Research map / Label-efficient and robust learning / Recognition and long-tail
HyCoRA: Hyper-Contrastive Role-Adaptive Learning for Role-Playing
Shihao Yang, Zhicong Lu, Yong Yang, Bo Lv, Yang Shen, Nayu Liu
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 |
|---|---|---|---|
| add hyper-contrastive learning (HCL) (ablation) | HyCoRA on ChatGLM2-6B · RoleBench Chinese, general and specific | BLEU-4: general 34.64 -> 35.09; specific 12.65 -> 13.40 when HCL added | Improved |
| use role-specific matrices for both A and B instead of shared B (ablation) | Hyper-Half LoRA · RoleBench general, ChatGLM2-6B | ROUGE-L: 50.69 -> 48.89 | Hurt |
| share matrix A and make B role-specific (swap roles of A and B) (ablation) | Hyper-Half LoRA · RoleBench general, ChatGLM2-6B | ROUGE-L: 50.69 -> 49.78 | Hurt |
| share matrix A and make B role-specific (ablation) | Hyper-Half LoRA · RoleBench specific, ChatGLM2-6B | BLEU-4: 13.40 -> 12.74 | Hurt |
| use HyCoRA instead of RoleLLM shared LoRA | Qwen2-7B, Chinese specific · RoleBench Chinese, specific | ROUGE-2: 20.63 -> 20.50 (Qwen2); 18.66 -> 19.78 (ChatGLM2) | Mixed |
| use HyCoRA instead of shared LoRA in Chinese scenario | ChatGLM2-6B, Qwen2-7B, Qwen2.5-7B · RoleBench Chinese | BLEU / ROUGE: ChatGLM2 specific BLEU 12.09 -> 13.40 | Improved |
| use HyCoRA instead of shared LoRA in English scenario | LLaMA-2-7B · RoleBench English | BLEU: general 25.50 -> 26.26 | Improved |
| use HyCoRA instead of independent LoRA per role | Qwen2-7B · RoleBench Chinese per role | ROUGE-L: Role-1 general 50.30 -> 53.58 | Improved |
| use HyCoRA instead of independent LoRA on Role-5 specific | Qwen2-7B · RoleBench Chinese specific Role-5 | ROUGE-L: 31.97 -> 30.88 | Hurt |
| use HyCoRA instead of LoRA, GPT-4 judged | ChatGLM2-6B, Qwen2-7B, Qwen2.5-7B, LLaMA-2-7B · RoleBench GPT-4 evaluation | avg win rate: ChatGLM2 0.36 -> 0.56 | Improved |
| use HyCoRA instead of LoRA on LLaMA-2-7B Role Knowledge | LLaMA-2-7B · RoleBench GPT-4 RK | win rate: 0.38 -> 0.37 | No clear effect |
| use HyCoRA instead of independent LoRA | trainable parameters · 5 roles | trainable parameters | Improved |
Closest papers by meaning
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.84)
- Attention-Guided Contrastive Role Representations for Multi-agent Reinforcement Learning (ICLR 2024 · similarity 0.83)
- MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents (ICLR 2025 · similarity 0.83)
- Codifying Character Logic in Role-Playing (NeurIPS 2025 · similarity 0.82)
- CoSER: Coordinating LLM-Based Persona Simulation of Established Roles (ICML 2025 · similarity 0.82)
- RoleAgent: Building, Interacting, and Benchmarking High-quality Role-Playing Agents from Scripts (NeurIPS 2024 · similarity 0.82)
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.
