@riverdong.bsky.social
⚖️ Personalization Can Protect Minority Viewpoints!
In diverse-user settings, personalization helps amplify underrepresented perspectives (User 8 in the Figure). Without personalization, models tend to default to majority opinions, sidelining minority viewpoints.
In diverse-user settings, personalization helps amplify underrepresented perspectives (User 8 in the Figure). Without personalization, models tend to default to majority opinions, sidelining minority viewpoints.
March 5, 2025 at 4:05 PM
⚖️ Personalization Can Protect Minority Viewpoints!
In diverse-user settings, personalization helps amplify underrepresented perspectives (User 8 in the Figure). Without personalization, models tend to default to majority opinions, sidelining minority viewpoints.
In diverse-user settings, personalization helps amplify underrepresented perspectives (User 8 in the Figure). Without personalization, models tend to default to majority opinions, sidelining minority viewpoints.
⚠️ Personalization can hurt model safety & reasoning by up to 30%.
March 5, 2025 at 4:04 PM
⚠️ Personalization can hurt model safety & reasoning by up to 30%.
📊 Key Findings:
(1) Performance can vary by up to 36%
(2) Fine-tuning per user is a strong baseline
(3) For the recently proposed algorithms: Personalized Reward Modeling (PRM) achieves best performance. Group Preference Optimization (GPO) show fast adaptation to new users.
(1) Performance can vary by up to 36%
(2) Fine-tuning per user is a strong baseline
(3) For the recently proposed algorithms: Personalized Reward Modeling (PRM) achieves best performance. Group Preference Optimization (GPO) show fast adaptation to new users.
March 5, 2025 at 4:04 PM
📊 Key Findings:
(1) Performance can vary by up to 36%
(2) Fine-tuning per user is a strong baseline
(3) For the recently proposed algorithms: Personalized Reward Modeling (PRM) achieves best performance. Group Preference Optimization (GPO) show fast adaptation to new users.
(1) Performance can vary by up to 36%
(2) Fine-tuning per user is a strong baseline
(3) For the recently proposed algorithms: Personalized Reward Modeling (PRM) achieves best performance. Group Preference Optimization (GPO) show fast adaptation to new users.