Temperature Scaling
A post-hoc calibration method that uses a single parameter (temperature) to adjust model confidence values.
Temperature scaling is the simplest calibration method: one parameter adjusts model confidence and controls creativity in LLMs.
Explanation
Temperature scaling divides logits by a learned temperature value T before softmax. T>1 makes predictions softer (less overconfident), T<1 makes them sharper. The simplest and often most effective calibration method.
Marketing Relevance
For LLM-driven marketing responses, temperature controls creativity vs. consistency – low for fact-based, high for creative outputs.
Example
GPT with temperature 0.1 delivers consistent, fact-based product descriptions; with 0.9 creative, varied marketing copy.
Common Pitfalls
Temperature calibration on validation data doesn't generalize under distribution shift. For LLMs: temperature ≠ calibration of factual accuracy.
Origin & History
Hinton et al. introduced temperature in the context of Knowledge Distillation (2015). Guo et al. (2017) showed that temperature scaling suffices for calibration, often making more complex methods unnecessary.
Comparisons & Differences
Temperature Scaling vs. Platt Scaling
Platt scaling learns two parameters (slope + intercept) on logits; temperature scaling uses only one parameter and is less prone to overfitting.
Temperature Scaling vs. Top-p Sampling
Temperature affects the probability distribution; Top-p limits selection to the most probable tokens.