Pseudo-label
You have many training examples, but nobody supplied correct answers. You ask the model to estimate them and use those estimates as guidance for further learning. What actually enters the data then?
A pseudo-label is a label obtained automatically, for example through a model prediction or a vote among several responses. A label is the result on which learning is based. “Pseudo” reminds us it is not an independently verified answer key.
In an illustrative vote, eight solutions give A three times, B twice, and three different other answers. A appears most often, although no majority supports it. Choosing the largest group is called plurality. Frequency does not prove correctness.
TTPO, §3.1 and §4, uses the selected result to divide responses among different training procedures. This is a specific way of using pseudo-labels, rather than a general property of the term. The main risk is simple: training on the model's own wrong answers as truth can reinforce errors.