Cmpnd 发布 dspy-wsp 用前沿 LLM 直接改写小模型权重
Cmpnd 开源 dspy-wsp,一种基于 DSPy 的权重空间编程优化器:由前沿写作模型在沙箱中读取学生模型张量、提出假设并直接打补丁,无需梯度、训练循环或微调数据。在 Typed Decisions 基准上,d1-omni-600M 单独使用 WeightSpace 从 50.9% 提升至 66.9%,叠加 ReAnchor 后达 69.1%,但 9B 模型仅提升 1.9 个百分点,且单次运行成本 5.42 至 7.63 美元、耗时 78 至 167 分钟。
报道正文
▲ Marktechpost AI · 2026-11-14 16
▲ Cmpnd just released dspy-wsp, a DSPy optimizer that lets a frontier LLM edit a small model's weights directly. No gradients, no training loop, no fine-tuning data.
The key idea is weight-space programming. A writer LLM (GPT-6-Astra) reads the student model's tensors in a sandbox, forms a hypothesis about what they compute, and patches a few. It tests each edit with forward passes from a fixed budget of 25. The best edit ships as a small, hashed patch.
On the Typed Decisions benchmark, d1-omni-600M rose from 50.9% to 66.9% with WeightSpace alone (+16.1), versus +9.3 for the ReAnchor threshold optimizer. Stacked together, it reached 69.1% (+18.3). It is a trade-off though: the 9B Clef-Flash gained only +1.9 from WeightSpace alone, and under the original paper's stricter rule d1-omni-600M gained just +2.8.
Each run cost $5.42 to $7.63 in writer calls and took 78 to 167 minutes. Every experiment ran on 1 Mac. MIT license, pip install dspy-wsp, needs DSPy 3.5.....
Full analysis:
