AVR-Chat

>
>as for the successive approximation, would the accuracy be higher if 
>the process were relatively steady and the change between readings 
>were rather small ?

Depends..
I used the M8 and M128 ADCs in a thermal printer, and I needed samples taken at a specific time, about 100uS after the burn starts. This is because the huge current draw pulls the battery voltage down during the print.  You can't calculate the burn energy accurately, unless you know the voltage during the burn.

For other inputs like the temperature, I just sampled them whenever convenient, and heavily filtered them in analog.

>Temperature is a very slow moving process so one would not expect to 
>see a process swing of more than a few percent over a few minutes.  
>So, would that mean internal ADC readings would be pretty accurate ?

The internal ADC is accurate, in that it reads out whatever it was when you sampled. 
The trick is that your voltage may have "hair" on it, that either is, or is not, important. 

Where it isn't important, filter it away, or you may read the peak, or valley of the hair, and mistake that for an average value. 

Where it is important, don't filter much at all, and sample at exactly the right time to read the hair you want.