跳到正文
arXiv cs.AI· Timo Oksanen·· 4 小时前AI 评分58

研究:GPT-6 Astra 与 Claude Fable 5.1 直接写入神经网络权重操控真实拖拉机

Weight-space programming for tractor guidance controller: large language models natively write neural network weights that steer a real tractor

AI 导读

一项 arXiv 研究将权重空间编程用于拖拉机自动导航:GPT-6 Astra 和 Claude Fable 5.1 在少量仿真评估预算下,直接读写一个 253 参数神经网络的权重,修复为另一台拖拉机训练的控制器,并从零权重写出可用控制器。

正文 · 原文

View PDF HTML (experimental)

Abstract:Weight-space programming is a process where a large language model reads and writes the weights of a neural network directly, guided by a natural language specification, without an optimiser or gradient descent. This study applies it to automatic tractor guidance, where a 253-parameter neural network steers along paths of lines and arcs and must be adapted when the machine changes, such as a slower steering actuator. GPT-6 Astra and Claude Fable 5.1 (also 5.2) edited the network for our research tractor using a small budget of simulated evaluations to repair a network trained for another tractor and write a controller into a network with zero weights. We compared them with an evolution strategy, gradient descent and pure pursuit in simulation and on the real tractor. Both models read the network and analysed the closed loop with control theory correctly, and Fable's repair without simulation succeeded only occasionally. With simulated feedback, both models repaired the trained network in single sessions and wrote a working controller from zero weights, whereas from zero the evolution strategy found none. The controllers written from zero were sparse and in a classical cascaded structure; one model edited the trained network while the other replaced it. Gradient descent reached higher accuracy with far more simulation, and all language model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit. Language models can program the weights of a neural controller directly, much as control engineers design one. We call this Weight-Space Engineering.
Comments: 20 pages, 8 figures, 8 tables (1 in the appendices); includes a field test on a real tractor
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO); Systems and Control (eess.SY)
Cite as: arXiv:2611.21958 [cs.LG]
  (or arXiv:2611.21958v1 [cs.LG] for this version)
  https://doi.org/10.48550/arXiv.2611.21958

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Timo Oksanen [view email]
[v1] Mon, 16 Nov 2026 22:39:25 UTC (4,671 KB)

来源:arXiv cs.AI · arxiv.org