LLM에서 Temperature란 무엇인가요?

LLM이 다음 토큰을 선택하는 방법

Softmax 함수

import numpy as np

# Raw logits from the model
logits = np.array([2.0, 1.0, 0.5, -1.0])  # scores for 4 tokens

# Standard softmax (temperature = 1)
def softmax(logits, temperature=1.0):
    scaled = logits / temperature
    exp_scaled = np.exp(scaled - np.max(scaled))  # subtract max for numerical stability
    return exp_scaled / exp_scaled.sum()

probs = softmax(logits, temperature=1.0)
print('Probabilities:', np.round(probs, 3))
# [0.567, 0.208, 0.129, 0.095]
# Token 0 is most likely at 56.7%

온도 = 0: 탐욕적 디코딩

# Temperature = 0 (greedy)
probs_temp0 = softmax(logits, temperature=0.01)  # near-zero
print('Probs at T=0.01:', np.round(probs_temp0, 4))
# [~1.0, ~0.0, ~0.0, ~0.0]

# In practice, temperature=0 is implemented as argmax:
def greedy_sample(logits):
    return np.argmax(logits)  # always returns the index of the highest logit

token_idx = greedy_sample(logits)
print(f'Selected token index: {token_idx}')  # always 0

온도 = 1: 표준 샘플링

# Temperature = 1 (standard)
probs_temp1 = softmax(logits, temperature=1.0)
print('Probs at T=1.0:', np.round(probs_temp1, 3))
# [0.567, 0.208, 0.129, 0.095]

# Sampling from this distribution:
def sample_token(probs):
    vocab = ['the', 'a', 'an', 'is']
    return np.random.choice(vocab, p=probs)

# Run 10 times to see variation
for _ in range(10):
    print(sample_token(probs_temp1), end=' ')
# Output varies each time, but 'the' appears most often

온도 = 2: 혼란스러운 샘플링

# Temperature = 2 (chaotic)
probs_temp2 = softmax(logits, temperature=2.0)
print('Probs at T=2.0:', np.round(probs_temp2, 3))
# [0.385, 0.261, 0.211, 0.143]
# Much flatter — the 4th token (logit=-1.0) now has 14.3% chance
# (vs 9.5% at T=1.0)

# Visualization: compare distributions
for temp in [0.1, 0.5, 1.0, 1.5, 2.0]:
    probs = softmax(logits, temperature=temp)
    print(f'T={temp}: {np.round(probs, 3)}')

분포의 첨예도를 조절하는 온도

import matplotlib
# Conceptual: what distribution shape looks like at different temps
temps = {'T=0.2 (sharp)': 0.2, 'T=1.0 (normal)': 1.0, 'T=2.0 (flat)': 2.0}
for label, temp in temps.items():
    probs = softmax(logits, temp)
    bar = '#' * int(probs[0] * 40)
    print(f'{label}: top token = {probs[0]:.1%} |{bar}|')
# T=0.2: top token = 97.2% |########################################|
# T=1.0: top token = 56.7% |######################|
# T=2.0: top token = 38.5% |###############|

온도와 결정성

import openai
client = openai.OpenAI(api_key='sk-...')

# Deterministic: temperature=0
resp1 = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': 'What is the capital of France?'}],
    temperature=0
)
resp2 = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': 'What is the capital of France?'}],
    temperature=0
)
print(resp1.choices[0].message.content == resp2.choices[0].message.content)  # True (usually)

온도가 상위 p와 상호 작용하는 방식

# Using temperature with top_p in OpenAI API
resp = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': 'Write a one-line haiku about code.'}],
    temperature=0.9,  # moderately diverse distribution
    top_p=0.95        # sample from top 95% of probability mass
)

과제별 실용적인 온도 값

TEMPERATURE_PRESETS = {
    'factual_qa': 0.0,
    'code_generation': 0.1,
    'summarization': 0.4,
    'chat': 0.8,
    'creative_writing': 1.1,
    'brainstorming': 1.3
}

def call_with_preset(task_type, prompt):
    temp = TEMPERATURE_PRESETS.get(task_type, 0.7)
    return client.chat.completions.create(
        model='gpt-4o',
        messages=[{'role': 'user', 'content': prompt}],
        temperature=temp
    )

API에서의 온도

# OpenAI: temperature 0 to 2
client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': prompt}],
    temperature=1.2  # Valid for OpenAI
)

# Anthropic Claude: temperature 0 to 1
import anthropic
claude = anthropic.Anthropic(api_key='sk-ant-...')
claude.messages.create(
    model='claude-opus-4-5',
    max_tokens=1024,
    temperature=0.8,  # Max is 1.0 for Claude
    messages=[{'role': 'user', 'content': prompt}]
)

온도만으로 충분하지 않을 때

# Seed for reproducibility (OpenAI API)
resp = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': prompt}],
    temperature=0,
    seed=42  # Guarantees same output across calls on same model version
)

# Note: same output only guaranteed with same model version
# A model update can change outputs even with the same seed

지식 확인

복습: LLM의 온도