Research map / Large language models / Alignment and preferences
HumanLM: Simulating Users with State Alignment Beats Response Imitation
Shirley Wu, Evelyn Choi, Arpandeep Khatua, Zhanghan Wang, Joy He-Yueya, Tharindu Cyril Weerasooriya, Wei Wei, Diyi Yang, Jure Leskovec, James Zou
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 |
|---|---|---|---|
| replace SFT-think with HumanLM latent-state alignment | HumanLM · Humanual-Avg. | response alignment score: 8.6 -> 13.2 | Improved |
| replace standard GRPO-think with HumanLM | HumanLM · Humanual-Avg. | response alignment score: 10.4 -> 13.2 | Improved |
| add reasoning (think) to SFT | SFT · Humanual-Avg. | response alignment score: 6.5 -> 8.6 | Improved |
| add reasoning (think) to Qwen3-8b | Qwen3-8b · Humanual-Avg. | response alignment score: 9.5 -> 8.4 | Hurt |
| add reasoning (think) to GRPO | GRPO · Humanual-Avg. | response alignment score: 10.3 -> 10.4 | No clear effect |
| use GRPO on alignment score instead of SFT | Qwen3-8b · Humanual-Email | response alignment score: 3.94 -> 4.78 | Improved |
| use SFT imitation instead of base model | Qwen3-8b · Humanual-Avg. | response alignment score: 9.5 -> 6.5 | Hurt |
| train HumanLM vs GRPO-think on state alignment exploration | HumanLM · Humanual-Opinion training | span of state alignment score: 23% higher | Improved |
| train HumanLM vs GRPO-think on response alignment exploration | HumanLM · Humanual-Opinion training | span of response alignment score: 104% higher | Improved |
| HumanLM vs Qwen3-8b-think embedding similarity | HumanLM · Humanual-Avg. | embedding similarity: 42.5 -> 45.7 | Improved |
| HumanLM vs GRPO-think in human study win rate | HumanLM · Real-time user study | win rate on overall similarity (%): 27.9% -> 41.4% | Improved |
| HumanLM vs Qwen3-8b-think in human humanlikeness | HumanLM · Real-time user study | share rated quite natural or above (%): 72.1% -> 76.6% | Improved |
Closest papers by meaning
- OdysSim: Building Foundation Models for Human Behavior Simulation (NeurIPS 2026 · similarity 0.89)
- RLTHF: Targeted Human Feedback for LLM Alignment (ICML 2025 · similarity 0.88)
- Human Simulacra: Benchmarking the Personification of Large Language Models (ICLR 2025 · similarity 0.88)
- Understanding the Learning Dynamics of Alignment with Human Feedback (ICML 2024 · similarity 0.87)
- Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors (NeurIPS 2026 · similarity 0.87)
- Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.87)
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.
