跳转至

PHASE-Tree：长程角色扮演对话中的角色状态演化建模

文章背景与核心概要

在长程角色扮演（Long-horizon role-playing）场景中，AI 角色不仅需要保持长期的身份可识别性，还必须随着剧情的发展而产生自然的演化。然而，传统系统往往难以胜任：其静态的档案表示无法在不破坏未变特征的前提下进行局部更新；同时，现有的评测基准主要测试表层的角色保持和记忆召回，而非真实演化状态下的对话能力。

为了解决这一痛点，本文作者推出了 PHASE-Tree——一个多时间尺度的角色状态树。它包含一个不可变的身份根节点，以及可变的个性、会话和瞬间层，从而支持跨剧集和剧集内部的局部更新。为了准确评估该能力，作者还推出了 LongEvoRoleBench 评测基准，在统一的下一轮话语（next-utterance）协议下，结合了长对话与短对话语料库。实验结果表明，PHASE-Tree 在角色级、语义级和嵌入指标上均显著优于内部变体及外部文本基线。

📌 Summary

Long-horizon role-playing requires AI characters to maintain recognition while organically evolving alongside a narrative. Traditional systems fail because their static profile representations cannot update locally without destabilizing unchanged traits, and existing benchmarks largely test superficial persona preservation and memory recall rather than authentic evolved-state dialogue.

To resolve this, the authors introduce PHASE-Tree, a multi-timescale character-state tree featuring an immutable identity root alongside mutable persona, session, and moment layers. This architecture allows for localized within- and cross-episode updates. To accurately evaluate this capability, they introduce LongEvoRoleBench, a benchmark pairing long-dialogue and short-dialogue corpora under a unified next-utterance protocol. Experimental results demonstrate that PHASE-Tree outperforms internal variants and external textual baselines across character-level, semantic, and embedding metrics.

长程角色扮演要求 AI 角色在随剧情演化时保持可识别性。然而，现有工作在两方面存在不足：表征通常是静态档案，无法在不破坏未变特征的情况下进行局部更新；评测基准主要测试角色保持和记忆 recall（召回），而不是模型是否从角色当前演化后的状态进行发言。我们同时解决了这两个问题。

Long-horizon role-playing requires characters to remain recognizable as they evolve with the narrative. Yet existing work falls short on two fronts: representations are typically static profiles that cannot be updated locally without destabilizing unchanged traits, and benchmarks mainly test persona preservation and memory recall rather than whether a model speaks from a character's currently evolved state. We address both.

PHASE-Tree 是一个多时间尺度的角色状态树，具有不可变的身份根节点以及可变的个性、会话和瞬间层，这使得每个可变字段都成为剧集内部及跨剧集局部更新的寻址目标。它通过显式文本提供或隐式参数自适应来条件化生成。

PHASE-Tree is a multi-timescale character-state tree with an immutable identity root and mutable persona, session, and moment layers, making each mutable field an addressable target for localized within- and cross-episode updates. It conditions generation through explicit textual provision or implicit parametric adaptation.

为了衡量演化状态的生成能力，我们推出了 LongEvoRoleBench，它将用于跨剧集演化的四个长对话语料库与用于场景内状态追踪检查的四个短对话语料库配对，并在统一的下一轮话语协议下运行。在长对话核心任务中，文本型 PHASE-Tree 在与内部变体的 12 个数据集-指标对照单元中赢得了 11 个第一，在与所有外部文本基线的 12 个单元中全部胜出，将角色级、语义级和嵌入得分分别提升了 19.7%、12.4% 和 15.1%。在包含 200 个回复的双盲研究中，人类评分与 GPT-4.1 裁判高度相关（皮尔逊 \(r=0.65\)）；在描述性的 \(n=10\) PT 和 NR 提示子集上，整体差异为 +0.20。长对话中的语义优势在各大 LLM 裁判和生成骨干网络中均保持一致。

To measure evolved-state generation, we introduce LongEvoRoleBench, which pairs four long-dialogue corpora for cross-episode evolution with four short-dialogue corpora as within-scene state-tracking checks, under a unified next-utterance protocol. On the long-dialogue core, textual PHASE-Tree ranks first in 11 of 12 dataset-metric cells against internal variants and all 12 cells against external textual baselines, improving character-level, semantic, and embedding scores by 19.7%, 12.4%, and 15.1% respectively. In a blinded 200-response study, human ratings correlate with the GPT-4.1 judge (Pearson \(r=0.65\)); on descriptive \(n=10\) PT and NR prompt subsets, the Overall difference is +0.20. The long-dialogue Sem advantage persists across LLM judges and generation backbones.