Reproduce “PHASE-Tree: Modeling Character-State Evolution in Long-Horizon Role-Playing Dialogue” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bo Tang, Jianan Yang, Junyi Zhu, Yiquan Wu, Rui Zhao, Zhengyu Yang, Yang Zhang, Feiyu Xiong and 2 more
A team led by MemTensor (Shanghai) Technology introduced PHASE-Tree, a multi-timescale character representation, and LongEvoRoleBench, a benchmark for evaluating character evolution in long-horizon role-playing dialogue. The system improves character consistency and contextual coherence in generated responses, with explicit textual conditioning achieving average gains of 19.7% in character score and 12.4% in semantic score over baselines.
Key insights
- Separating immutable identity from mutable traits across different timescales (persona, session, moment) in the PHASE-Tree provides a cleaner and more robust conditioning signal for character evolution.
- Transient contextual cues, captured by the session and moment layers, are crucial for significantly improving character consistency and contextual coherence in dialogue generation.
- While implicit parametric adaptation offers high token efficiency, the profile-to-LoRA mapping acts as a compression bottleneck, limiting the fine-grained state detail that can be effectively conveyed parametrically.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.