See an interesting discussion on benefits and shortcomings of “least used resource” mode of load balancing at High Scalability.
Here is what I think:
When dealing with TCP-based services (say http), my solution is not to use this load balancing mode at all precisely for this reason (you bring a new server online - and all of a sudden all connections are going to this new server). When backend servers change their state very fast, load balancers are always trying to catch up. The higher the load, the more likely they will fall behind - if that happens, they will be distributing connections based on state information from some point in the past (and you never know how far in the past). Round robin (possibly weighted) is much simpler and more reliable, imho.