pith. sign in

arxiv: 2606.27025 · v1 · pith:PS2EWONGnew · submitted 2026-06-25 · 💻 cs.CL

Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization

Zhenhua Xu , Dongsheng Chen , Jian Li , Yitong Lin , Zhebo Wang , Jiafu Wu , Yizhang Jin , Chengjie Wang
show 2 more authors
Meng Han Yabiao Wang
This is my paper

Pith reviewed 2026-06-26 04:52 UTC · model grok-4.3

classification 💻 cs.CL
keywords role-playing agentschain-of-thoughtpolicy optimizationmutual informationreinforcement learningcharacter fidelityreward hackingpsychology-grounded reasoning
0
0 comments X

The pith

Psychology-grounded chain-of-thought plus profile-token mutual information weighting lets models portray characters more faithfully from profiles.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Supervised fine-tuning for role-playing agents encourages surface mimicry without internal thought processes tied to the character profile, which limits generalization to new situations. Psy-CoT decomposes pre-response reasoning into Interaction Perception, Psychological Empathy, and Logical Construction so the model derives responses dynamically from the profile. RAPO then applies reinforcement learning that uses profile-token mutual information to weight gradients asymmetrically, amplifying role-specific tokens under positive advantage and attenuating them under negative advantage to block accumulation of generic reward-hacking phrases. Experiments show Psy-CoT beats prior role-playing CoT methods and RAPO beats GRPO across model scales on CoSER, CharacterBench, and CharacterEval. A reader would care because the combination moves agents from scripted imitation toward profile-driven, human-like character behavior.

Core claim

The paper establishes that structured three-step reasoning from psychology allows the model to think dynamically from the profile rather than mimic surface patterns, and that asymmetric gradient weighting via profile-token mutual information prevents both generic and role-specific phrases from receiving identical signals under LLM reward models, producing higher character fidelity on the three benchmarks.

What carries the argument

Psy-CoT's three role-specific reasoning steps and RAPO's profile-token mutual information used to weight gradients asymmetrically for role-specific tokens.

If this is right

  • Agents should generalize better to out-of-distribution profiles because reasoning starts from the profile itself rather than learned mimicry.
  • Reward hacking should decrease because role-specific tokens receive stronger positive and weaker negative updates than generic phrases.
  • The gains should appear consistently across different model scales on the reported benchmarks.
  • The separation of token types should reduce the accumulation of phrases that exploit the reward model without advancing character fidelity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The mutual-information weighting approach could reduce reward hacking in other reinforcement-learning settings where reward models favor generic safe outputs.
  • The three-step reasoning structure might transfer to tasks such as multi-turn dialogue or story continuation that also require profile or context consistency.
  • Applying the same asymmetric weighting to other forms of preference optimization could test whether mutual information provides a general tool for distinguishing signal from noise in token-level updates.

Load-bearing premise

Profile-token mutual information reliably separates genuinely role-specific tokens from generic reward-hacking phrases without introducing new biases or requiring profile-specific tuning.

What would settle it

Train identical models with RAPO versus standard GRPO and measure the rate at which generic phrases versus profile-unique phrases appear in responses to held-out profiles; equal rates would show the weighting fails to create separation.

Figures

Figures reproduced from arXiv: 2606.27025 by Chengjie Wang, Dongsheng Chen, Jiafu Wu, Jian Li, Meng Han, Yabiao Wang, Yitong Lin, Yizhang Jin, Zhebo Wang, Zhenhua Xu.

Figure 1
Figure 1. Figure 1: Overview of RAPO. The upper part illustrates the sampling stage: given an input sample, the policy model rolls out multiple completions, each of which is scored by the reward function, and group-relative advantages are then computed. The lower part depicts the token-level gradient modulation applied to each completion, where role-agnostic and role-specific tokens receive differentiated gradient signals con… view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of different policy optimization methods on Qwen2.5-7B-Instruct with Psy-CoT. (a) GRPO and DAPO exhibit early entropy collapse indicative of premature convergence, while GSPO partially recovers but with limited exploration space remaining; RAPO maintains a stable exploration–exploitation balance throughout training. (b) RAPO achieves the highest reward among all methods, confirming that i… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of different CoT strategies on strong instruction-following models. Both Psy-CoT and CB-CoT improve CoSER scores, but applying CoT to less capable models can degrade performance on CharacterBench. (a) CharacterBench (b) CharacterEval [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average performance across models on CharacterBench and CharacterEval. RAPO achieves the best average score among all RL variants on Qwen2.5-7B-Instruct; SFT-based methods harm generalization. See [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human evaluation results. We recruit 5 PhD candidates from different universities who are familiar with role-playing tasks, compensated at $15/hour. For each of the three backbone models (Qwen2.5-7B-Instruct, Qwen3-4B-Instruct, Llama-3.1-8B-Instruct), we use the same set of 20 randomly sampled CoSER benchmark samples to generate trajectories, ensuring a fair comparison across meth￾ods. Annotators judge whi… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on µ− in RAPO. A positive µ− reduces gradients on role-specific tokens under negative advantage, promoting exploration; a negative µ− amplifies them, leading to entropy collapse. results confirm that a positive µ− is essential for preserving exploration: by shrinking the gradient on role-specific tokens when the advantage is negative, RAPO lowers the penalty for role-related exploration, giving th… view at source ↗
read the original abstract

