Hello everyone, We are excited to present our latest survey paper, titled "Controllable Text Generation for Large Language Models: A Survey," which delves into the controllable generation of text by large language models (LLMs).(1/6)

Controllable Text Generation for Large Language Models: A Survey paper:

Better, steerable, and controllable text generation is key to applying LLMs in the real-world. While there are many issues with LLMs, there is still a sense of progress in areas like safety, consistency, style, and helpfulness. These areas account for most of the AI work I do with companies today. It's great to see a detailed report specifically on methods for controllable text generation in LLMs.

The paper discusses the field of Controllable Text Generation (CTG) for Large Language Models (LLMs). CTG aims to generate text that meets specific requirements, such as safety, sentiment, and style, while maintaining high quality in terms of fluency, coherence, and diversity. The paper presents a detailed classification of CTG methods into two main stages: the training stage and the inference stage. It discusses the key techniques used in each stage, such as retraining, fine-tuning, reinforcement learning, prompt engineering, latent space manipulation, and decoding-time intervention. full paper: