FWIW,
I used a multi-plexing chip with several known resistances attached to calibrate Thermistor outputs. The manufacturer of a quality Thermistor (US Sensors, et al) publishes resistances for their thermocouples across a wide temperature range. So, once you know what the ADC should be producing for a given resistance, you can use a least squares approach to figuring out how to compensate for offset and gain errors (I happen to use five precision resistors to calibrate with). For a computational approach, do a search for Least Squares here.
Once ADC outputs have been corrected, plug the corrected results into a lookup / map table with data you've previously calculated (I'm planning on using flash memory for this) and from what I can tell, the Arduino should be able to get thermal accuracy to within 0.2C while using a nominally +/- 0.1C accurate thermistor. The resolution is significantly higher than that, so Delta-T calculations should be significantly more accurate. For now, the line fit shows R^2 to be around 0.99988, which is pretty close to 1.