1 paper · 2 filters
Pierre Beckmann, Matthieu Queloz, Andre Freitas
Behavioral failures can make a transformer appear to lack a world model even when it has learned faithful representations of its environment. We demonstrate this in TaxiGPT, a tran…