Cmpnd has published DSPy-WSP, an open-source optimizer that plugs into DSPy. It lets a language model inspect a smaller model and directly change selected weights, the numerical parameters that shape the smaller model’s behaviour.
How the optimizer works
The larger model acts as a writer in a sandboxed Python session. It can run the student model on practice examples, inspect what happens and propose targeted edits. The workflow uses a fixed budget of forward passes, so the writer has a limited number of opportunities to test its ideas. The editing step uses no gradient-based training.
How to tell if an edit helped
The repository describes holding back some examples until the final check. It runs the original and edited models on that same holdout set with the same seeds. That gives developers a way to see whether the edits help on examples the writer did not see. The repository reports results for specific test tasks. Results for a new task depend on the model, data and evaluation metric.
Sources: Cmpnd’s DSPy-WSP repository and implementation notes and MarkTechPost’s November 14 coverage.
Illustrative programming photograph by Negative Space. Wikimedia Commons, CC0 1.0, cropped and colour graded.