Research map / Large language models / Reasoning and chain of thought
Know Thyself, Know Thy User: Intrinsic Dual-Perspective Reasoning for Role-Playing LLMs
Haotong Sun, Jianye Xie, Bocheng Xu, Yinghui Jiang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes intrinsic dual-perspective reasoning for role-playing LLMs that realizes character-authenticity and user-satisfaction duality within the core attention mechanism rather than through prompts, auxiliary modules, or inference-time reflection.
Problem
Role-playing LLMs must balance character authenticity with user satisfaction; prior dual-process and dual-perspective approaches realize this duality externally (prompt conditioning, auxiliary modules, or inference-time reflection).
Approach
Realizes the dual perspectives (self/character and user) intrinsically inside the model's attention mechanism instead of externally via prompts, auxiliary modules, or reflection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune base on same data and recipe | Qwen3-4B-Thinking · CharacterBench AVG (zh) | average score: 3.69 -> 3.80 | Improved |
| add KSKT architecture over data-matched fine-tuning (ablation) | Qwen3-4B-Thinking + FT · CharacterBench AVG (zh) | average score: 3.80 -> 3.92 | Improved |
| use wider unified attention with proxy losses instead of dual-stream factorization (ablation) | KSKT · CharacterBench AVG (zh) | average score: 3.83 vs 3.92 | Hurt |
| add Dual-Stream Axial Attention alone (ablation) | Qwen3-4B-Thinking · CharacterBench Emotion | delta over base: +0.39 | Improved |
| add Mutual-Understanding Position Encoding alone (ablation) | Qwen3-4B-Thinking · CharacterBench Emotion | delta over base: +0.41 | Improved |
| add Self-Awareness MoE alone (ablation) | Qwen3-4B-Thinking · CharacterBench Persona | delta over base: +0.28 | Improved |
| add Bipolar Reasoning Module alone (ablation) | Qwen3-4B-Thinking · CharacterBench Believability and Knowledge | delta over base: +0.18 each | Improved |
| add Self-Awareness MoE alone (ablation) | Qwen3-4B-Thinking · CharacterBench Memory | delta over base: -0.05 | Hurt |
| add KSKT to base model | Qwen3-4B-Thinking · SOTOPIA Relationship | Relationship score: 1.45 -> 1.73 | Improved |
| add KSKT to base model | Qwen3-4B-Thinking · SOTOPIA Secret | Secret score: -0.07 -> -0.18 | Hurt |
| add 3-shot prompting (ablation) | Qwen3-4B-Thinking · SOTOPIA Overall | Overall score: 3.08 -> 3.28 | Improved |
| add 3-shot prompting (ablation) | KSKT · SOTOPIA Overall | Overall score: 3.30 -> 3.48 | Improved |
Experimental setup
Builds on: dual-process theory dual-perspective reasoning
Closest papers by meaning
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.88)
- Concept Incongruence: An Exploration of Time and Death in Role Playing (NeurIPS 2025 · similarity 0.85)
- ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind (AAAI 2025 · similarity 0.85)
- Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.85)
- MindForge: Empowering Embodied Agents with Theory of Mind for Lifelong Cultural Learning (NeurIPS 2025 · similarity 0.84)
- The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.84)
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.
