Let’s look at an example. I’m going to use a sample rate of 50 throughout this post (i.e. the functions are sampled at 0, 0.02, 0.04, 0.06, …) and simple sine waves of the form $f_a(t) = \sin (2\pi\cdot a\cdot t)$, where $a$ is the frequency of the sine wave. For example, this is a plot of $f_3$:

The red impulses indicate the sample values. The green line, on the other hand, is the continuous function that is being sampled. The sample values all lie on this curve, and it goes through three complete cycles, as one would expect.
Aliasing
Because the sample rate is 50, anything above the Nyquist frequency of 25 is going to alias. As an example, here is a plot of $f_{53}$; mind the axes, I zoomed in to one tenth of the above, because otherwise the image would have become quite crowded.

There is slightly less than one sample per cycle, so the sampled values don’t follow the continuous waveform very closely. Actually, if all you saw were the sample values without the green curve, like so …

… you might be tempted to draw a much slower wave. Here is what happens when we plot not only $f_{53}$, but $f_3$ at the same time:

While the two sine waves are quite dissimilar in appearance, they happen to coincide at each and every sample point. There is absolutely no way to tell them apart given only the sample values. That’s aliasing. In fact, not only the samples of $f_{53}$ are exactly the same as those of $f_3$, but $f_{103}$, $f_{153}$, $f_{-47}$ etc. alias in the same way.
Put differently, due to sampling at a frequency of 50, the ranges 0–25, 50–75, 100–125, etc. collapse. What about the “upper half”, from the Nyquist frequency to the sampling frequency? It is mirrored; $f_{47}$ is the same as $-f_3$, and in general the frequency range 25–50 aliases to 25–0, with the samples inverted (or, equivalently, a phase shift by $\pi$).
This means, for your question, that the effect you are observing is not due to you using signals that exceed the Nyquist frequency; you can replace any higher-than-Nyquist-frequency signal with its alias in the Nyquist range and will get exactly the same results.
The “beating”
So what causes the apparent “beating”? Here is the very first plot of $f_3$ again, but this time it shows the sample values and linear interpolation between them (i.e. straight line segments from each sample point to the next), not the actual continuous waveform that was sampled:

The interpolated waveform is a bit edgy, but overall a good approximation of the original sine wave. Unfortunately, this is only because the frequency is very low relative to the sample rate. Here is $f_7$:

Urgh. It worsens as we get to higher frequencies, closer to the Nyquist frequency; this is $f_{23}$:

Unfortunately, even if we don’t draw those straight lines, our eye will try to connect the dots all on its own, and the impression is similar: It looks like a slow oscillation modulated onto a fast one. And that impression actually isn’t completely beside the point:

The same sample values, from $f_{23}$, can equally well explained by a ring modulation of $f_2$ and $f_{25}$ (with some phase adjustment), namely, $g(t) = -f_2(t)\cdot f_{25}(t+0.01)$. Thus, $g$ and $f_{23}$ are aliases, when sampled at a rate of 50. Here is a close-up again, showing that $g$ and $f_{23}$ in fact meet at 0, 0.02, 0.04 etc.:

A different way to write $g$ is $g(t) = 0.5 f_{23}(t) - 0.5 f_{27}(t)$, a sum of two sine waves of similar frequency, exactly the case where beating could be expected to happen. So it is beating after all, isn’t it? Not really, because $f_{27}$ is beyond the Nyquist frequency; as outlined above, it is an alias of $-f_{50-27}=-f_{23}$, and so $g(t)$ becomes $0.5 f_{23}(t) - 0.5 f_{27}(t) = 0.5 f_{23}(t) + 0.5 f_{23}(t) = f_{23}(t)$.
This is, within the baseband from 0 to 25, the only possible interpretation of those sample values.
Conclusion
Watch out for improper interpolation! When reconstructing a continuous signal from samples, the limited bandwidth must be taken into account (band-limited interpolation). Unfortunately, our eyes are not good at this; and many software waveform editors and resamplers aren’t, either.