TL;DR
Mixture-of-Agents (MoA) is a technique where several "experts" (the same LLM in different roles) generate answers independently, and then the model synthesizes them into a final output. The key principle: there are always more parallel generations than synthesis rounds — the optimum is reached with N experts and N−1 aggregation layers.
The most important finding: "Improve your answer" — a popular trick — actually hurts quality. Self-refinement performed worse than a plain CoT request at every compute budget. A model that iteratively edits itself doesn't get more accurate — it drifts and degrades. Users who type "now make it better" three times in a row burn tokens and end up with something worse than the first draft.
MoA solves this differently: instead of iteratively editing one answer — several independent attempts, each with no memory of the others, and a single final synthesis. This removes accumulated errors and gives the model a "fresh pair of eyes" on every generation. In the study, MoA added +7.1 percentage points of accuracy over baseline CoT and +2.7 pp over self-consistency at the same budget.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1 (in parallel, in one prompt):
Expert A → independent answer
Expert B → independent answer
Expert C → independent answer
[N experts — no references to each other]
STEP 2 (synthesis, once):
Aggregator → reads all answers → final output
Rule: N experts > N synthesis rounds by 1
Optimum: 3 experts + 1 round | 4 experts + 1 round | 5 experts + 1 round
Both steps run in a single prompt — the model generates both the experts and the synthesizer itself.
Example in Action
Task: Alex is launching his own YouTube channel about investing and can't decide whether to make short videos for Shorts or long deep-dives. It's a tricky call: he has to factor in the algorithm, the audience, monetization and his own bandwidth.
Prompt:
I'm launching a YouTube channel about investing for beginners in the US.
The key strategy question: focus on Shorts (under 60 seconds)
or long deep-dives (15–30 minutes)?
Step 1. Three experts give independent answers — each
from their own role, without considering the others' opinions:
Expert A (YouTube algorithms): [a detailed recommendation
strictly from the standpoint of reach and promotion]
Expert B (monetization and business model): [a detailed
recommendation strictly from the standpoint of revenue and channel growth]
Expert C (audience behavior in the investing niche):
[a detailed recommendation strictly from the standpoint of
what people watch and subscribe to]
Step 2. Synthesizer: read the three answers above, find
the contradictions, weigh the arguments and give a final
recommendation with a strategy for the first 3 months.
Result: The model will generate three independent expert positions in sequence — each emphasizing its own angle without looking over its shoulder at the others. In the final block, the synthesizer will compare all three, explicitly flag where the experts disagree and why, and give a balanced recommendation. This is fundamentally different from a single "balanced" answer — the contradictions between experts are visible and explained.
Why It Works
The LLM's weakness: When a model "improves" its previous answer, it works like an editor with tunnel vision. It has already "seen" the first version — and the new answer unconsciously gets built around it. Errors in the first answer don't get fixed; they carry over into the second and third. The study shows this isn't a hypothesis — it's a measurable decline.
The LLM's strength: The model is good at looking at one problem from different angles at the same time. When each "expert" is generated with no memory of the others, every attempt is a fresh one. Errors in one branch don't infect the rest.
How MoA uses this: The parallel generations are isolated — no expert sees the others. The final synthesis receives N independent signals and picks the best or weighs them. Control levers:
- Number of experts → more for complex tasks, 2–3 for simple ones. Beyond 5, returns diminish
- Expert roles → specific ones (financier, marketer, skeptic) cut sharper than a faceless "expert 1"
- Synthesize once → a second aggregation round isn't needed and often hurts
- Task difficulty → on easy questions (a capital, a fact, a formula) MoA is overkill. It shines on tasks with real uncertainty
Prompt Template
Task: {task description}
Step 1. Three experts answer independently — each strictly
from their own role, without considering the others' opinions:
Expert A ({role A}): [answer strictly from the standpoint of {angle A}]
Expert B ({role B}): [answer strictly from the standpoint of {angle B}]
Expert C ({role C}): [answer strictly from the standpoint of {angle C}]
Step 2. Synthesis: read the three answers above. Find
the contradictions and points of agreement. Give a final conclusion
with reasoning — what matters more and why.
What to fill in:
- {task description} — a specific question or situation
- {role A/B/C} — specific professional roles suited to the task: financier, lawyer, marketer, engineer, skeptic, end user
- {angle A/B/C} — which aspect of the task this expert looks at
You can cut the number of experts to two on simple tasks. Going beyond four isn't worth it.
🚀 Quick start — paste into the chat:
Here is a template for the Mixture-of-Agents technique. Adapt it to my task: {your task}.
Ask clarifying questions to fill in the fields.
[paste the template above]
The LLM will ask about your task, choose roles and angles for the experts, and fill in the template — because for the method to work, you need a clear picture of which three independent perspectives give the widest coverage of your specific task.
Limitations
⚠️ Easy tasks: If the answer is obvious, three experts will say roughly the same thing. The method is overkill. Use MoA only when there's real uncertainty.
⚠️ Self-critique doesn't work: Don't add "now each expert critiques the others" to the template — the study shows that iterative rounds with memory of previous answers make results worse. Synthesis happens once, at the end.
⚠️ More ≠ better: The optimum is N experts and N−1 synthesis rounds. Adding one more aggregation layer reduced accuracy in almost every configuration.
⚠️ Tasks with a single right answer: The method is built for tasks with uncertainty. On questions like "what is 2+2" it's pointless.
Resources
Paper: "Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scaling"
Authors: Florian Valentin Wunderlich, Lars Benedikt Kaesberg, Jan Philip Wahle, Terry Ruas, Bela Gipp
Institution: University of Göttingen, Germany
Contact: [email protected]
Code: https://github.com/Multi-Agent-LLMs/lm-evaluation-harness
Benchmarks: MMLU-Pro, BBH
Related work: Du et al. 2023 (debate), Wang et al. 2024a (MoA), Madaan et al. 2023 (self-refinement), Wang et al. 2023 (self-consistency)