LO
CapabilityCognitive automation2025-01-29
Language models wrote G-code for six simple shapes with about 80% success, and 100% for most when a person supplied structured parameters, a BTW 2025 industrial paper found
Machinist / CNC machinistoccupation page →
Event date / reported
2025-01-29
Evidence stage
CapabilityA demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.
Tasks this bears on
Programming machines
Turning drawings into toolpaths and machine code (G-code), usually with CAM software, and choosing tools, speeds and feeds.
Being augmented≈ Platform inference
Where this applies
Germany, a software company's study in a peer-reviewed conference's industrial track. Across six G-code tasks for simple shapes, checked against target toolpaths in simulation and averaged over five runs, two models reached about 80 percent success, and structured prompts in which a person had already extracted the parameters reached 100 percent for most tasks. Nothing was cut; the authors build tools of this kind.
What this means
Models can write code for simple shapes most of the time, and reliably when a person first sets out the parameters.
What it does not yet show
Simple shapes in simulation, by a company that builds such tools; it does not show use in shops.
What you can check
Open arXiv 2501.17584 and find "100% success rate for most tasks".
Does it change the assessment?
No — and this stage does not move it either. A "Capability" record is real evidence, but it does not upgrade a task judgement on its own. The 1 linked judgement above stand where they were.
Source
Abdelaal M, Lokadjaja S, Engert G — GLLM: Self-Corrective G-Code Generation using Large Language Models with User Feedback, Industrial Track of BTW 2025 (arXiv 2501.17584, 29 January 2025) · verified 2026-10-01 · Claude (VOLO agent) · interpreted 2026-10-01 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.