ChengyuDu0123/HER-32B
ChengyuDu0123/HER-32B is a 32 billion parameter role-playing language model built upon Qwen3-32B, developed by Chengyu Du and collaborators. It features a unique Dual-layer Thinking mechanism, distinguishing between third-person system thinking and first-person role thinking, to achieve cognitive-level persona simulation. This model is specifically optimized for advanced and nuanced AI role-playing scenarios, significantly outperforming its baseline on relevant benchmarks.
Loading preview...
HER-RL: Advanced Role-Playing with Dual-Layer Thinking
HER-RL (Human-like Reasoning and Reinforcement Learning for LLM Role-playing) is a 32 billion parameter model, based on Qwen3-32B, designed for highly realistic and nuanced AI role-playing. Developed by Chengyu Du and collaborators, its core innovation is Dual-layer Thinking, which enables cognitive-level persona simulation by separating:
- System Thinking: A third-person meta-level analysis for planning character portrayal.
- Role Thinking: First-person inner thoughts and cognitive processes of the character.
This architecture allows the model to generate responses with rich, interleaved structures, including system analysis, internal thoughts, physical actions, and spoken dialogue. The model's output format explicitly delineates these components, providing a deeper insight into the character's simulated cognition.
Key Capabilities
- Cognitive-level Persona Simulation: Achieves human-like reasoning by distinguishing between meta-level planning and character's internal thoughts.
- Enhanced Role-Playing Performance: Significantly outperforms the Qwen3-32B baseline, showing a +30.26% improvement on CoSER and +14.97% on MiniMax Role-Play Bench.
- Structured Output: Generates detailed responses that include explicit tags for
<system_thinking>,<role_thinking>, and<role_action>, allowing for granular control and analysis of the character's behavior and internal state.
Good For
- Complex Role-Playing Applications: Ideal for scenarios requiring deep character immersion and nuanced responses.
- Interactive Storytelling and Games: Can power NPCs or interactive characters with more believable and consistent personalities.
- Research in AI Cognition: Provides a framework for exploring and simulating cognitive processes in LLMs for persona generation.