I have some useful background here that leads me to strongly disagree with you. I did well at these international science olympiads in high school. To do that I studied for the particular shape of the exam questions. I became friends with many other people who were high performers at these olympiads. Later on I got a PhD from a prestigious institution and I am now a scientist at another one. Throughout this time, I was very involved in the pedagogical work at these institutions.
One very important piece of insight I gained throughout that time is that after you use some exam to get the top 25% (ish) percent of students, the exam performance really does not matter: it does not correlate with scientific output or originality of work. If anything, I had to unlearn some of the skills that made me good at olympiads because they were severely limiting my creativity. At an olympiad you know there is a solution, while true research problems might be unsolvable and need to be approached differently.
TL;DR: Exams (selective or not, hard or not) are great at giving you the top 20% of students, but they are inherently terrible at telling you who among the top 20% will be a productive scientist or engineer.
Your anecdotal evidence is contradicted by wider evidence.
In particular the famous Benbow study [1] that established that even among the top percentile of the population there are massive differences: the top quarter percentile (99.75-99.99) was 2-3 times more likely to have authored academic research later in life, and about 1.5x more likely to have attended postgrad education, to have gone for a STEM degree, to have gone for a PhD than the bottom quarter of the top percent (99.00-99.24).
To update on the sibling comment I posted, after reading this study, I think it is pretty flawed. They indeed have an interesting population they have chosen to test, but to claim that they are the 1% top performers is incredibly misleading. The other 99% of their peers were never tested in the same fashion, which leads to obvious sampling and bias problems. This special population was treated differently from the other 99% of students from the very start of the study, so there is no surprise that new internal dynamics will appear among them, simply by virtue of being observed and treated as a separate group.
To start testing the validity of the claims in this paper a study needs to be performed where the strength of effect has to be considered, when comparing a similar one-percenter group and a larger ten-percenter group. My prediction is that the strength of effect would show only negligible differences, confirming my hunch that the special treatment[1] is what created the new subdivision in this new group.
More reading material on this topic would certainly be interesting, if you have anything in mind.
[1]: The special treatment in this paper being telling a kid "you are only a 7th grader, but you are as smart as a high schooler".
This is interesting, thanks for sharing it! It would take me some time to read it, but given that you seem the have already looked into this, I was wondering whether you have suggestion for a wider set of studies, a meta-study maybe.
I am asking, because if I have to choose between my empirical observations (anecdotal as they are, given it is n=1 observers), and a single piece of research without followups, I do feel justified to stick to what I have seen myself. But I am open to be convinced otherwise.
One very important piece of insight I gained throughout that time is that after you use some exam to get the top 25% (ish) percent of students, the exam performance really does not matter: it does not correlate with scientific output or originality of work. If anything, I had to unlearn some of the skills that made me good at olympiads because they were severely limiting my creativity. At an olympiad you know there is a solution, while true research problems might be unsolvable and need to be approached differently.
TL;DR: Exams (selective or not, hard or not) are great at giving you the top 20% of students, but they are inherently terrible at telling you who among the top 20% will be a productive scientist or engineer.