also rebut points it disagrees with, providing a justification.
  4. Iterative Refinement: This cycle of critique and revision can be repeated multiple times, with each round producing a more refined and robust output.
  5. Final Synthesis: Finally, a "judge" or "synthesizer" persona can be prompted to review the entire debate and produce a final, balanced conclusion.

A Practical Example: Refining a Policy Proposal​

Problem: You want to generate a well-thought-out policy proposal for reducing traffic congestion in a city.

Step 1: The Proposer​

Prompt:

You are an optimistic urban planner, "Proposer Pat." Your goal is to propose a bold and innovative solution to reduce traffic congestion.

**Proposal:**
My proposal is to implement a "Congestion Pricing" system, where drivers are charged a fee to enter the downtown core during peak hours. This will create a strong financial incentive to use public transport, carpool, or travel at off-peak times. The revenue generated can be invested in improving the subway and bus systems.

Step 2: The Critic​

Prompt:

You are a skeptical and pragmatic city comptroller, "Critic Chris." Your job is to find the flaws in Proposer Pat's plan. Please provide a harsh but fair critique of the "Congestion Pricing" proposal.

**Critique:**
Proposer Pat's plan is a regressive tax that will disproportionately harm low-income workers who have no choice but to drive into the city for their jobs. The political capital required to implement this would be immense, and public backlash is almost certain. Furthermore, there is no guarantee that the revenue will actually be used to improve public transport; it could easily be absorbed into the city's general fund. The plan also fails to account for the economic impact on downtown businesses.

Step 3: Revision and Rebuttal​

Prompt:

You are Proposer Pat. You have received the following critique from Critic Chris. Please revise your proposal to address his concerns.

**Revised Proposal:**
I acknowledge the valid concerns raised by Critic Chris. To address these, I propose a revised "Congestion Pricing and Equity" plan.
1.  **Equity Program:** Low-income residents who can demonstrate need will be granted a full exemption from the congestion fee.
2.  **Revenue Lockbox:** All revenue generated will be placed in a legally protected "transportation lockbox" that can only be used for specific, publicly-audited improvements to the public transit system.
3.  **Business Support:** We will launch a "Shop Downtown" campaign and offer tax credits to small businesses within the congestion zone to mitigate any negative economic impact.

This revised proposal is vastly superior to the original. It is more nuanced, practical, and politically viable, all thanks to the structured adversarial process.

When to Use Adversarial Self-Critique​

This is an advanced and resource-intensive technique. It is best reserved for:

  • High-stakes decision-making: When you need an extremely robust and well-vetted plan.
  • Complex policy or strategy questions.
  • Creative writing, to refine a plot or character arc.
  • Any task where "groupthink" or a single point of view is a significant risk.

Key Takeaways​

  • Adversarial Self-Critique involves a structured debate between two or more opposing AI personas.
  • It separates the role of idea generation ("proposer") from idea evaluation ("critic"), reducing cognitive bias.
  • The iterative cycle of critique and revision leads to highly robust and well-vetted outputs.
  • This is an advanced technique best suited for complex, high-stakes tasks where nuance and rigor are paramount.

What's Next?​

Adversarial critique is excellent for refining a single idea. But what happens when the model generates multiple, conflicting pieces of information? How can it resolve contradictions in its own knowledge base? In the next article, we will explore Iterative Contradiction Resolution, a technique for prompting a model to identify and resolve conflicts in its own outputs.

By mastering the art of the AI debate, you can ensure that your final outputs have survived the trial by fire, emerging stronger, smarter, and more resilient.