Work out softmax probability instantly with clear inputs, formula shown and shareable results.
Softmax exponentiates each logit and normalises by the sum, turning arbitrary real scores into a probability distribution. Subtracting the maximum logit first is the standard trick that prevents overflow. Temperature divides the logits before exponentiation: below 1 the distribution sharpens toward the argmax, above 1 it flattens toward uniform, which is exactly how sampling temperature controls generation diversity.
Softmax with temperature
p_i = exp((z_i - max z) / T) / sum_j exp((z_j - max z) / T); entropy = -sum p_i log2(p_i)
In the limit it becomes greedy argmax decoding: all probability collapses onto the largest logit. Implementations special-case it rather than dividing by zero.
Entropy summarises how uncertain the distribution is in bits. A near-zero entropy means a confident prediction; entropy close to log2(classes) means the model is guessing.