Yes, we agree that it should not be taken literally, as the game might have some confounders that affect the performance in the game and that might be different in the real world. We do not yet know the size of this effect, and we will do additional experiments to shed more light on it in the revision.
As for the GTA analogy, we respectfully disagree: we are not training the model on any objective, but evaluating it in an RL environment similar to its training environment. We do this in a multi-agent setting to make it more realistic. Nothing in the game rewards a kill. In an ideal world, models should also show compassion in these scenarios. Most evaluation scenarios do not reflect the real world, but we still think their results provide useful evidence.
Yes, we agree that it should not be taken literally, as the game might have some confounders that affect the performance in the game and that might be different in the real world. We do not yet know the size of this effect, and we will do additional experiments to shed more light on it in the revision.
As for the GTA analogy, we respectfully disagree: we are not training the model on any objective, but evaluating it in an RL environment similar to its training environment. We do this in a multi-agent setting to make it more realistic. Nothing in the game rewards a kill. In an ideal world, models should also show compassion in these scenarios. Most evaluation scenarios do not reflect the real world, but we still think their results provide useful evidence.