Weight-space programming for tractor guidance controller: large language models natively write neural network weights that steer a real tractor
Organizations: Chair of Agrimechatronics, Technical University of Munich (TUM), Germany · Munich Institute of Robotics and Machine Intelligence (MIRMI), Technical University of Munich (TUM), Germany · Aalto University, Department of Electrical Engineering and Automation, Finland
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.
Figures & tables
| Era | Who writes the controller | Form | Examples |
| 1 | Human | Control law, few hand-tuned parameters | Pure pursuit ( Coulter, 1992 ) , PID, LQR ( O’Connor, 1997 ) |
| 2 | Human and numerical optimiser | Model and cost, action computed online | NMPC ( Backman et al., 2012 ; Soitinaho and Oksanen, 2023 ) |
| 3 | Numerical optimiser | Network weights from data or reward | Deep reinforcement learning ( Baby et al., 2024 ; Pindl et al., 2024 ) |
| 4 | Large language model | Network weights from understanding, no gradients | This paper |
| Parameter | Value |
| Plant: AMX G-trac simulation | |
| Model | kinematic bicycle, reference point on the rear axle |
| Wheelbase | 2.42 m |
| Speed | 1 m/s |
| Time step | 0.05 s |
| Steering-rate command, | 15 ∘ /s |
| GPT-6 Astra | Claude Fable 5.1 | |
| Vendor | OpenAI | Anthropic |
| Model identifier | gpt-6-astra | claude-fable-5-1 |
| Interface | Responses API | Messages API, streaming |
| Reasoning | effort high, summary auto | thinking always on, effort high |
| Output tokens per turn, maximum | 100 000 | 128 000; 64 000 in turn 1 of S1; lower in late turns of S0 (cost cap) |
| Conversation state | chained on the server | full conversation re-sent each turn, prompt caching |
| S1 Astra | S1 Fable | S0 Astra | S0 Fable | |
| Turns | 25 | 25 | 25 | 25 |
| Valid proposals | 24 | 24 | 22 | 23 |
| API time, total [s] | 934 | 2512 | 1676 | 4157 |
| API time per turn, median [s] | 25 | 62 | 53 | 56 |
| API time, first turn [s] | 168 | 868 | 252 | 1435 |
| Output tokens [thousand] | 70 | 197 | 78 | 322 |
| Controller | Budget | Failures | Max. | Non-zero | |
| Start points | |||||
| S1 start | – | 190.53 | 352 | 204 | 252 |
| S0 start (zero weights) | – | 268.09 | 512 | 204 | 0 |
| Repair of S1 (zero-shot: RQ1) | |||||
| AS-1 (Astra) | 25 | 7.82 | 0 | 87 | 220 |
| FA-1 (Fable) | 25 | 9.36 | 1 | 201 | 25 |
| Controller | [Hz] | |||||
| S1 start | 0.120 | |||||
| AS-1 (Astra, repair) | 0.187 | |||||
| FA-1 (Fable, repair) | 0.072 | |||||
| AS-0 (Astra, from zero) | 0.141 | |||||
| FA-0 (Fable, from zero) | 0.119 | |||||
| GD-T (PPO from scratch) | 0.059 |
| Element | Weight-space programming | Weight-Space Engineering | This study |
| Core act | An LLM reads, explains, repairs or writes weights directly | The same act, as one step of a process | The LLMs read the network, repaired S1 and wrote S0, without gradients or code execution |
| Domain | Any task: games, text, images, data | A technical system with physical consequences, e.g. a feedback controller | The steering loop of a tractor with actuator dead time and lag |
| Input from the human | A task | A specification and the knowledge of the field; not the meaning of single weights | Plant, objective and limits in the prompt; control theory as the field knowledge; no hint on the direction of the edit |
| Cost | Rarely stated | Explicit costs (evaluations, tokens, time) against the alternatives | Evaluation budget of 25 per session; tokens and time per session (Table 4 ); compared with equal-budget search and with PPO |
| Acceptance and lifecycle | A working demonstration | Verification and validation with the theory of the field, a test on the machine, deployment and maintenance, traceable reasoning | Unseen test paths; exact linearised closed loop; stated control law checked against measured gains; retrofit to a slower actuator (S1); test on the real tractor (Section 5 ) |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Repair of S1 | From zero | ||||||
| Unit | Input | S1 | AS-1 | FA-1 | GD-1 | AS-0 | FA-0 |
| 2.0505 | 2.051 | 0 | 2.1852 | 0.6 | 1 | ||
| 0.7088 | 0.2481 | 1 | 0.5242 | 0 | 0 | ||
| 0.1233 | 0.1233 | 0 | 0.1541 | 0 | 0 | ||
| 0.7771 | 0.7771 | 0 | 0.7798 | 0 | 0 | ||
| 0.3802 | 0.6463 | 0 | 0.3663 | 0.2 | 0 | ||