Microsoft researchers introduce Jigsaw, a new tool that can help these big language models perform better. Jigsaw is a Python Pandas API code generator that accepts multi-modal inputs. Jigsaw uses post-processing techniques to decipher the syntax and semantics of programs and then uses user feedback to improve future performance. [SOURCE]
2 Likes
I wonder if anyone has researched embedding a symbolic computation engine like KLEE into feedback loop with tools like Codex and this.
In theory it should let you build models faster because rather than needing to physically execute the entire program, you can symbolically compute what would have happened and programmatically manipulate the application state, I think?
Easier said than done.
EDIT: The only reference to something I’ve found along these lines was a 2020 paper