Prompt-based methods

Methods that guide language models toward specific tasks through careful instruction-writing (prompts) rather than task-specific fine-tuning. Prompt-based approaches leverage the implicit knowledge and capabilities learned during pretraining, enabling zero-shot and few-shot task performance. The effectiveness of these methods depends critically on prompt engineering—choosing the right instruction wording, examples, and format.

Key papers

• Toxicity Detection with Generative Prompt-based Inference — Explores generative vs. discriminative prompt formulations for toxicity detection; shows sensitivity to prompt wording and the importance of prompt design for achieving strong performance

Related topics

• Language Models — the models being prompted
• Zero-shot learning — performing tasks without task-specific training data
• Few-Shot Learning — using a few examples to guide model behavior