Language models write neural network weights that steer a real tractor
Researchers report that large language models directly edited and wrote the weights of a small steering controller, repairing a trained network and producing controllers that ran on a physical tractor without intervention.
In plain words
A neural network is a controller whose behaviour is set by numbers called weights. Instead of training those weights with an optimiser, the researchers had large language models read and rewrite them directly from a plain-language specification. The resulting controllers steered a real tractor along line and arc paths.
What was announced
The study, an arXiv abstract, describes weight-space programming applied to automatic tractor guidance: a 253-parameter network steers along paths of lines and arcs and must be adapted when the machine changes, for example with a slower steering actuator. GPT-6 Astra and Claude Fable 5.1 (also 5.2) edited the network for the authors' research tractor using a small budget of simulated evaluations, repairing a network trained for a different tractor and writing a controller from zero weights without an optimiser or gradient descent.
In simulation and on the real tractor, the models were compared with an evolution strategy, gradient descent and pure pursuit. Both models read the network and correctly analysed the closed loop with control theory, though Fable's repair without simulated feedback succeeded only occasionally. With simulated feedback, both repaired the trained network in single sessions and wrote working controllers from zero weights, while the evolution strategy found none from zero. The from-zero controllers were sparse and classically cascaded; one model edited the trained network while the other replaced it. Gradient descent reached higher accuracy with far more simulation, and the language-model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit. The authors call this Weight-Space Engineering.
What it does not show
- The evidence is an abstract only; full experimental details, metrics and peer review are not available in the supplied passage.
- Results rest on the authors' own simulations and a single research tractor; no independent confirmation is reported.
- Gradient descent achieved higher accuracy with more simulation, so the reported advantage is evaluation efficiency, not accuracy.
For security and IT teams
- Scope
- The reported work covers a 253-parameter guidance controller for line-and-arc paths on one research tractor, not a general-purpose method validated across platforms.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.