Logits

Logits are the raw real-valued scores a classifier produces before normalisation. They can be any real number. Passing them through softmax turns them into a probability distribution; for a binary sigmoid output, the logit is the log-odds log(p / (1 − p)), which is where the name comes from.

Only the differences between logits matter to softmax. Scores of (2, 0) and (102, 100) give the same probabilities. The spread matters a great deal, though: the wider apart the logits, the more peaked the distribution. That is why dividing logits by a temperature is enough to soften or sharpen a model's confidence (see temperature scaling).

Logits in Laya

Laya's decision head produces one logit per option. The encoder reads the whole sequence; the head adds a question-type embedding, runs two extra transformer layers, gathers the hidden state at each option marker and passes it through a small scorer (LayerNorm, linear, GELU, linear) that outputs a single number. Padding positions are masked to -10,000 so they never receive probability.

Three things happen to the logits at inference time:

1. They are divided by the clamped temperature for the question's (type, option-count) bucket.
2. They are softmaxed into the probabilities you see in the response.
3. During training only, RLCD adds zero-mean Gaussian exploration noise to them to sample alternative reports, and scores those with a proper rule.

The API does not return logits, only probabilities and derived fields. If you self-host the open weights you can read them directly, which is useful for fitting your own temperatures. See option-marker scoring.