Research map / Large language models / Agents, code and math
Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning
Read the paperCheck your idea against 108,967 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 Role-Aware Reasoning (RAR), which distills character-grounded, style-controlled reasoning traces from a large reasoning model into LLaMA-3-8B via role identity activation (SFT) and reasoning style optimization (preference loss) to improve role-playing agents.
Problem
Role-playing agents built on LLMs with explicit dialogue data lack deep, character-consistent internal thought; directly applying large reasoning models suffers from attention diversion (forgetting the role) and style drift (overly formal reasoning).
Approach
Role Identity Activation injects character elements (emotion, experience, standpoint, motivation) as a reasoning prompt during teacher trace generation and distills them via SFT; Reasoning Style Optimization builds positive/negative style-scenario pairs using two system prompts and trains with a sigmoid preference loss so the student adapts reasoning style to context.
Key results
- RAR achieves the highest CharacterBench average of 3.69 versus 3.57 for Distill and 3.24 for the best Vanilla variant.
- On SocialBench, RAR reaches an average of 65.4 versus 61.1 for Distill and 57.2 for Vanilla.
- Ablating RIA drops the CharacterBench average to 3.49 and ablating RSO to 3.60.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add role-aware reasoning with RIA and RSO | LLaMA-3-8B role-playing agent · CharacterBench | Avg.: 3.57 -> 3.69 | Helped |
| add role-aware reasoning with RIA and RSO | LLaMA-3-8B role-playing agent · SocialBench | Avg.: 61.1 -> 65.4 | Helped |
| enforce extended reasoning with MoreThink | Distill · CharacterBench | Avg.: 3.57 -> 3.05 | Hurt |
| enforce extended reasoning with MoreThink | Distill · SocialBench | Avg.: 61.1 -> 61.5 | Helped |
| add RSO (ablation) | RAR · CharacterBench | Avg.: 3.60 -> 3.69 | Helped |
| add RSO (ablation) | RAR · CharacterBench | Coherence: 2.78 -> 2.86 | Helped |
| add RSO (ablation) | RAR · CharacterBench | Effectiveness: 3.74 -> 3.92 | Helped |
| add RIA (ablation) | RAR · CharacterBench | Avg.: 3.49 -> 3.69 | Helped |
| add RIA (ablation) | RAR · CharacterBench | Role Relevance: 3.61 -> 3.83 | Helped |
| add RIA (ablation) | RAR · CharacterBench | Human-likeness: 2.30 -> 2.78 | Helped |
| add RIA (ablation) | RAR · CharacterBench | Engagement: 2.22 -> 2.93 | Helped |
| enforce extended reasoning with MoreThink | Distill · CharacterBench reasoning traces | Coherence: 2.71 -> 2.53 | Hurt |
Surprising findings
- MoreThink (forced extended reasoning) degrades several CharacterBench persona and memory metrics relative to Distill.
- The paper reports RAR is less concise than Distill and its ablations, which it argues is a beneficial trade-off.
Limitations
- Performance depends on the capability of the teacher LRM used for distillation.
- Experiments limited to LLaMA-3-8B as base model and a single teacher LRM.
- RSO relies on a predefined set of scenario types and style prompts.
Experimental setup
Datasets: RoleBench-Train
Benchmarks: CharacterBench SocialBench
Metrics: CharacterBench dimension scores (1-5 scale) SocialBench accuracy by dimension reasoning trace scores (coherence, relevance, effectiveness,
Baselines: Vanilla RAG (zero-shot, one-shot, few-shot CoT) Distill ZeroThink LessThink MoreThink Neeko Character-GLM
Builds on: RoleBench LLaMA-Factory LoRA Distill from large reasoning model
Closest papers by meaning
- Codifying Character Logic in Role-Playing (NeurIPS 2025 · similarity 0.88)
- MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents (ICLR 2025 · similarity 0.86)
- CoSER: Coordinating LLM-Based Persona Simulation of Established Roles (ICML 2025 · similarity 0.86)
- Concept Incongruence: An Exploration of Time and Death in Role Playing (NeurIPS 2025 · similarity 0.85)
- RLVER: Reinforcement Learning with Verifiable Emotion Rewards for Empathetic Agents (ICLR 2026 · similarity 0.84)
- RoleAgent: Building, Interacting, and Benchmarking High-quality Role-Playing Agents from Scripts (NeurIPS 2024 · 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.