Reproduce “Characteristic AI Agents via Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xi Wang, Hongliang Dai, Shen Gao, Piji Li
The paper introduces the "Character100" dataset and a comprehensive benchmark for evaluating Large Language Models (LLMs) in simulating specific real-life individuals as characteristic AI agents. It evaluates LLMs across zero-shot, few-shot, and fine-tuning settings, demonstrating that few-shot learning improves factual consistency but can reduce style consistency, while fine-tuning generally enhances factual recall at a slight cost to style.
Key insights
- Few-shot learning consistently improved LLM performance in background knowledge consistency across models, indicating the value of in-context examples for leveraging profile information.
- Instruction-tuned LLMs generally outperformed their base counterparts in embodying characteristic agents, highlighting the benefits of prior alignment and conversational training.
- A trade-off was observed between background knowledge consistency and style consistency; optimizing for factual accuracy through fine-tuning often led to a slight decrease in stylistic adherence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.