dspy-wsp releases prototype optimizer that edits model weights via LLM

dspy-wsp released a prototype DSPy optimizer, dspy-wsp, that lets a writer LLM such as GPT-6-Astra edit another model's weights without gradients or training. On five open decision models from 0.6B to 9B parameters, tested with the Typed Decisions benchmark, WeightSpace alone added 1.9 to 16.1 points of accuracy on 400 held-out cases, and 6.5 to 20.0 points when combined with the ReAnchor optimizer, the project says. The tool requires DSPy 3.5 and is described by its authors as very experimental. Each run used 25 forward passes at maximum reasoning effort for roughly $5-10; edited models run locally through MLX on Apple silicon or PyTorch on NVIDIA GPUs. The authors caution that editing runs on text-generating LLMs are long and costly and may not produce compelling results. Weight-space optimization on d1-omni-600M runs on a Mac or a GPU with at least 4 GB of memory.

Generated from reports · updated 2 h ago

Not enough continuous observations yet to draw a trend.