What Has Cmpnd Released?
Cmpnd has released dspy-wsp, an experimental optimizer that introduces weight-space programming, or WSP, into the DSPy ecosystem. According to MarkTechPost, the package is distributed under the MIT license, is installed via pip, and requires DSPy 3.5. The central component is called WeightSpace and plugs into any DSPy program and metric.
The approach upends classical fine-tuning: no gradients, no training loop, no dedicated dataset. A frontier model—in the tests, GPT-6-Astra—reads the weight tensors of another model, reasons about what they compute, and modifies them to achieve an objective expressed in natural language. The idea stems from MIT research titled Sparks of Weight-Space Programming, which the same source describes as an emergent capability rather than a codified method.
How It Works and What Was Measured
The dspy-wsp workflow involves three actors: a writer, which proposes a brief explanation followed by a code block at each turn; a forward-pass mechanism with a fixed budget, set to 25 steps in the tests; and a hidden validation set that compares original and modified weights on the same examples and seeds. The resulting code is a patch: the writer can apply targeted modifications, such as the one to the backbone of the 2-billion-parameter model at layer 23 or to the scorer portion of the decision head of d1-omni-600M.
The trials were conducted on Typed Decisions, a benchmark with 400 cases and 2,000 typed probabilistic decisions. In the cited data, WeightSpace alone added 1.9 to 16.1 points; combined with ReAnchor, an earlier optimizer that recalibrates response thresholds without LLM calls, gains rose to between 6.5 and 20.0 points. The best case was d1-omni-600M, which rose from 50.9% to 69.1% with both tools. The worst was Clef-Flash at 9 billion parameters, stuck at +1.9 with WeightSpace alone.
Limitations and open questions
The source flags several reasons for caution. Under the stricter rule used in the original work, which prevents the writer from recalculating the decision head in its own kernel, d1-omni-600M gained only 2.8 points, and adding ReAnchor did not outperform ReAnchor alone. The choice of writer matters: with gpt-6.1-sol, results were +5.8, +12.1, and -0.6 on three models, while Astra produced +16.1, +13.3, and +13.4. Each run takes 78 to 167 minutes. The experiments were run on a single Mac, with about 6 GB of memory for the 2-billion-parameter models and at least 4 GB for d1-omni-600M.
Cmpnd also presents the strengths: low cost—between $5 and $10 in GPT-6-Astra calls to improve each model tested—and small, verifiable patches, with sha256 hashes of the starting and modified tensors that allow the weights to be reconstructed exactly. The picture remains experimental, however: gains are weaker on large models, and the original research shows that the modification can remove an obstacle but cannot add a missing capability. The same source notes that the concept does not exclude gradients, although they were banned in the experiments.
