It has been mentioned before but surely it is strange that Atmel never had given any good recommendation of the values of the capacitor at AREF and the inductor to AVCC. 10µH is mentioned but not a word about DC resistance.
Read this text from the datasheet of the linear regulator TPS73028 and you will understand how much there is to know about putting in a correct value of the AREF capacitor:
The internal voltage reference is a key source of noise in an LDO regulator. The TPS730xx has an NR pin which is connected to the voltage reference through a 250kΩ internal resistor. The 250kΩ internal resistor, in conjunction with an external bypass capacitor connected to the NR pin, creates a low pass filter to reduce the voltage reference noise and, therefore, the noise at the regulator output. In order for the regulator to operate properly, the current flow out of the NR pin must be at a minimum, because any leakage current creates an IR drop across the internal resistor thus creating an output error. Therefore, the bypass capacitor must have minimal leakage current. The bypass capacitor should be no more than 0.1μF to ensure that it is fully charged
during the quickstart time provided by the internal switch shown in the Functional Block Diagrams.
I think the topic about the output current should be the same for the AVR Vref so from that we can learn that a low leakage capacitor would be to prefere if we need to minimize the error.
The capacitor value is more complicated to figure out as we know nothing about the internal circitry in AVR so if there is something similar like the 250k resistor we can only guess. A too small capacitor will result in unnecessary noise. A too large capacitor will mean it will take longer time to charge (and decharge in case of swithing to another reference voltage). Other resistors in the AVR like reset and port pull-up are 20-60 kohm so if there are anything like a resistor in the Vref circuitry it MAY be something like this. 100 nF is usually recommended but most probably there is not any problem to increase this value but 100nF could just as well be the perfect value.