Inherent researcher @DamonFalck on why the industry is wrong that GRPO needs verifiable tasks: they made it work on paper replication with a hindsight based rubric judge that scores individual turns "Everyone's trying to do GRPO RL for longer and longer horizon tasks, and they're running into problems. The solution has been, well, we need to constrain to verifiable tasks." "We decided that really wasn't the route we're taking, so we've had to come up with a bunch of innovations to make this…