New PHASE-Tree system enhances character evolution in long-form role-playing dialogues

Researchers have developed PHASE-Tree, a novel system designed to improve character state evolution in long-horizon role-playing dialogues. This system utilizes a multi-timescale tree structure with distinct layers for identity, persona, session, and moment, allowing for localized updates without destabilizing the character's core traits. To evaluate its effectiveness, a new benchmark called LongEvoRoleBench was introduced, which pairs long-dialogue corpora for cross-episode evolution with short-dialogue corpora for within-scene state tracking. PHASE-Tree demonstrated significant improvements in character-level, semantic, and embedding scores, outperforming existing textual baselines and showing strong correlation with human and GPT-4.1 evaluations. AI

IMPACT Improves character consistency and state tracking in long-form AI-driven narratives, potentially enhancing interactive storytelling and virtual world experiences.

RANK_REASON Academic paper detailing a new model and benchmark. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• GPT-4.1
• Hugging Face
• LongEvoRoleBench
• PHASE-Tree
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →