The GenToka paper builds a shared model for measurements from a tokamak. The authors use the same backbone for tasks that request different outputs, including a time window of a signal, a plasma profile, a field or a scalar. Their evaluation uses records from MAST. The preprint was submitted to arXiv on November 5, 2026.
How the model handles different outputs
Each task has a schema defining the signals that are available, the outputs requested and the time periods each covers. The model encodes available diagnostics and actuator measurements as context. It represents each requested output with a noisy target token and uses a shared conditional diffusion Transformer to denoise the targets. Output heads then map the representation back to the required signal format.
Pretraining reconstructs signals recorded at the same time while randomly masking parts of the context. The authors then adapt the shared backbone and task-specific interfaces to the inputs and outputs each downstream task requires.
What the MAST comparison found
The benchmark has 20 tasks in six groups. Relative to TokaMind, the authors report a 4.7% reduction in macro-average normalized root mean squared error across the first four groups and a 17.7% reduction in Group 5, which addresses actuator recovery. For the disruption group they report an 8.4% increase in macro-average window-level F1 at a common threshold selected on validation data.
How to read the benchmark
These are aggregate comparisons within the paper’s MAST evaluation. The task definitions specify what information a model can use and which part of the discharge it must reconstruct or predict. Those details matter to anyone reusing the benchmark. A prediction from earlier observations, recovery of a control signal and a disruption score answer different operational questions. The paper reports those groups separately, giving readers a way to match the results to the kind of output they need.
Comments