Tricks for LLM diversity
•1 min read
- The more context, the less slop. But too much context results in mode collapse.
- Personas: "Write like Ernest Hemingway"
- Rubrics as Rewards RL
- Blending/Ensembles
- Verbalized sampling: "Generate [artefacts] about [topic] and their corresponding probabilities"
- Paraphrase and randomize your prompts, sample, repeat.
- Don't use structured outputs.
- Don't use code or markup in your prompt, only simple markdown.