LLMs have learned to read and edit neural network weights

What weight-space programming is

Usually a neural network's behavior is changed by training: you collect data, compute gradients and wait. The authors propose another way — give a strong language model the network's weights as a set of numbers and ask it to work with them like code: read them, find the defect and fix it.

In eight experiments the GPT-6 Astra model did this without gradient descent, without tools and without specially trained helpers. The authors called the approach Weight-Space Programming.

Four experiments

• Recognize the task from the weights. The model was given, one by one, the weights of four game policies with the same architecture, inputs and actions. With no hints it correctly named all four games: Breakout, Pong, Pac-Man and Freeway.
• Repair a network. A car-driving policy got stuck on a hill. From the weights, a scene description and one failed trajectory, the model wrote an edit of 3.5% of the weights in a single answer — after it the task was solved in every attempt.
• Write from scratch. From the rules and code of Breakout the model wrote out the weights of a network straight away, and it scored 20 times more points than a PPO policy after 500 million training steps.
• Improve another LLM. The small Qwen3.5-0.8B is known for looping in its reasoning. Over 50 runs on training tasks the model found the cause and edited 0.016% of the weights: according to the authors, math improved up to six times and instruction following up to three.

How it was re-checked

What convinces more than the paper itself is that people have started reproducing the result. One participant in the alphaXiv discussion applied the published edit to Qwen3.5-0.8B and got similar numbers.

Another participant repeated the first experiment with Claude models: Opus 5.5 also recognized all four games, even without a description of the environment, while smaller models recognized none. So the ability exists only in the strongest models.

Why it is too early to celebrate

• A hint in the prompt. The request to the model explicitly mentioned that the small model often repeats itself and does not finish its answer. The result without such a hint has not been tested yet.
• Knowledge instead of understanding. The model may be transferring known heuristics into the weights, or may have seen similar networks in training, rather than reasoning about these particular weights.
• An unfair comparison. Comparing with PPO, which learns from scratch, means comparing different tasks. The authors themselves call the work a demonstration and are preparing a benchmark and a harness for systematic testing.

What it means in practice

For small models that companies run inside their own perimeter, this is potentially a cheap way to fix known behavior defects — such as looping — without collecting data or fine-tuning. For now it is a research demonstration, not a tool: we are waiting for the benchmark and independent checks on other models.