🧠 Core Concept

Multi-Candidate Generation

Why asking AI once is a gamble — and how generating multiple candidates at different temperatures covers the solution space.

The Problem

Single-Shot AI Is Unreliable

When you ask a standard AI a question, you get one answer. That answer is shaped by a parameter called temperature — a randomness dial that controls how "creative" the model is.

Low temperature (0.1–0.3): Deterministic, focused, but sometimes stuck in local patterns
Medium temperature (0.4–0.7): Balanced, but can still miss the right answer
High temperature (0.8–1.0): Creative and diverse, but more likely to hallucinate

The problem? No single temperature is optimal for every question. A math problem needs precision. A creative task needs divergence. And you don't know in advance which one you're dealing with.

💡 The Key Insight

Temperature is randomness. Each time you sample at a given temperature, you're rolling the dice. Ask 100 times at temperature 0.7, and you'll get different answers each time. Some right. Some wrong. The question is: how do you find the right one?

The Solution

Generate N Candidates, Cover the Space

Instead of betting on a single sample, Mikoshi AI Turbo generates multiple candidates at different temperatures. Each candidate explores a different region of the solution space:

Focused

0.3

Precise, conservative

Balanced

0.6

Best of both worlds

Creative

0.9

Divergent, exploratory

By sampling across the temperature spectrum, we get diverse perspectives on the same problem. The focused candidate captures the obvious answer. The balanced candidate finds nuance. The creative candidate might discover an unconventional but correct approach.

// Turbo candidate generation strategy
const temperatures = [0.3, 0.6, 0.9];

const candidates = await Promise.all(
  temperatures.map(async (temp, i) => ({
    id: i,
    temperature: temp,
    response: await llm.generate(prompt, { temperature: temp }),
  }))
);
// Now we have 3 diverse answers to evaluate
      

Example

Three Candidates, One Question

Consider the question: "What is 47 × 83?" — the correct answer is 3,901.

💬 "What is 47 × 83?"

↓

t=0.3

↓

t=0.6

↓

t=0.9

🎯 Candidate A — temp 0.3

"3,901"

Precise calculation, step by step

✅ Correct

⚖️ Candidate B — temp 0.6

"3,901"

Quick mental math approach

✅ Correct

🎲 Candidate C — temp 0.9

"3,891"

Creative reasoning, made an error

❌ Incorrect

With single-shot AI, if you happened to sample at high temperature, you'd get the wrong answer and never know. With multi-candidate generation, two out of three candidates agree — and more importantly, the verification step will prove which ones are right.

Theory

Why Multiple Candidates Works

The mathematics behind multi-candidate generation are compelling:

Error reduction: If a single model has 80% accuracy, generating 3 independent candidates and selecting the best gives you ~99.2% accuracy (assuming at least one is correct and verification works)
Coverage: Different temperatures activate different neural pathways in the model, producing genuinely different reasoning strategies
Consensus: When multiple candidates agree, confidence is high. When they diverge, that disagreement itself is useful signal
Fault tolerance: If one generation fails, crashes, or returns garbage, the others can still succeed

// Probability of at least one correct answer
// with N candidates each having accuracy p
const P_at_least_one = 1 - Math.pow(1 - p, N);

// p = 0.80, N = 3
// P = 1 - (0.20)³ = 1 - 0.008 = 0.992 → 99.2%

// p = 0.70, N = 5
// P = 1 - (0.30)⁵ = 1 - 0.00243 = 0.997 → 99.7%
      

This is the same principle behind ensemble methods in machine learning — multiple weak learners combined outperform any single strong learner.

Configuration

Tuning the Candidate Pool

Mikoshi AI Turbo lets you configure the number of candidates (default: 3) and the temperature spread. More candidates = higher accuracy but slower response time:

// Default: 3 candidates (fast + accurate)
{ candidates: 3, temperatures: [0.3, 0.6, 0.9] }

// High accuracy: 5 candidates
{ candidates: 5, temperatures: [0.2, 0.4, 0.6, 0.8, 1.0] }

// Speed mode: 2 candidates
{ candidates: 2, temperatures: [0.3, 0.8] }
      

The optimal configuration depends on your use case. For financial calculations, use 5 candidates. For casual questions, 2 may suffice.

See multi-candidate generation in action

⚡ Try Turbo in Synapse

← Back to Turbo Code Verification →