Research map / Large language models / Alignment and preferences
Personalized Reasoning: Just-in-time Personalization and Why LLMs Fail at It
Shuyue Stella Li, Avinandan Bose, Faeze Brahman, Simon Shaolei Du, Pang Wei Koh, Maryam Fazel, Yulia Tsvetkov
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 |
|---|---|---|---|
| attempt preference discovery by asking questions (vs generic baseline) | 21 frontier LLMs · PrefDisco, 10 tasks | normalized preference alignment: 61 of 210 model-task combinations negative (29.0%) | Mixed |
| ask more questions before answering | discovery mode LLMs · PrefDisco, 10 tasks | preference alignment: r=0.445 | Improved |
| provide full ground-truth preferences (oracle) in the prompt | frontier LLMs · PrefDisco, 10 tasks | task accuracy (%): 65.2 -> 61.8 | Hurt |
| add interactive preference discovery | frontier LLMs · PrefDisco, 10 tasks | task accuracy (%): 65.2 -> 60.1 | Hurt |
| add personalization constraints | frontier LLMs · AIME | task accuracy (%): 12.1% loss | Hurt |
| add personalization constraints | frontier LLMs · CommonsenseQA | task accuracy (%): 5.4% gain | Improved |
| add personalization constraints | frontier LLMs · mathematical reasoning (summary) | task accuracy (%): 3.5% accuracy loss | Hurt |
| add personalization constraints | frontier LLMs · social reasoning (summary) | task accuracy (%): 3.1% gain | Improved |
| force a fixed number of questions (2, 4 or 8) instead of letting the model stop (ablation) | discovery mode LLMs · MATH and MascQA vs CommonsenseQA and SocialIQA | preference alignment: Auto: 1.4 turns vs Fixed: 8.0 turns | Mixed |
| use newer model (Claude Sonnet-4 instead of Claude 3-Opus) | Claude models · PrefDisco discovery alignment | normalized preference alignment | Mixed |
Closest papers by meaning
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.91)
- How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants (NeurIPS 2026 · similarity 0.90)
- Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs (ICLR 2025 · similarity 0.89)
- PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.89)
- HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.89)
- Think-While-Generating: On-the-Fly Reasoning for Personalized Long-Form Generation (ICLR 2026 · similarity 0.88)
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.
