HER Framework Enhances LLM Role-Playing with Cognitive Simulation
Researchers propose HER, a unified framework for cognitive-level persona simulation in LLM role-playing. HER introduces dual-layer thinking, distinguishing characters' first-person thinking from LLMs' third-person thinking. The framework addresses two key deficiencies: lack of high-quality reasoning traces and lack of reliable reward signals aligned with human preferences. To bridge these gaps, the team curated reasoning-augmented role-playing data via reverse engineering and constructed human-aligned principles and reward models. The work is detailed in arXiv paper 2601.21459.
Key facts
- HER is a unified framework for cognitive-level persona simulation in LLM role-playing.
- It introduces dual-layer thinking: characters' first-person vs. LLMs' third-person.
- Addresses lack of high-quality reasoning traces and reliable reward signals.
- Uses reverse engineering to curate reasoning-augmented role-playing data.
- Constructs human-aligned principles and reward models.
- Paper available on arXiv with ID 2601.21459.
- LLM role-playing is used in companionship, content creation, and digital games.
- Current models capture character tones but struggle with inner thoughts.
Entities
Institutions
- arXiv