Generative AI models, including large language models (LLMs) and diffusion-based image generators, are characterized by their non-deterministic nature. Unlike traditional systems, which might rely on a single predetermined output, these models generate responses by selecting from a probability distribution of likely outcomes based on user prompts. For example, asking an LLM for the capital of France could yield various forms of the answer “Paris,” highlighting the flexibility of these models.
However, frequent users have observed that the outputs can sometimes be repetitive, demonstrating a phenomenon known as mode collapse. This occurs when the models, during post-training alignment, gravitate towards safer, more common responses, limiting their creativity. A collaborative research effort from Northeastern University, Stanford University, and West Virginia University has introduced a method called Verbalized Sampling (VS) to address this limitation. By simply adding the instruction to generate multiple responses along with their probabilities, the researchers found that the models produced a broader range of outputs.
The VS technique enables models such as GPT-4, Claude, and Gemini to access a greater diversity of potential responses without the need for additional training. When prompted in this fashion, models articulate their internal distribution of possibilities, which enhances the variety of outputs substantially.
The research tested this method across several tasks, including creative writing, dialogue simulation, open-ended question answering, and synthetic data generation. Results indicated significant increases in output diversity without sacrificing quality. Furthermore, the method allows for tunable diversity by adjusting probability thresholds, enhancing its usability across varying model sizes.
Available as a Python package, VS can be integrated with applications like LangChain, facilitating easier access for users. While some initial errors may occur, adjustments to the prompting structure can typically resolve these issues. In summary, Verbalized Sampling provides a straightforward solution to increase the diversity and quality of generative AI model outputs, potentially benefiting numerous fields such as writing and education.
Source: https://venturebeat.com/ai/researchers-find-adding-this-one-simple-sentence-to-prompts-makes-ai-models