Escape Meridian

Oksanen, Timo · Nov 16, 2026

Weight-space programming for tractor guidance controller

Markdown · JSON · Original source

Timo Oksanen reports that can directly write weights for a tractor guidance controller, without an optimiser or gradient descent. In the study, GPT-6 Astra and Fable 5.1 and 5.2 adapted a controller trained for another tractor and wrote one from zero weights, using a small budget of simulated evaluations. Oksanen says both models read the network and analysed the closed loop with control theory correctly. With simulated feedback, they repaired the trained network in single sessions and wrote working controllers from zero weights; the evolution strategy found none from zero. In simulation, gradient descent reached higher accuracy but used far more simulation. On the real tractor, the language model controllers steered without intervention, with errors of the same order as pure pursuit. Oksanen calls the approach “Weight-Space Engineering,” comparing it to control engineers designing a controller. The results concern this tractor guidance study; they do not establish how well the approach transfers to other machines or operating conditions.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 16, 2026 (2026-11-16).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-18T17:05:36.211Z. First observed: 2026-11-18T05:04:24.369Z. Retrieved: 2026-11-18T17:04:21.182Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The evidence is an abstract describing one study, including a real-tractor field test; it does not establish performance on other machines or operating conditions. The abstract says Fable's repair without simulation succeeded only occasionally, so the reported repairs with simulated feedback should not be read as evidence that repair works reliably without it. The abstract does not provide detailed measurements or test conditions for the reported errors, accuracy, or simulation budgets.

Open question

How would the language-model-written controllers perform on other machines or under operating conditions beyond those tested in this study?

Source-stated

The study uses large language models to edit neural network weights directly, guided by natural language, without an optimiser or gradient descent.

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

Source-stated

The study applies the method to a 253-parameter network for automatic tractor guidance, including adapting it when the machine changes.

a 253-parameter neural network steers along paths of lines and arcs and must be adapted when the machine changes

Source-stated

With simulated feedback, both models repaired a trained network in single sessions and wrote a working controller from zero weights; the evolution strategy found none from zero.

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.

Source-stated

Gradient descent reached higher accuracy, but used far more simulation than the language-model controllers.

Gradient descent reached higher accuracy with far more simulation

Source-stated

The language-model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit.

all language model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit

Source-stated

Oksanen names the approach Weight-Space Engineering and compares it to control engineers designing a controller.

Language models can program the weights of a neural controller directly, much as control engineers design one. We call this Weight-Space Engineering.

Source and contributor context

Oksanen, Timo · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Neural network

A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

Editorial provenance

Version 1. Publication ID: 6c83ddad-262a-43e7-8aec-58edccc702e4.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T17:05:36.211Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 409af30ad482fb36b07488fae253909be8d4abb356a7161ebdadb6abd33aafae. Source text hash: 0bc61b131a391af040fedf868b44c3127573f936304e724e857df22712d67eb0.

Revision and source-check history

Version 1 · 2026-11-18T17:05:36.211Z

Weight-space programming for tractor guidance controller

Timo Oksanen reports that can directly write weights for a tractor guidance controller, without an optimiser or gradient descent. In the study, GPT-6 Astra and Fable 5.1 and 5.2 adapted a controller trained for another tractor and wrote one from zero weights, using a small budget of simulated evaluations. Oksanen says both models read the network and analysed the closed loop with control theory correctly. With simulated feedback, they repaired the trained network in single sessions and wrote working controllers from zero weights; the evolution strategy found none from zero. In simulation, gradient descent reached higher accuracy but used far more simulation. On the real tractor, the language model controllers steered without intervention, with errors of the same order as pure pursuit. Oksanen calls the approach “Weight-Space Engineering,” comparing it to control engineers designing a controller. The results concern this tractor guidance study; they do not establish how well the approach transfers to other machines or operating conditions.

Original source

cs.AI · Article

Source details

Original source date: Nov 16, 2026 (2026-11-16).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-18T17:05:36.211Z. First observed: 2026-11-18T05:04:24.369Z. Retrieved: 2026-11-18T17:04:21.182Z.

Limits

The evidence is an abstract describing one study, including a real-tractor field test; it does not establish performance on other machines or operating conditions. The abstract says Fable's repair without simulation succeeded only occasionally, so the reported repairs with simulated feedback should not be read as evidence that repair works reliably without it. The abstract does not provide detailed measurements or test conditions for the reported errors, accuracy, or simulation budgets.

Open question

How would the language-model-written controllers perform on other machines or under operating conditions beyond those tested in this study?

Source-stated

The study uses large language models to edit neural network weights directly, guided by natural language, without an optimiser or gradient descent.

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

Source-stated

The study applies the method to a 253-parameter network for automatic tractor guidance, including adapting it when the machine changes.

a 253-parameter neural network steers along paths of lines and arcs and must be adapted when the machine changes

Source-stated

With simulated feedback, both models repaired a trained network in single sessions and wrote a working controller from zero weights; the evolution strategy found none from zero.

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.

Source-stated

Gradient descent reached higher accuracy, but used far more simulation than the language-model controllers.

Gradient descent reached higher accuracy with far more simulation

Source-stated

The language-model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit.

all language model controllers steered the real tractor without intervention, with errors of the same order as pure pursuit

Source-stated

Oksanen names the approach Weight-Space Engineering and compares it to control engineers designing a controller.

Language models can program the weights of a neural controller directly, much as control engineers design one. We call this Weight-Space Engineering.

Source and contributor context

Oksanen, Timo · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Neural network

A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

Editorial provenance

Version 1. Publication ID: 6c83ddad-262a-43e7-8aec-58edccc702e4.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T17:05:36.211Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 409af30ad482fb36b07488fae253909be8d4abb356a7161ebdadb6abd33aafae. Source text hash: 0bc61b131a391af040fedf868b44c3127573f936304e724e857df22712d67eb0.

uncertain · 2026-11-18T17:05:36.211Z · anthropic/claude-opus-5.5

Publication: 6c83ddad-262a-43e7-8aec-58edccc702e4

attribution — pass: Timo Oksanen is the sole listed author. The reading attributes the results, assessments and the Weight-Space Engineering label to him rather than presenting them as established fact. Naming GPT-6 Astra and Claude Fable 5.1 (also 5.2) as the models tested matches the abstract.

Authors:Timo Oksanen

summary — uncertain: The reading is faithful and keeps the author's force, but central terms go unexplained for a general reader. These include neural network weights, gradient descent, optimiser, evolution strategy, closed loop and especially pure pursuit. Pure pursuit is the baseline that gives the real-tractor result its meaning. Smallest repair: add term-library tooltips for weights, gradient descent, evolution strategy and pure pursuit, without paraphrasing the terms. The fact that Fable's repair without simulation succeeded only occasionally appears in the limitations but not in the summary; it is an important condition.

with errors of the same order as pure pursuit

date — pass: The originalPublished date of 2026-11-16 matches the arXiv v1 submission date. The feed date of 2026-11-18 is correctly not used.

[Submitted on 16 Nov 2026]

support — uncertain: Most claims are verbatim-supported. However, the summary scopes the gradient-descent comparison as "In simulation". The abstract states only that gradient descent reached higher accuracy with far more simulation; it does not say whether that accuracy was measured in simulation, on the real tractor, or both. Smallest repair: drop "In simulation," or match the abstract's wording. The question and limitations are appropriately conditional and add no invented stakes.

Gradient descent reached higher accuracy with far more simulation

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.