Concepts

These pages are explanation: how Laya Vision works and why it is designed the way it is. They are for understanding, not step-by-step instructions (see the How-to guides for those).

The big picture

Laya Vision turns "ask a vision-language model a question" into "score the options you wrote". The image, the state text, the question and its options go through the backbone once; a small head reads one logit per option; a per-type temperature turns those logits into probabilities. Nothing is generated, so the cost is one forward pass per batch of questions, whatever the answers are.

Explanations

  • How it works: the sequence, the per-option readout, and how causal and bidirectional backbones differ.
  • Architecture: every backbone variant and the shared head, with diagrams, the training objective, the shared-prefix cache and the checkpoint layout.
  • Calibration: what the temperatures do, why they are fitted after training, and what refitting them can and cannot change.
  • Training data: the prepared datasets, from The Cauldron to the held-out evaluation sets.
  • Training data for score questions: the four rubric-scored datasets behind the score head, and how they were cleaned.
  • Caching the fixed question in game play: what caching the constant question could save, and the CUDA-graph path that was built instead.