Adversarial Closed-Loop Curriculum for Evolving Role-Playing Agents
Source
arxiv.org
Author
Zheng Zhang, Liu Liu, Qi Chai, Deheng Ye, Peilin Zhao, Mao Zheng, Hao Wang
Date
Context
AdvRole changes a role-playing model's training scenarios as its weaknesses change. One model acts out a character; a second rewrites character profiles and dialogue contexts to create cases that lower the current actor's score. Those cases then become training material for the actor.
The rewriter starts from the original scenarios each round, helping keep new challenges connected to the source material. The authors report stronger overall benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → results than their comparison methods, though not every scored dimension improved in every comparison. The contribution is an evolving curriculum: training keeps targeting situations the current actor handles poorly instead of relying on an unchanged scenario pool.
Key takeaways · AI-distilled
AdvRole alternates two components: an Actor that learns to role-play and a Rewriter that edits character profiles and dialogue contexts into scenarios that are hard for the current Actor.
The Rewriter is rewarded by the performance gap, favoring rewrites that lower the Actor's score relative to the original scenario, so the scenario pool keeps shifting toward under-mastered regions of the character-context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → space.
The authors report consistent gains over baselines on three role-playing benchmarks in English and Chinese plus a new multilingual benchmark they release; the abstract does not give margins.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
articleAgentic Data Engineering for LLMs: System Design and a Controlled Empirical StudyRuomeng Ding, Qianli Shen, Zhaoyang Han, Meixin Chen, Daoyuan Chen, Yaliang Li
articleDeepInstructor: An Agentic AI Instructor for Experience-Driven Idea EvaluationRongcan Pei, Fang Guo, Qinglin Qi, Qi Zhu, Yun Luo, Jianhao Yan, Minjun Zhu, Qiujie Xie, Dehong Zheng, Yue Zhang
articleCoaching Qwen3 Coder 30B to Think Like a CodeClash Arena AgentIvy Ning Zhang