AnalysisDevelopersNovember 14, 2026

Cmpnd releases dspy-wsp, weight-space programming for DSPy

Read original source →marktechpost.com

The MIT-licensed optimizer lets a frontier LLM edit a student model's weights directly — no gradients, no training loop, no fine-tuning data — for roughly $5–$10 of GPT-6-Astra calls per run. Tested on 0.6B–9B decision models, it added 1.9–16.1 points on Typed Decisions, or 6.5–20.0 with ReAnchor.

1 source

More stories today

Open the live feed