dspy-wsp Lets a Frontier LLM Edit Small Model Weights Directly: +18.3 Pts for ~$5-8 per Run

stanfordnlp · x · 2026-11-16

Cmpnd open-sourced dspy-wsp, bringing weight-space programming to DSPy: a frontier writer LLM reads a small model's tensors, hypothesizes what they compute, and patches weights directly — no gradients, training loops, or fine-tuning data. Each run verifies edits with 25 forward passes and ships the best one as a sha256-hashed patch (MIT license).

All tested 0.6B–9B decision models improved: d1-omni-600M rose 50.9%→69.1% (+18.3 stacked with ReAnchor) on Typed Decisions; d1-2B hit 77.9%; NanoJev 0.6B +17.3
WeightSpace alone beat the ReAnchor optimizer on every model charted, but scaled poorly to 9B Clef-Flash (+1.9 alone)
Cost: $5.42–7.63 per run, 78–167 minutes; all experiments ran on a single Mac (6 GB for 2B models)

Original post →

dspy-wsp Lets a Frontier LLM Edit Small Model Weights Directly: +18.3 Pts for ~$5-8 per Run

More from coding & agent