Making LLMs Better at Creative Writing Using Entropy (www.countbayesie.com)

Recent advancements in the field of AI writing have focused on enhancing large language models (LLMs) by improving their sampling techniques, particularly to foster creativity in generated content. The new approach involves a modified sampling method that incorporates future entropy—an indicator of the diversity of potential next tokens—into the text generation process. This innovation directly addresses the common criticism that LLM outputs often feel generic or predictable, as pointed out in a viral discussion featuring Ben Affleck. By considering not just the current token's probability but also how it influences future choices, the future-entropy sampler aims to produce outputs that feel less formulaic and more engaging. The future-entropy sampler works by calculating the potential entropy of the token distribution that would follow a chosen token, and this value is then used to adjust the selection probabilities. The approach introduces a variable, α, that allows for fine-tuning the balance between the original probabilities and future entropy in the sampling process. This level of customization can lead to richer and more varied outputs, potentially overcoming the limitations of traditional LLM sampling methods. The results indicate a noticeable improvement in the quality of creative writing generated by LLMs, encouraging further exploration into tailored samplers and their impact on AI-generated content.