PR-CoT (Poly-Reflective Chain-of-Thought): Four-Lens Self-Review Instead of One Vague Reflection

STEP 1: Initial answer
Prompt: [task] + "Solve step by step" → CoT and initial answer

STEP 2: Four specialized reflections
  PERSPECTIVE 1: Logical integrity → finds contradictions, leaps in logic
  PERSPECTIVE 2: Completeness of information → finds missing factors, data
  PERSPECTIVE 3: Ethics and biases → finds unjustified assumptions, unfairness
  PERSPECTIVE 4: Alternative solutions → finds other paths that weren't considered

STEP 3: Synthesis
Prompt: [task] + [initial CoT] + [all 4 critiques] → Final improved answer

You can run it: - As one prompt: all steps in sequence inside a single request - Separately: each step as its own message (more control, but requires copy-pasting)

Task: You're launching an educational project on copywriting. There are two strategies: (A) a mass-market "Copywriting from Scratch" course at $299 for beginners, (B) an elite mastermind at $1,499 for experienced copywriters. There's one budget for the first launch, so you have to pick one strategy.

Prompt:

Task: I'm launching an online copywriting school. Two options for the first product:
A) A "Copywriting from Scratch" course at $299, for beginners
B) A mastermind at $1,499, for experienced copywriters

There's one launch budget. I need to pick ONE option.

STEP 1: Analyze the situation and propose a solution. Show your reasoning step by step.

STEP 2: Check your solution from four angles:

PERSPECTIVE 1 — Logical integrity:
Are there contradictions in the reasoning? Does every step follow logically from the previous one? 
Am I using circular logic?

PERSPECTIVE 2 — Completeness of information:
What important factors might I have missed? What data, context, or constraints did I overlook? 
Are all assumptions stated explicitly?

PERSPECTIVE 3 — Ethics and biases:
Are there unjustified assumptions or stereotypes? Is the decision unfair in any way? 
Did I account for ethical consequences?

PERSPECTIVE 4 — Alternative solutions:
What other options exist? Why didn't I consider them? 
Is there a better path?

STEP 3: Based on all the checks, give a final decision that accounts for the problems found.

Result:

The model will show the full process: an initial decision with a rationale (it'll likely pick A or B with some market logic) → then 4 blocks of critique:

• Perspective 1 will find logical inconsistencies (if there's a conflict between "we need fast money" and "long-term reputation")
• Perspective 2 will surface forgotten data (market competition, an existing audience, customer acquisition costs)
• Perspective 3 will check the assumptions (is "beginners are willing to pay $299" actually justified?)
• Perspective 4 will propose other paths (maybe a phased launch? or a test sale of both formats?)

The final synthesis won't be just "pick A" or "pick B" but a multi-factor analysis that accounts for the blind spots of the initial reasoning. You get a decision that has gone through a systematic review from every side.

The LLM weakness: The model is good at generating coherent reasoning but bad at noticing specific types of errors inside it. A "check yourself" request is too vague: the model doesn't know WHAT exactly to look for. It might find an obvious logical hole but miss a data gap or an ethical issue. Different kinds of errors need different review frameworks, and without a clear focus the model checks only superficially.

The LLM strength: The model follows structured instructions really well and can analyze text (including its own) against given criteria. Give it a clear framework ("check the logic specifically," "check the data completeness specifically") and it focuses and finds problems it would have missed in a generic reflection. It's like handing an expert a checklist instead of a vague "assess the quality".

How the method works: PR-CoT uses decomposed reflection. Instead of one fuzzy "check yourself," it gives four narrow checks, each with its own focus. The model makes 4 specialized passes, each time through a new "lens." The insights accumulated across all four perspectives give a 3D picture of the initial answer's weak spots. At the end, the model doesn't just "fix the bug"; it synthesizes a new solution that accounts for every nuance uncovered. It's a rework, not a patch.

Control knobs:

Number and mix of perspectives: the original uses 4 standard ones. You can drop ethics for purely logical tasks → save tokens. Or add a fifth: "Practical feasibility" for business tasks, "Performance" for code. Tailor the perspectives to your domain.

Order of perspectives: Logic → Data → Ethics → Alternatives isn't random. You can reorder: Alternatives first (to widen the solution space), then the rest. Order matters for how the model synthesizes: the last perspective is "fresher" in context.

Depth of each perspective: the template uses short questions. You can expand them: "Find ALL logical inconsistencies and rate the severity of each from 1 to 10" → deeper review, more tokens, more detail.

Critique output format: you can ask for a numeric score after each perspective, "how critical is the problem found: 1-10" → this adds prioritization during synthesis. Or ask for concrete edits instead of general remarks.

One-shot vs. multi-turn: you can pack everything into one prompt (saves time) or run each perspective as a separate message → more control, and you can adjust the next perspective based on the previous critique, but it requires copy-pasting.

Task: {describe the task or dilemma}

STEP 1: Analyze the situation and propose a solution. Show your reasoning step by step.

STEP 2: Check your solution from four angles:

PERSPECTIVE 1 — Logical integrity:
Are there contradictions in the reasoning? Does every step follow logically from the previous one? 
Am I using circular logic or unjustified leaps?

PERSPECTIVE 2 — Completeness of information:
What important factors might I have missed? What data, context, or constraints did I overlook? 
Are all the necessary assumptions stated explicitly?