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm -- supervised fine-tuning -- encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose \textbf{Psy-CoT}, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps -- \emph{Interaction Perception}, \emph{Psychological Empathy}, and \emph{Logical Construction} -- so that the model \emph{thinks dynamically} from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals -- this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose \textbf{Role-Aware Policy Optimization (RAPO)}, which uses profile--token mutual information to weight gradients asymmetrically -- amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that Psy-CoT, a psychology-grounded chain-of-thought framework decomposing reasoning into Interaction Perception, Psychological Empathy, and Logical Construction steps, enables dynamic profile-based thinking and outperforms prior role-playing CoT methods. It further claims that RAPO, which applies profile-token mutual information to asymmetrically weight gradients (amplifying role-specific tokens under positive advantage and attenuating under negative), mitigates reward hacking in LLM-based rewards and consistently surpasses GRPO across model scales on CoSER, CharacterBench, and CharacterEval.

Significance. If the empirical claims hold with proper validation, the work offers a structured psychological reasoning approach and a targeted fix for reward-model exploitation in role-playing RL, potentially improving out-of-distribution fidelity for general agents. The multi-benchmark, multi-scale evaluation provides a reasonable testbed for the methods.

major comments (2)
  1. [RAPO formulation] The RAPO description (method section) provides no estimation procedure, sampling details (joint vs. marginal), or token count for computing profile-token mutual information. This is load-bearing for the central claim that RAPO surpasses GRPO, because the asymmetric weighting mechanism requires that MI scores cleanly separate role-specific tokens from generic reward-hacking phrases without introducing new biases or profile-specific retuning.
  2. [Experiments] No ablation studies, error bars, dataset statistics, or per-benchmark breakdowns are visible to support the headline result that RAPO consistently beats GRPO. Without these, the cross-scale and cross-benchmark superiority cannot be verified as robust rather than an artifact of the specific reward model or evaluation setup.
minor comments (1)
  1. [Method] Notation for mutual information and advantage weighting should be formalized with an equation to clarify the asymmetric update rule.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point-by-point below and will revise the manuscript to incorporate additional methodological details and empirical analyses.

read point-by-point responses

  1. Referee: [RAPO formulation] The RAPO description (method section) provides no estimation procedure, sampling details (joint vs. marginal), or token count for computing profile-token mutual information. This is load-bearing for the central claim that RAPO surpasses GRPO, because the asymmetric weighting mechanism requires that MI scores cleanly separate role-specific tokens from generic reward-hacking phrases without introducing new biases or profile-specific retuning.

    Authors: We agree the current description is high-level and lacks implementation specifics. In the revision we will add an explicit subsection detailing the MI estimator (including the use of Monte Carlo sampling over the policy distribution to approximate the joint p(profile, token) and marginals, the number of tokens sampled per profile, and any smoothing or thresholding applied). This will make the separation of role-specific versus generic tokens transparent and allow readers to assess potential biases. revision: yes

  2. Referee: [Experiments] No ablation studies, error bars, dataset statistics, or per-benchmark breakdowns are visible to support the headline result that RAPO consistently beats GRPO. Without these, the cross-scale and cross-benchmark superiority cannot be verified as robust rather than an artifact of the specific reward model or evaluation setup.

