In our study, we conducted two experiments assessing the effectiveness of our method: one using proxy human preferences and the other based on real human feedback.
Proxy human preferences, sourced from EUREKA, offered a quantitative evaluation of our approach, while ensuring that the LLM never observed the actual reward values.
Despite the reliability of proxy preferences in assessing our method, they may not effectively capture human difficulties such as ranking issues and intransitive preferences.
The second experiment involved real human subjects engaging in tasks without definitive reward functions, further testing the boundaries of our method's applicability.
Collection
[
|
...
]