SC-MoA (Self-Consistent Mixture of Agents): Synthesize the Reasoning Instead of Voting on the Answer

Phase 1 — Diversity (1 request or N separate ones)
  Question → REPHRASE into 5 semantically identical variants
  Each variant → full chain of reasoning + answer

Phase 2 — Anchored refinement (N−k requests)
  Find the majority (which answer shows up most often)
  Versions with a different answer → REFINE them using the majority chain
  The majority stays frozen

Phase 3 — Synthesis (1 request — always, even when unanimous)
  All N reasoning chains → SYNTHESIZE into a single answer
  Explicitly pull out the unique intermediate steps from each version

You can run all three phases one after another in a single chat or fold them into one big prompt.

The task: You run a small marketing agency and you're wondering whether to launch a SaaS product built on your own methodology. It's a tough call: there are legal, financial, operational and market angles, and each one needs its own chain of reasoning.

Prompt (one request, all three phases):

I need the highest-quality analysis possible. Use SC-MoA: 
synthesize reasoning instead of giving a single answer.

**Question:** Should a small marketing agency (8 people, 
Chicago, B2B segment, $1.5M annual revenue) launch a SaaS product 
based on its content methodology?

**Step 1 — Reframes.**
Generate 5 semantically identical versions of this question. 
Each one should push toward a different angle of analysis 
(financial, market, operational, risk, strategic). 
Show all 5.

**Step 2 — Reasoning chains.**
For each of the 5 versions — a detailed chain of reasoning 
(at least 5 steps) and a final conclusion (yes / no / conditional).

**Step 3 — Anchored refinement.**
Find the majority: which conclusion shows up most often?
For versions with a different conclusion — adjust the reasoning 
using the strongest arguments of the majority. 
Keep the unique intermediate steps the others don't have.

**Step 4 — Synthesis (always, even if everyone agrees).**
Read ALL 5 reasoning chains. 
Assemble the final analysis, explicitly pulling unique insights 
from each chain — especially the ones the majority lacks. 
Don't vote on the answer — integrate the best intermediate steps.

The result: The model shows 5 rewordings of the question — each one stressing a different facet of the problem. Then five detailed lines of reasoning with intermediate steps and conclusions. After that come the refined "minority" versions with the majority's arguments stitched in. And at the end — a synthesized analysis that says explicitly: "this argument comes from the financial chain, that one from the risk chain," building a picture that no single version delivered on its own.

The LLM weakness: focus mode. Modern models are trained to predict the next token — and on top of that they're drilled to produce the most probable chain of reasoning. One wording of the question = one track of thinking. If that track has a blind spot, asking again won't fix it — it'll just cement it.

The key paradox. Even when all versions arrive at the same answer, synthesis still improves the result. This is the "aggregation paradox": the correct intermediate steps often live in the minority of versions that voting throws away. A synthesizer model reading all the chains finds those steps and integrates them. Voting knows nothing about those steps — it only deals with final answers.

The LLM strength: pattern matching across text. Models are great at reading several texts and spotting complementary arguments in them. That's exactly what synthesis does: instead of picking a "winner," it assembles a mosaic out of the correct pieces of different chains. Changing the wording of the question is the cheapest way to trigger different thinking paths. The content of the rephrase doesn't matter — what matters is that a surface change in form triggers structurally different reasoning.

Control knobs: - Number of reframes (N): 3 for quick decisions, 5 as the standard, 7+ for the hardest problems. Bigger N = more tokens, but a richer synthesis. - Reframe angles: set them explicitly (financial, legal, user-facing) → you control which aspects are guaranteed to get covered. - Anchored refinement: you can skip it on simple tasks and go straight to synthesis. - Synthesis instruction: add "note which version each key argument came from" → you get a transparent audit trail of the reasoning.

Use SC-MoA for the analysis: synthesize the reasoning instead of voting on the answer.

**Task:** {your question or task}

**Step 1 — Reframes.**
Generate {number_of_versions} semantically identical rewordings of this question.
Each one should push toward a different angle of analysis: {angle_1}, {angle_2}, {angle_3}.
Show all versions.

**Step 2 — Reasoning chains.**
For each version — a detailed chain of reasoning (at least {number_of_steps} steps)
and a final conclusion.

**Step 3 — Anchored refinement.**
Find the majority: which final conclusion shows up most often?
For versions with a different conclusion — adjust them using the majority's arguments.
Keep the unique intermediate steps the others don't have.

