Trajectory have generally impressed me with their tasteful execution on ambitious goals. on our Continual Learning track, @rronak_ gave a very thoughtful overview on how they're tackling the main data problems left in CL, including why GRPO isn't enough and they had to go on-policy.... and then subsequently fix all the issues that come up with it
nice overview from one of the early leaders in this field!
(see the rest of the track for more, this one was quite stacked)
It's time to rethink RL. Translating real world use into model improvements requires redesigning post-training algorithms for non-verifiable, per token rewards....