Research map / Large language models / AI and society
A Computational Framework for Evaluating Human-likeness in LLMs' Open-ended Human Behaviors
Yuxuan Lei, Jianxun Lian, Defu Lian, Jincenzi Wu, Tianfu Wang, Xing Xie
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 |
|---|---|---|---|
| scale Qwen2.5 from 3B to 72B | Qwen2.5 simulator · Amazon Movies&TV | consistency: 0.2100 -> 0.8160 | Improved |
| scale Llama-3.1 from 8B to 405B | Llama-3.1 simulator · Amazon Movies&TV | average score: 0.3282 -> 0.4560 | Improved |
| use GPT-4o instead of Llama-3.1-405B | behavior simulator · Amazon Movies&TV | average score: 0.4560 -> 0.5052 | Improved |
| use real user behavior instead of GPT-4o | behavior simulator · Stackexchange Worldbuilding | average score: 0.6247 -> 0.7504 | Improved |
| use real user behavior instead of GPT-4o | behavior simulator · Amazon Movies&TV | distinguishability: 0.3800 -> 0.4820 | Improved |
| use E5-large instead of E5-base as user encoder (ablation) | evaluation framework · Stackexchange Worldbuilding (user encoder ablation) | consistency of Qwen2.5-72B: 0.7610 -> 0.7695 | Improved |
| use Bge-base instead of E5-base as user encoder (ablation) | evaluation framework · Stackexchange Worldbuilding (user encoder ablation) | consistency of Qwen2.5-72B: 0.7610 -> 0.7570 | Hurt |
| use 2-layer MLP instead of 1-layer MLP classifier head (ablation) | evaluation framework · Amazon Movies&TV (classifier head ablation) | consistency of GPT-4o: 0.4000 -> 0.8800 | Improved |
| use 2-layer MLP instead of joint user encoder training (ablation) | evaluation framework · Amazon Movies&TV (classifier head ablation) | consistency of GPT-4o: 0.6600 -> 0.8800 | Improved |
| use 2-layer MLP instead of XGBoost on fixed embeddings (ablation) | evaluation framework · Amazon Movies&TV (classifier head ablation) | consistency of Qwen2.5-72B: 0.6780 -> 0.8160 | Improved |
| use 3-layer MLP instead of 2-layer MLP (ablation) | evaluation framework · Amazon Movies&TV (classifier head ablation) | consistency of Qwen2.5-72B: 0.8160 -> 0.8340 | No clear effect |
| increase simulated sequence length bucket | Llama-3.1 simulators · Amazon Movies&TV | predictability | Hurt |
Closest papers by meaning
- Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.90)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.89)
- OdysSim: Building Foundation Models for Human Behavior Simulation (NeurIPS 2026 · similarity 0.89)
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.89)
- Human Simulacra: Benchmarking the Personification of Large Language Models (ICLR 2025 · similarity 0.88)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · 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.
