FusioN (Fusion-of-N): Synthesis Instead of Selection, a New Way to Aggregate LLM Answers

STEP 1: Build the candidate pool
├─ Test-time: N samples from one model (temperature > 0)
└─ Synthetic data: 1 sample each from N different teachers

STEP 2: Synthesis (in a single prompt)
└─ Fusor receives: {prompt} + {all candidates}
    → Analyzes the strengths/weaknesses of each
    → Writes a new answer that combines the best parts

Important: Both steps can be done in a regular chat: first generate the variants, then hand them to the fusor for synthesis.

Task: You are the founder of an EdTech startup, an online SAT prep platform with AI tutors. You need a pitch for your pre-seed round. Instead of one version of the pitch, you generate 3 different angles: (1) market and metrics, (2) technology and product, (3) team and vision. Then you synthesize a final pitch that takes the strongest parts of each.

Prompt to generate the pool:

Write a pre-seed pitch for the startup "SAT.AI", an SAT prep platform with AI tutors.

Variant 1: Focus on market size, unit economics and growth metrics.
Variant 2: Focus on what makes the technology unique, the product and the UX.
Variant 3: Focus on the team, the vision and the long-term mission.

Each pitch should be 200-250 words.

Synthesis prompt (FusioN):

I have 3 versions of an investor pitch. Your job is to create a final pitch that combines the strengths of each version.

Original request: {paste the SAT.AI task}

Variant 1 (market-driven):
{paste the text of variant 1}

Variant 2 (product-driven):
{paste the text of variant 2}

Variant 3 (vision-driven):
{paste the text of variant 3}

Analyze what each variant does well and what could be improved. Then write a final pitch (~250 words) that:
- Takes the most persuasive arguments from each variant
- Drops weak or repetitive points
- Builds a coherent narrative that works better than any of the three variants on its own

The final pitch must read as one cohesive text, not a patchwork of fragments.

Result:

The fusor will analyze all three pitches and pull out:

• From variant 1: concrete market numbers (N million high school students, average order value, CAC/LTV)
• From variant 2: a description of the unique technology (ML-driven personalization, adaptive practice problems)
• From variant 3: a strong vision (democratizing high-quality education)

The final synthesized pitch weaves these elements into one coherent text that is stronger than any of the originals. Instead of three lopsided pitches you get one balanced pitch that sells the market, the product and the vision.

The LLM weakness: When generating a long answer to a complex prompt, a model cannot cover every aspect equally well. It makes trade-offs, focusing on some details at the expense of others. As a result, different samples (at temperature > 0) cover different facets of the problem. That is exactly why Best-of-N works at all: some variants turn out better than others.

The LLM strength: Modern large models are excellent at comparative analysis: they can evaluate several texts, list the pros and cons of each and see where one text beats another. This is a core capability of modern LLMs as judges.

How the method leverages that strength: FusioN gives the fusor a task of synthesis (polylithic evaluation) instead of selection (monolithic quality scoring). The fusor sees all variants at once, picks the best fragments of each (not whole answers) and writes a new text. Because quality is polylithic (different parts of different answers are good), the synthesis can beat any original variant.

Control knobs:

• N (number of candidates): 2-5 for simple tasks, 5-10 for complex ones. More = pricier, but more diverse.
• Sampling temperature: 0.7 for balance, higher for creative tasks, lower for factual ones.
• Instructions to the fusor: You can ask it to "keep the style of variant 2" or "prioritize accuracy over creativity".
• Output format: "No explanations" for a clean result, or "Show your reasoning" to see how the fusor made its choices.

I have {N} candidate answers to a request. Your job is to create a final answer that combines the strengths of each candidate.

Original request: {task description}

Variant 1:
{text of variant 1}

Variant 2:
{text of variant 2}

[... remaining variants]

Variant N:
{text of variant N}

Analyze what each variant does well and what could be improved. Then write a final answer that:
- Takes the highest-quality elements from each variant
- Drops weak, repetitive or incorrect points
- Produces a coherent result that works better than any single variant on its own

{extra format/style requirements, if any}

How to fill it in:

• {N}: the number of variants (usually 3-5)
• {task description}: the original prompt the variants were generated for
• {text of variant X}: each generated variant in full
• {extra requirements}: for example, "The final answer must be no longer than 300 words" or "Keep the formal tone of variant 1"

⚠️ Fusor size matters: Small models (4B-8B) are bad at synthesis. It works well with models from 27B up, even better at 111B+. For test-time scaling with a small model, stick with classic Best-of-N.

