Research map / Probabilistic and scientific modeling / Molecules and proteins
Cold-Start Personalization via Bayesian Adaptive Questioning
Avinandan Bose, Shuyue Stella Li, Faeze Brahman, Pang Wei Koh, Simon Shaolei Du, Yulia Tsvetkov, Maryam Fazel, Lin Xiao, Asli Celikyilmaz
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 |
|---|---|---|---|
| learn offline world model of preference correlations (latent structure) instead of none (ablation) | CAPE · PrefDisco | gap closed to ground truth | Improved |
| use adaptive querying instead of random querying (ablation) | CAPE · PrefDisco | gap closed to ground truth at T=5: 24% vs 21% at T=5 | Improved |
| use adaptive question selection (ablation) | CAPE · PrefDisco | number of questions needed: saves one question per interaction (T=5 matches random at T=6) | Improved |
| replace end-to-end GRPO RL questioning policy with offline world model plus Bayesian inference | preference elicitation · PrefDisco average over four domains | preference alignment: 68.5% -> 80.8% | Improved |
| replace GRPO with CAPE | preference elicitation · CSQA | preference alignment (% of Oracle): 55.2 -> 78.2 | Improved |
| replace GRPO with CAPE | preference elicitation · query efficiency | queries for equal alignment: 2.5x on AIME, 7x on MedQA, >15x on SocialIQA and CSQA | Improved |
| GRPO end-to-end RL training vs population average | Llama-3.1-8B-Instruct · SocialIQA, CSQA | preference alignment (% of Oracle): SocialIQA 82.1 -> 71.3; CSQA 72.4 -> 55.2 | Hurt |
| CollabLLM generic interactivity training | Llama-3.1-8B-Instruct · PrefDisco | preference alignment (% of Oracle): 20-26% alignment despite 28-75% adaptivity | Hurt |
| use CAPE instead of GRPO on adaptivity | question selection · PrefDisco | adaptivity (%): GRPO 0-28% vs CAPE 39-62% | Improved |
Closest papers by meaning
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.88)
- Deep Bayesian Active Learning for Preference Modeling in Large Language Models (NeurIPS 2024 · similarity 0.87)
- Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization (NeurIPS 2026 · similarity 0.87)
- Eliciting Human Preferences with Language Models (ICLR 2025 · similarity 0.87)
- HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.86)
- Improving Context-Aware Preference Modeling for Language Models (NeurIPS 2024 · 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.
