Recent research has uncovered some interesting techniques for prompting large language models (LLMs) like GPT-3 to produce higher quality responses. Strategies like "let's think through this step-by-step", "tree of thought", and "take a deep breath" aim to activate the latent knowledge embedded in the enormous parameter space of LLMs. However, these approaches lack a unified framework for understanding how to systematically activate latent knowledge. The key is latent space activation.
Latent space refers to the vast repository of implicit knowledge contained within the parameters of a trained LLM like GPT-3. At any given moment, only a small portion of this knowledge is "active" and influencing the model's responses. Latent space activation refers to techniques for selectively recruiting more of the model's implicit knowledge to bear on the current context.
This parallels how human cognition works. We have intuitions and gut reactions, but can also engage in deliberative, systematic thinking to reason more effectively. Psychologist Daniel Kahneman's popular book "Thinking Fast and Slow" distinguishes these two modes of thought as System 1 (fast, intuitive) vs System 2 (slow, analytical). A single inference from an LLM is akin to System 1 intuition. Prompt strategies aim to activate System 2-like latent knowledge.
How can prompt engineers actually implement latent space activation? The key is to prime the model to step through reasoning processes similar to how humans logically think through problems. For example, ask the model questions like:
- What information do I already know about this topic?
- What techniques or methods could help answer this question?
- How will I integrate what I know to discuss the question and give my final answer?
This prompts the model to actively recall relevant knowledge and "step through" logical reasoning before rendering a final response. Each step expands the context window, surfacing more latent knowledge to bring to bear. I implemented a simple prototype of this approach [link to Github repo] which shows it can improve reasoning.
An even more powerful technique is to prompt the model to brainstorm lists of targeted search queries relevant to the question at hand. Humans intuitively do this all the time in something called the "brainstorm, search, hypothesize, refine loop" (BSHR). We brainstorm questions and searches which, when followed up on, provide key facts to refine our understanding iteratively. Large language models can generate comprehensive search queries by prompting them to "employ everything you know about information foraging and information literacy." This allows tapping deep latent knowledge about how to strategically search for and synthesize information.
In summary, latent space activation through prompting models to mimic human reasoning and search strategies shows great promise. Rather than relying on narrow techniques, prompt engineers should focus on principles for systematically directing models to tap into their vast latent knowledge, just as our minds tap into our full mental potential. With the right prompts, extraordinary latent potential can be activated!