The Wigner's friend thought experiment can be used to understand non-realism in quantum mechanics. For anyone not familiar, the thought experiment involves two researchers observing an experiment at different times, let's say it's an electron spin. The question is, how should the second researcher to observe the experiment treat the system between his observation and his friend's observation?

In a non-realist interpretation we would say that the second researcher should continue to time evolve the system with the Schrodinger equation until he observes the system himself, but now he must include his friend as part of the quantum system as well. So, the two observers see the collapse of the wavefunctions at different times, which is fine because the time of wavefunction collapse is not given by a linear Hermitian operator so they need not agree about it.

My question concerns what happens if the additional time evolution that the second researcher observes changes the probability. For instance, say there is a 50% probability of spin up when the first researcher observes the experiment. Can this probability change from the second observer's perspective in the time between observations? If it can, how do we explain the fact that if they repeat the experiment 100 times the first researcher would expect to see spin up 50 times and the second researcher would expect to see something else. If it can't, what is the purpose of the additional unitary time evolution that the second researcher uses to describe the situation, couldn't he just use a description where the wavefunction collapses when the first researcher observes the experiment and get the same answer?

Once the first researcher measures the system, the state of the system becomes entangled with the state of it's environment (which includes the first researcher). The information of the measurement result propagates very quickly into the environment (photons from the measurement device spread, etc.), in a process called decoherence, creating a redundancy in which many copies of the result are imprinted in all that surrounds. The total state of the system plus the environment can be effectively described as (assuming the system is a single spin like in your question): $$\frac{1}{\sqrt{2}} \left( |\uparrow \rangle |\text{environment after spin up result} \rangle + |\downarrow \rangle |\text{environment after spin down result} \rangle \right)$$ The second researcher, which doesn't know the outcome of the measurement yet, would now describe the system as a mixed state, with the density matrix: $$\rho= \frac{1}{2} \left( |\uparrow \rangle \langle \uparrow| + |\uparrow \rangle \langle \uparrow| \right)$$ The same goes with how the second researcher would describe the first researcher after the measurement: in a mixed state, and not in a coherent superposition of having measured up and having measured down. This is because the information of the result has many copies beside the one in the first researcher's brain, and any such copy will be part of the big state I gave of the system plus the environment. Any description that includes only part of all the environment that was affected by the measurement, will necessarily be of a mixed state and not of a superposition.