Making LLMs Better at Creative Writing Using Entropy
The article explains how entropy, a measure of randomness or unpredictability in language models, can be adjusted to improve the quality of creative writing outputs from large language models (LLMs). It suggests that carefully modulating entropy levels helps balance coherence and novelty, allowing models to produce more engaging and less predictable creative text.
Background
- **LLMs** (e.g., GPT-4, Claude) generate text by picking the most likely next word. Standard methods (greedy sampling, top-k, top-p) favor safe choices, making output generic — fine for facts, bad for fiction.
- **Entropy** measures how "surprised" the model is. Low entropy = one word is very obviously next; high entropy = many words could fit, so there's room for creativity.
- The article proposes using entropy in real-time to adjust **temperature** (a randomness dial). When entropy is high, increase randomness to explore more surprising continuations; when low (e.g., a proper name), stay conservative.
- Why it matters: this simple, inference-time trick could make LLMs produce less repetitive, more stylistically varied creative writing — poetry, dialogue, literary prose — without retraining the model.