Paper Walkthrough: Controllable Neural Text Generation (https://lilianweng.github.io) #gpt #safety

Large pretrained language models can generate fluent text, but often lack control over attributes like topic and style. Researchers have explored techniques to steer outputs during decoding, optimize prompts to elicit desired text, fine-tune models on downstream tasks, train small tunable components alongside base models, and directly alter training objectives. Each approach has tradeoffs. Key developments include guided decoding strategies, learning prompt and discriminator models for preferential generation, controlled fine-tuning with reinforcement learning, and modifying model internals to enable intervention. More work is likely needed to unlock both wide coverage knowledge and precise control over free-form text generation.