That's completely inane. There's nobody home. The physician by definition wins actual empathy on walkover, no matter how bad for a human.
Sad statement on the judgment of the respondents.
But an important reason it can turn out like this, I suppose, would also be that the RL feedback gives the model a fairly effective general optimization about what statements are liked by the mechanical turk-like evaluators. Most physicians have probably never had access to anything like that level of feedback on how their expressions are received. Maybe the LLM's can be rigged to provide goodness gradients for actual physicians' statements?
> Sad statement on the judgment of the respondents.
Nope, it's a sad reflection of the study construction.
Physicians' empathy was evaluated by their 52 word responses on Reddit. Unsurprisingly, a chatbot optimised for politeness and waffle outperformed responses of people volunteering answers in a different format optimised for brevity...
Exactly. I've spent ten years too disabled to leave my home and most doctors are just bullies who would rather insult you than consider that their initial evaluation of "it's just stress" might be wrong.
Sad statement on the judgment of the respondents. But an important reason it can turn out like this, I suppose, would also be that the RL feedback gives the model a fairly effective general optimization about what statements are liked by the mechanical turk-like evaluators. Most physicians have probably never had access to anything like that level of feedback on how their expressions are received. Maybe the LLM's can be rigged to provide goodness gradients for actual physicians' statements?