Prototype brings weight-space programming to DSPy
Cmpnd has released dspy-wsp, a prototype optimizer that adds weight-space programming to DSPy. Its WeightSpace component lets a frontier language model inspect and directly modify the weights of a smaller model, then evaluate each change. The process uses no gradients, training loop or fine-tuning dataset. The package is MIT-licensed, can be installed with pip, and requires DSPy 3.5.
In reported tests, the system was applied to open decision models ranging from 0.6B to 9B parameters. A run used up to 25 forward passes and cost about $5 to $10 in GPT-6-Astra calls. WeightSpace alone improved Typed Decisions scores by 1.9 to 16.1 points; combined with ReAnchor, gains reached 6.5 to 20.0 points. The largest cited improvement raised d1-omni-600M from 50.9% to 69.1%.
The experiments ran on a Mac, with 2B models using about 6 GB of memory. Runs took 78 to 167 minutes, and the largest model showed the weakest improvement. Cmpnd describes the software as very experimental. Under the original research paper’s stricter evaluation rule, d1-omni-600M improved by only 2.8 points.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.