This parameter tends to give RNA its "resolution". The value should not be large. Good values are generally 3 or 4 bps in general. On a few rare occasions, as large as 5 bps has worked, but such cases are quite rare and caution is strongly advised. When the average stem length can be considered longer, this can help speed up the program considerably because there is no need to search the short stems to find out if they are sufficiently contiguous. However, when too large a value is used for the average number of contiguous stems, one may end up filtering out meaningful information. Much of the problem is probably still our ignorance about appropriate values for stacking interactions in context. Currently little context is used other than the nearest neighbor contributions. In some cases it is both better, and faster to use a longer average stem length, but in most cases, it may only be "faster" and not really "better".
The minimum stem length is limited to a maximum value of 6 bp. Just select what you like, but keep in mind that small values give detail and large values do not.