**Step 4 — Synthesis (always do this, even when unanimous).**
Read ALL {number_of_versions} reasoning chains.
Assemble the final answer, explicitly pulling unique insights from each chain —
especially the ones the majority lacks.
Don't vote on the answer — integrate the best intermediate steps.

What to plug in: - {your question} — a complex problem with several angles: a business decision, a technical choice, a situation analysis - {number_of_versions} — 3 (fast) or 5 (deep) - {angle_1–3} — set the perspectives explicitly: financial / risk / user / operational / strategic - {number_of_steps} — 4–6 for most tasks

🚀 Quick start — paste into the chat:

Here is an SC-MoA template. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask which angles of analysis matter and how deep a reasoning chain you want — because those are the parameters that decide how diverse the reframes turn out and how rich the final synthesis will be.

⚠️ Not for simple questions: If the task is clear-cut ("translate a word," "what's 2+2"), the method just burns tokens for nothing. SC-MoA is for cases where different entry points lead to different routes of thinking.

⚠️ Token cost: The full pipeline in the study takes 11–21 requests/calls. Run by hand in a chat it's one long prompt with 4 steps, but the response will be big.

⚠️ A weak synthesizer wipes out the gains: Ask a weak model to synthesize and it may just pick the majority instead of integrating. The instruction "state explicitly what was taken from each version" lowers that risk.

⚠️ Subjective tasks — not as good: The method works best where there's a correct answer or objective quality criteria. For "write something beautiful" it's less useful.

⚠️ Several different models don't add an edge: The study showed that a single model with rephrasings beats a pool of different models without this technique. Paying for a model "zoo" is pointless.

The Bioscope AI team asked a simple but non-obvious question: what if, instead of reading the agents' final answers, you read their full reasoning? They took one model (gpt-oss-120b), ran it through five rewordings of the same question with strictly deterministic generation (temperature 0, so randomness wouldn't muddy the results), and then compared two aggregation options on the exact same data: voting on the answer vs. synthesizing the full chains.

The surprising finding: synthesis beat voting even in cases where the answer was unanimously wrong. On PhD-level chemistry and biology problems (GPQA-Diamond), synthesis fixed errors almost 2.5 times more often than it introduced new ones. On structured reasoning tasks (BBH-3) — 3.2 times more often. The researchers tested the obvious hypothesis: maybe rephrasings just produce uncorrelated errors? Nope — error correlations were identical for diverse and identical prompts. The real reason is that minority chains contain correct intermediate steps that the majority skipped.

One more unexpected result: the content of the rephrase doesn't matter. Hand-written "personas," automatic paraphrases from GPT and structural linguistic variations produced statistically indistinguishable results. That's good news for users — no need to build elaborate prompt systems; just ask the model to reword the question five different ways itself.

🔧 A lighter version without anchored refinement — for quick tasks

If the task is medium-hard and you don't have time for the full pipeline, skip Phase 2. Just ask for 3 reframes with reasoning, then a synthesis. You'll lose ~10–15% of the quality but get your answer twice as fast.

Answer the question three ways — from different angles: {angle_1}, {angle_2}, {angle_3}.
For each: a full chain of reasoning + a conclusion.
Then synthesize all three — pull the best arguments from each chain.

Question: {your question}

🔧 Explicit names instead of faceless "versions" → sharper role-play

Instead of "Version 1, Version 2," give each reframe an expert's name or a role:

Analyze the question from four positions:
- David Carter (venture investor, looking for scale)
- Laura Bennett (financial regulator, looking for risks)  
- Emma, a user who runs an Amazon storefront (looking for practicality)
- The Skeptic (looking for why it won't work)

For each — a full chain of reasoning. Then a synthesis of all four.

Roles force the model to genuinely change its entry point into the problem instead of just rewording things on the surface.

🔧 Make the synthesis auditable → see where every claim came from

Add this to the synthesis instruction:

In the final analysis, for each key claim note in parentheses: 
(source: Version 2, risk reasoning)

You get not just a conclusion but a provenance map for every argument — handy for auditing decisions or explaining the logic to other people.

Beyond Consensus: Trace-Level Synthesis in Mixture of Agents Shreyas Fadnavis, Praitayini Kanakaraj, Felix Wyss — Bioscope AI [email protected]

Related work: Wang et al. 2023 (Self-Consistency), Wang et al. 2025 (Mixture of Agents), Gao et al. 2024 (SPUQ — semantics-preserving paraphrases), Choi et al. 2025 (debate as a martingale)