Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Variational Preference Alignment with Gumbel-Softmax Prior for Personalized User Preferences
Jabin Koo, Hoyoung Kim, Minwoo Jang, Jungseul Ok
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedVPA-GP instead of FedBiscuit | Qwen-2 0.5B, 10 clients · HH-RLHF | Helpful win-rate: 48.85 -> 66.45 | Improved |
| Use FedVPA-GP instead of FedBiscuit | Gemma-2B, 10 clients · HH-RLHF | Harmless win-rate: 82.45 -> 96.34 | Improved |
| Add Federated Mixture Prior and orthogonal loss to FedVPL | Qwen-2 0.5B, 100 clients · HH-RLHF | Helpful win-rate: 53.05 -> 55.18 | Improved |
| Add Federated Mixture Prior and orthogonal loss to FedVPL | Gemma-2B, 10 clients · HH-RLHF | Helpful win-rate: 66.82 -> 73.21 | Improved |
| Increase clients from 10 to 100 | FedVPA-GP Qwen-2 0.5B · HH-RLHF | Helpful win-rate: 66.45 -> 55.18 | Hurt |
| Increase clients from 10 to 100 | FedVPL Gemma-2B · HH-RLHF | Harmless win-rate: 89.15 -> 80.42 | Hurt |
| Evaluate on unseen clients instead of seen clients | FedVPL · HH-RLHF unseen clients | Harmless win-rate: 83.82 -> 75.21 | Hurt |
| Evaluate on unseen clients instead of seen clients | FedVPA-GP · HH-RLHF unseen clients | Harmless win-rate: 94.25 -> 91.23 | Hurt |
| Evaluate on unseen clients instead of seen clients | FedBiscuit · HH-RLHF unseen clients | Helpful win-rate: 47.32 -> 47.62 | No clear effect |
| Use FedVPA-GP instead of FedBiscuit on unseen clients | unseen clients · HH-RLHF unseen clients | Helpful win-rate: 47.62 -> 63.16 | Improved |
| Use FedVPA-GP instead of FedBiscuit at 70/30 client split | Qwen-2 0.5B, 10 clients · HH-RLHF client ratio | Helpful win-rate: 49.12 -> 68.12 | Improved |
| Use FedVPA-GP instead of FedBiscuit at 20/80 client split | Qwen-2 0.5B, 10 clients · HH-RLHF client ratio | Harmless win-rate: 78.21 -> 90.32 | Improved |
Closest papers by meaning
- Towards Federated RLHF with Aggregated Client Preference for LLMs (ICLR 2025 · similarity 0.88)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.86)
- Differentially Private Sparse Reward Estimation with Preference Feedback (NeurIPS 2026 · similarity 0.86)
- Federated Generalised Variational Inference: A Robust Probabilistic Federated Learning Framework (ICML 2025 · similarity 0.86)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.86)
- Pairwise Calibrated Rewards for Pluralistic Alignment (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
