QbitAI · WeChatMay 10, 2026, 11:52 UTC+8

Zhejiang University introduces AdaMARP, an AI role-playing framework with scene direction

浙大推出让AI会「导演」的角色扮演框架！四通道消息沉浸式交互｜ACL 2026

Zhejiang University and Tencent Youtu proposed AdaMARP for immersive role-playing, using a four-channel message format and a scene manager; its data pipeline includes 81 literary works, 20 synthetic themes, and AdaptiveBench with 100 evaluation seeds.

Why it matters: ACL 2026 role-play agent work brings four-channel messaging, a scene manager, and an 81-book dataset, clearing HKR-H/K. Narrow use cases and missing open-source or production evidence keep it at threshold featured.

Read the original ↗Export Markdown

Take

AdaMARP is useful because it turns role-play into trainable interfaces, rather than asking one prompt to carry the whole story. The four channels are Thought, Action, Environment, and Speech. The scene manager has five actions: init_scene, pick_speaker, switch_scene, add_role, and end. That is a cleaner engineering split than Character.AI-style long chat, because speaker choice and scene changes become explicit events.

I don’t buy the “immersive role-playing is solved” framing. The data pipeline uses 81 literary works and 20 synthetic themes, while AdaptiveBench has only 100 evaluation seeds and relies on model-based scoring. The hard failures in role-play are long memory, adversarial user turns, and world-state consistency across hours. The article gives no human blind test, retention data, pricing, or stress test over long sessions. ACL acceptance validates the research setup, not the product experience.

Source: QbitAI · WeChat · mp.weixin.qq.com
Published: May 10, 2026, 11:52 UTC+8 · May 10, 2026