⚠️ Not for every task: On strictly formal tasks (math with a single correct answer) FusioN shows mixed results. It shines on open-ended tasks where there is room to combine different approaches: writing, analysis, translation, reasoning.

⚠️ Pool quality matters: If all variants are weak or nearly identical, synthesis won't help. The method pays off when there is diversity in the pool. For synthetic data, use different teachers; for test-time scaling, sample at temperature > 0.

⚠️ Language imbalance: The fusor may do worse in languages it supports poorly. The study saw the best results in English. For other languages, make sure the fusor is strong enough in that language.

A team from Cohere Labs ran a large-scale evaluation of FusioN vs Best-of-N in two main scenarios.

Test-time scaling: They took two models, Aya Expanse 8B and Command A 111B. For each prompt they generated N=5 variants at temperature 0.7. For Best-of-N, a strong reward model picked the winner. For FusioN, Command A served as the fusor. They tested 11 languages across three tasks: Arena (open-ended prompts), WMT (translation) and GeoFactX (reasoning).

Synthetic data generation: They assembled a pool of 5 strong open models of different sizes: Gemma3-27B, Kimi-K2, Qwen3-235B, DeepSeek-V3, Command A. For each of 10k prompts (translated into 9 languages) they took one answer from each teacher. Then they applied either Best-of-N (selection via RM) or FusioN (synthesis via Command A). The resulting data was used to fine-tune a 111B student model, and the fine-tuned models were evaluated on the same benchmarks.

Why the results came out this way: FusioN consistently beat BoN: on Arena by +2.5% win-rate on average, on translation by +0.8 XCometXL. The most interesting part: in some cases FusioN surpassed the Oracle (the best answer in the pool by the ground-truth metric). For example, on the de, ru and zh translation directions FusioN produced an answer better than any of the original 5 variants. That backs the central claim: selection is not the upper bound, synthesis can go further.

The surprise: Even when the teacher pool was weakened (a strong model swapped for a weak one), FusioN stayed more robust than Best-of-N. When one teacher was removed (4 instead of 5), BoN dropped harder. This shows that FusioN makes better use of diversity, even when individual contributors are weak.

Practical insight: The method is more sample-efficient: FusioN with N=2 gives the same gain as BoN with N=4. So for the same budget (in tokens), FusioN squeezes out more. Especially valuable when resources are tight.

💡 Adaptation: different roles instead of different models

If you don't have a pool of different models, you can use one model in different roles:

Write 3 versions of a review of the book "{title}":

Variant 1: as a literary critic (focus on style, structure, language)
Variant 2: as a reader from the target audience (emotions, engagement, usefulness)
Variant 3: as a professional editor (weak spots, what could be improved)

Then:
[paste the FusioN template above]

Each role nudges the model toward different aspects, and the synthesis gives you a balanced review.

💡 Adaptation for B2B sales copy

Example: writing a sales proposal:

Write 3 versions of a sales proposal for a B2B client (an analytics service for Amazon sellers):

Variant 1: Straightforward: hard ROI numbers, case studies, guarantees
Variant 2: Expert: a deep dive into the client's problem, market insights
Variant 3: Emotional: a success story, how the client's business gets transformed

[Then synthesize with FusioN]

🔧 Technique: controlling how much to synthesize

You can control how aggressively the fusor should synthesize:

For conservative synthesis (minimal changes):

Write the final answer while preserving the original variants' wording as much as possible.
Combine their fragments as-is, editing only the seams between them.

For aggressive synthesis (create something new):

Do not copy fragments verbatim. Take the ideas from each variant 
and rewrite the final answer entirely in your own words, weaving those ideas together.

💡 Extrapolation: cascaded FusioN for very complex tasks

For tasks that need many different perspectives (say, a one-year product strategy):

Step 1: Generate N=5 variants with different focuses
Step 2: FusioN produces a first version
Step 3: Generate N=3 more alternative takes on the result of step 2
Step 4: A final FusioN combines the version from step 2 with the alternatives from step 3

This gives you two-level synthesis: first a baseline merge, then refinement through fresh perspectives.

Making, not Taking, the Best of N by Ammar Khairi, Daniel D'souza, Marzieh Fadaee, Julia Kreutzer (Cohere Labs)

Data and evaluations: The team published the synthetic datasets and pairwise evaluations on HuggingFace:

• CohereLabs/fusion-synth-data-ufb
• CohereLabs/fusion-synth-data-geofactx
• CohereLabs/fusion-pairwise-evals-test-time-scaling
• CohereLabs/fusion-pairwise-evals-finetuned