PERSPECTIVE 3 — Ethics and biases:
Are there unjustified assumptions or stereotypes? Is the decision unfair in any way? 
Did I account for ethical consequences?

PERSPECTIVE 4 — Alternative solutions:
What other solutions exist? Why didn't I consider them? 
Is there a better path I missed?

STEP 3: Based on all four checks, give a final decision that accounts for the problems found.

What to fill in: - {describe the task or dilemma}: your specific situation, whether a business decision, career choice, technical architecture, or ethical dilemma. The more complex and multi-factor it is, the more you get out of the method.

Adapting to the task type:

For technical tasks (code, architecture), replace Perspective 3:

PERSPECTIVE 3 — Performance and scalability:
Are there bottlenecks? How will the solution behave under load? 
Did I account for scaling requirements?

For creative tasks (copy, design, content), replace it with:

PERSPECTIVE 3 — Target audience and context:
Does the solution fit the target audience? Did I account for the context of use? 
Is there a gap between the idea and the customer's reality?

For personal decisions (career, relocation, purchases), leave it as is: ethics and biases matter and help catch emotional distortions.

🚀 Quick start: paste into the chat:

Here is a PR-CoT template for multi-level decision review. Adapt it to my task: [your task]. 
Ask clarifying questions so you can fill in the details correctly.

[paste the template above]

The LLM will ask about the details of your situation and context, which it needs so each perspective checks the relevant aspects. For example: - For a business decision it'll ask about budget and audience (for Perspective 2) - For an ethical dilemma, about the affected parties (for Perspective 3) - For a technical task, about performance requirements

It will take the pattern from the template (the structure of 4 perspectives + synthesis) and fill it with specific criteria for your task.

⚠️ Token cost: The method requires 4 rounds of reflection + synthesis, roughly 3-4 times more tokens than plain CoT. For simple tasks ("what's 2+2," "fix this typo") it's overkill and expensive. Use it only for complex multi-factor tasks where errors are costly.

⚠️ Overkill on simple tasks: If the task is straightforward ("capital of France," "basic arithmetic"), the extra perspectives add nothing. The initial CoT is already right, and four checks just spin their wheels. The method shines on tasks with trade-offs and non-obvious factors.

⚠️ Hallucinated critique: The model may "invent" problems where there are none, especially in Perspective 3 (ethics). If the initial decision is good, excessive critique can make it worse ("but what if this offends someone?"). Use common sense when reading the final synthesis; not every remark is justified.

⚠️ Conflicting perspectives: Different angles can give contradictory recommendations (logic says A, ethics says B, alternatives say C). Step 3 (synthesis) doesn't always resolve the conflict well and may water the decision down to "it depends on your priorities" instead of a clear choice.

A team from the University of Brasilia took GPT-3.5 and GPT-4 and ran them through four task types: arithmetic, commonsense reasoning, ethical dilemmas, and logic puzzles. They compared three approaches: plain CoT ("solve step by step"), Multiplex CoT with one generic reflection ("check yourself"), and their own PR-CoT with four specialized perspectives.

They measured two metrics: logical consistency (the percentage of solutions with no internal contradictions) and error correction rate (how many initially wrong answers became correct after reflection). That matters more than plain accuracy: it shows not "did it guess right" but "did it fix its own mistakes."

The results showed a clear pattern: PR-CoT beat single reflection by 2-3% on simple tasks (arithmetic: 94% vs. 92% consistency, commonsense: 87% vs. 85%), but pulled sharply ahead on complex ones. On ethical dilemmas: 84% consistency vs. 81% for Multiplex CoT and 74% for plain CoT, a 10% gap! On logic puzzles: 93% vs. 90% and 82%.

The error correction rate was highest on complex tasks: 21% in ethics (vs. 18% for MCoT), 23% in puzzles (vs. 20%). In simple arithmetic the difference was minimal: 17% vs. 15%.

Why this result? Simple tasks have one clear path to the answer; there, the initial CoT is already close to optimal and the extra perspectives find little (there's no room to maneuver). But complex tasks like ethics and puzzles involve many factors and trade-offs. There, a single reflection catches only some of the problems (usually the logical ones, the most obvious), while PR-CoT systematically walks through every dimension: logic, data, values, alternatives. It's exactly in these complex domains that multiple perspectives deliver a leap.

They also ran a human evaluation: experts rated reasoning quality on a 1-5 scale for 100 examples from ethics and puzzles. PR-CoT scored 4.4 for logical coherence vs. 4.0 for MCoT and 3.4 for CoT. For completeness of argumentation: 4.3 vs. 3.8 and 3.1. People see a quality difference even where the accuracy metrics are close, which means it's not just about the right answer but about the quality of the path to it.

The practical takeaway: The method doesn't work equally well everywhere. If the task is straightforward (calculations, factual questions), save your tokens and skip the four perspectives. If the task is complex, with non-obvious trade-offs and ethical nuances, PR-CoT delivers the biggest gains exactly there. It's not a universal "make it better" button; it's a tool for multi-factor decisions.

"Enhancing Self-Correction in Large Language Models through Multi-Perspective Reflection"
Mariana Costa, Alberlucia Rafael Soarez, Daniel Kim, Camila Ferreira
University of Brasilia