Jeffrey Binder on PromptArray
Jeffrey Binder has a fascinating project on encouraging LLMs to explore the vector space through boolean expressions that nudge them toward (or away from) certain directions. He calls this PromptArray and it uses boolean syntax to essentially (1) create multiple versions of prompts, (2) then combine them into a single response.
PromptArray runs the model using multiple variations of the prompt, but it only generates one continuation, produced by composing together the model’s predictions for the different prompt variants.
The analogy to something like RLHF is clear, in that the goal of both is to encourage specific behaviour through approaches beyond pre-training adn data curation, but here that kind of guidance is during inference time. RLHF also seems that what you're training the model to do is learn from analogy across prompt types, but this mode of prompting is more explicit and situated. It's a programmatic, scaffolded way of trying to steer a model's inference. It feels less anthropomorphic in spirit, which just dovetails with my idea of using LLMs as tools.
In his blogpost, Jeff talks through adapting this approach from its GPT-2/3 origins to the more recent GPT-OSS variants that are open source albeit have thinking modes that make some of the boolean variants trickier.
I also thought this was an interesting note on the interest in Boolean operators, from the Git repo:
This project also takes inspiration from some old insights from linguistic anthropology. Structuralists such as Claude Lévi-Strauss maintained that language is based on difference: in order to understand what it means for something to be raw, we also must understand the meaning of "cooked." Yet it is not always self-evident which words are opposites. Perhaps it is clear that the opposite of day is night, but what is the opposite of a narrative? Is it, as Lev Manovich once posited, a database? A scientific analysis? A photograph? Silence? The as opposed to operator makes it possible to specify which opposite one has in mind, thus guiding the generator with more precision.
Forward links:
- This pointed me to his article The Artifice of AI: Why Machine-Generated Text Seems So Fake, which is now on my to-read list.
- The repo for this project links to the inspiration for it: Jean Sammet's vision for English as a programming language. Seems like it would also be an interesting read.
- This reading from Lev Manovich on databases and film? I am not quite sure but seeing a film journal linked on a Git repo is a surefire signal that it will be fun.