AI Aulis

AI NEWS IN ENGLISH · 15 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

26 stories · next issue tomorrow morning · How the paper is made →

← Issue 15 November 2026Paper

Frontier model improves five decision models’ accuracy by editing their weights in Cmpnd’s tests

AI-generated contentWritten by AI

github.com, marktechpost.com, cmpnd.ai, aikraft.ru, tldr.tech · 5 sources · Sources published: 13–14 November 2026 · Some source publication dates are unknown

With Cmpnd’s dspy-wsp optimiser, a frontier model reads and edits a small model’s weights without training. In the company’s own tests, the accuracy of five decision models rose by 1.9–16.1 percentage points.

Drew Breunig, CEO of Cmpnd, introduced dspy-wsp on the company’s blog on 13 November. dspy-wsp is an optimiser for the DSPy library, and the company itself describes it as highly experimental. DSPy is an open-source library for building programs that use language models and for optimising them against a metric. The new WeightSpace optimiser does not edit prompts. It changes a model’s weights without gradients or training.

The method comes from Sparks of Weight-Space Programming, a blog post by MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. In it, GPT-6-Astra edited 0.016 per cent of Qwen3.5-0.8B’s weights, and the model’s MATH-500 score rose from 7.9 per cent to 48.1 per cent. In the researchers’ experiments, editing removed obstacles but did not add new capabilities: Gemma 3 270M’s score barely changed. Cmpnd says it has not yet reproduced the study’s full result.

The frontier model tests and edits

The frontier model, which Cmpnd calls the writer, reads the small model’s weights with Python code in a sandbox. It runs the user’s program on training examples a limited number of times, works out what is going wrong in the neural network and saves a targeted change. The result is a small patch file containing only the changed values and checksums. It can be used to modify a public base model.

Cmpnd says the method works best with decision models. A decision model reads a situation description and typed questions, then assigns a probability to each answer option. It does not generate text.

Results in Cmpnd’s own tests

The table shows how much the accuracy of five decision models rose on 400 cases from the Typed Decisions benchmark. These cases were not used in optimisation. ReAnchor is an experimental DSPy optimiser that adjusts answer thresholds without touching the weights.

Model Baseline ReAnchor WeightSpace Both
d1-omni-600M 50.9% +9.3 +16.1 +18.3
NanoJev 39.2% +11.6 +13.3 +17.3
d1-2B 60.6% +9.5 +13.4 +17.3
Strands Decider 2B – +13.7 +10.4 +20.0
Clef-Flash (9B) 68.8% +1.5 +1.9 +6.5

Change in percentage points compared with the unedited model. The metric is the accuracy of the answers the program returns. The writer was GPT-6-Astra, which could run each model 25 times during its optimisation. Standard error is 0.7–1.4 points. Source: Cmpnd; NanoJev’s baseline and its ReAnchor and combined results from MarkTechPost, which cites Cmpnd’s results report.

The results do not support the conclusion that editing weights is always the best approach. On Strands Decider 2B, ReAnchor alone improved the result more. MarkTechPost’s claim that WeightSpace beat ReAnchor on every model applies only to the four models in its table.

The choice of writer model also affected the result. Edits by GPT-6.1-Sol changed accuracy by +5.8 percentage points on d1-omni-600M, +12.1 on NanoJev and −0.6 on d1-2B.

By default, the writer may compute the decision layer’s outputs itself in the sandbox, which the original study did not allow. Under the study’s rules, d1-omni-600M’s accuracy rose by only 2.8 ± 0.7 percentage points, and the edit did not improve on ReAnchor’s result.

Costs and privacy

According to the repository’s documentation, optimising a decision model with 25 runs took 68–167 minutes, and the writer model’s calls cost $5.42–10.78. MarkTechPost’s figures (78–167 minutes, $5.42–7.63) are based on three models.

The sandbox stops the writer’s code from using the network or reaching the home directory. The writer model’s provider still sees the model weights and training examples as the writer reads them.

dspy-wsp is on GitHub under the MIT licence and requires Python 3.11 or later, DSPy 3.5 and an API key for the writer model. The d1-omni-600M model weights can be optimised on a Mac or on a GPU with at least 4 GB of memory; installation instructions are in the repository.

Report an error in this story