핵심 개념
OPEN framework combines LMs and BOED to optimize user preferences efficiently.
초록
Introduction emphasizes the importance of understanding user preferences for automation tasks.
Challenges in preference learning include quantifying uncertainty and asking informative questions.
OPEN framework integrates LMs and BOED for optimal preference elicitation.
Featurization, initializing user preferences, selecting optimal questions, verbalizing queries, and updating preferences are key steps in OPEN.
Experimental setup includes baselines, hyperparameters, and human participant details.
Evaluation metrics include TIDA scores comparing OPEN with different question generation and prediction methods.
Results show OPEN outperforms LM-only approaches in predicting and eliciting user preferences.
Qualitative analysis highlights user feedback on different elicitation methods.
Future work includes exploring other preference-learning domains and incorporating open-ended questions.
Ethical considerations focus on aligning AI systems with user values and collecting preference data responsibly.
User study conducted on Prolific platform with positive feedback from participants.
Reproducibility statement ensures codebase and anonymized data will be released on GitHub.
통계
최근, 언어 모델(LMs)를 사용하여 인간의 선호도에 대한 정보를 수집하는 데 관심이 증가했습니다.
OPEN은 기존의 LM 및 BOED 기반 선호도 추출 방법을 능가합니다.
OPEN은 사용자 선호도를 최적화하기 위해 LMs와 BOED를 결합합니다.
인용구
"OPEN can optimize the informativity of queries while remaining adaptable to real-world domains."
"OPEN outperforms both LM- and BOED-based preference elicitation approaches."
"Understanding user preferences in underspecified environments is crucial to avoiding real-world issues with AI systems."