Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Adaptive Querying with AI Persona Priors
Kaizheng Wang, Yuhang Wu, assaf zeevi
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 persona-based greedy querying instead of CAT-GRM | greedy persona model · WorldValuesBench real users (T=5) | log loss: 1.074 -> 1.019 | Improved |
| use persona-based greedy querying instead of CAT-GRM | greedy persona model · Synthetic users (T=5) | log loss: 1.207 -> 1.089 | Improved |
| use greedy instead of random question selection | persona model · Synthetic users (T=5) | log loss: 1.115 -> 1.089 | Improved |
| use non-adaptive design instead of greedy at larger budget | persona model · WorldValuesBench real users (T=20) | log loss: .999 -> .977 | Improved |
| use greedy instead of nonadaptive at small budget | persona model · WorldValuesBench real users (T=5) | log loss: 1.039 -> 1.019 | Improved |
| use multidimensional IRT (MGRM) instead of unidimensional GRM | CAT baseline · Synthetic users (T=5) | log loss: 1.207 -> 1.164 | Improved |
| cluster persona dictionary to 50 personas (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> .977 | Hurt |
| cluster persona dictionary to 200 personas (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> .963 | No clear effect |
| replace elicited distributions with deterministic answer plus noise eps=0.1 (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> 1.291 | Hurt |
| replace elicited distributions with deterministic answer plus noise eps=0.3 (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> 1.106 | Hurt |
| sharpen LLM distributions with temperature 0.5 (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> 1.021 | Hurt |
| soften LLM distributions with temperature 2 (ablation) | nonadaptive persona model · WorldValuesBench real users (T=20, 10k users) | log loss: .962 -> 1.074 | Hurt |
Closest papers by meaning
- Does Your AI Agent Get You? A Personalizable Framework for Approximating Human Models from Argumentation-based Dialogue Traces (AAAI 2025 · similarity 0.86)
- Adaptive Elicitation of Latent Information Using Natural Language (ICML 2025 · similarity 0.85)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.85)
- When Can Digital Personas Reliably Approximate Human Survey Findings? (NeurIPS 2026 · similarity 0.84)
- PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.84)
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.84)
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.
