Reinforcement Learning and the Language Bottleneck

The language bottleneck can help potentially like reduce these spurious correlations that the model picks up on exactly as well as well and then that would help with generalization yeah. In terms of policy learning and reinforcement learning you did apply this ideaYeah I mean honestly the reinforcement learning experiments in that paper are kind of a gross hack and I look forward to being able to point to something better to replace them with but the story there in reinforcement learning right is that it's not like the form of the experience that you get is here's an example of the agent going to its goal figure out what it was trying to do and then do it again because if that were the case you could just memorize the policy and