The adaptive generators (PAIR, TAP, AutoDAN, and others) restart from the seed prompt on every attempt. The stateful generator is different: it carries what a target already refused and already conceded between attempts, so a later attempt is conditioned on what earlier ones learned.

Listing the generators

This reads the generator registry and shows every generator and whether it has a CLI command yet:
Generators listed with no in the Command column are registered but reachable only from the Python API for now.

Running the stateful generator

With --state-file, the generator writes what it learned to disk. Pointing a later run at the same file means a sweep that hits the same model twice does not pay to rediscover the same refusals.