    Authors: The manuscript already reports results on three benchmarks and multiple model scales, yet we concur that further disaggregation would strengthen the claims. We will add (i) an ablation isolating the MI-weighting component of RAPO, (ii) error bars computed over at least three random seeds where compute permits, (iii) basic dataset statistics (size, profile length distribution, etc.), and (iv) per-benchmark tables with mean and variance. These additions will be included in the revised experimental section. revision: yes

Circularity Check

0 steps flagged

No circularity; claims are proposals without self-referential derivations

full rationale

The abstract and description introduce Psy-CoT as a three-step reasoning framework and RAPO as an asymmetric gradient weighting scheme based on profile-token mutual information. No equations, derivations, or fitted-parameter renamings are present. No self-citation chains or uniqueness theorems are invoked to justify core choices. The improvements are framed as empirical outcomes on external benchmarks (CoSER, CharacterBench, CharacterEval) rather than reductions to inputs by construction. This is the expected honest non-finding for a methods paper whose central contributions are algorithmic proposals evaluated externally.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are described or derivable from the provided text.

pith-pipeline@v0.9.1-grok · 5812 in / 1037 out tokens · 23581 ms · 2026-06-26T04:52:15.875840+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

143 extracted references · 86 canonical work pages · 3 internal anchors

  1. [1]

    2026 , eprint=

    Improving Search Agent with One Line of Code , author=. 2026 , eprint=

    2026

  2. [2]

    The Bell system technical journal , volume=

    A mathematical theory of communication , author=. The Bell system technical journal , volume=. 1948 , publisher=

    1948

  3. [3]

    arXiv preprint arXiv:2506.13131 , year=

    Alphaevolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

    Pith/arXiv arXiv

  4. [4]

    2026 , eprint=

    Policy of Thoughts: Scaling LLM Reasoning via Test-time Policy Evolution , author=. 2026 , eprint=

    2026

  5. [5]

    2026 , eprint=

    ICPO: Illocution-Calibrated Policy Optimization for Multi-Turn Conversation , author=. 2026 , eprint=

    2026

  6. [9]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  7. [12]

    Findings of the

    Gao, Zhenpeng and Xing, Xiaofen and Xu, Xiangmin , editor =. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-emnlp.288 , eventtitle =

    work page doi:10.18653/v1/2025.findings-emnlp.288 2025

  8. [15]

    doi:10.48550/ARXIV.2308.09597 , urldate =

    Li, Cheng and Leng, Ziang and Yan, Chenxi and Shen, Junyi and Wang, Hao and Mi, Weishi and Fei, Yaying and Feng, Xiaoyang and Yan, Song and Wang, HaoSheng and Zhan, Linkang and Jia, Yaokai and Wu, Pingyu and Sun, Haozhen , year = 2023, journal =. doi:10.48550/ARXIV.2308.09597 , urldate =. arXiv , copyright =:2308.09597 , primaryclass =

    work page doi:10.48550/arxiv.2308.09597 2023

  9. [18]

    Xintao Wang and Heng Wang and Yifei Zhang and Xinfeng Yuan and Rui Xu and Jen-tse Huang and Siyu Yuan and Haoran Guo and Jiangjie Chen and Shuchang Zhou and Wei Wang and Yanghua Xiao , booktitle=. Co. 2025 , url=

    2025

  10. [19]

    HER: Human-like Reasoning and Reinforcement Learning for LLM Role-playing

    Du, Chengyu and Wang, Xintao and Chen, Aili and Li, Weiyuan and Xu, Rui and Liu, Junteng and Huang, Zishan and Tian, Rong and Sun, Zijun and Li, Yuhao and Feng, Liheng and Ding, Deming and Zhao, Pengyu and Xiao, Yanghua , year = 2026, number =. doi:10.48550/arXiv.2601.21459 , urldate =. 2601.21459 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.21459 2026

  11. [20]

    2026 , eprint=

    AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing , author=. 2026 , eprint=

    2026

  12. [25]

    Character-

    Tang, Yihong and Chen, Kehai and Bai, Xuefeng and Wang, Benyou and Liu, Zeming and Wang, Haifeng and Zhang, Min , year = 2026, number =. Character-. doi:10.48550/arXiv.2601.04611 , urldate =. 2601.04611 , primaryclass =

    work page doi:10.48550/arxiv.2601.04611 2026

  13. [26]

    Theory of Mind in Large Language Models: Assessment and Enhancement

    Chen, Ruirui and Jiang, Weifeng and Qin, Chengwei and Tan, Cheston. Theory of Mind in Large Language Models: Assessment and Enhancement. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1522

    work page doi:10.18653/v1/2025.acl-long.1522 2025

  14. [27]

    1991 , publisher=

    Emotion and adaptation , author=. 1991 , publisher=

    1991

  15. [28]

    Review of general psychology , volume=

    The emerging field of emotion regulation: An integrative review , author=. Review of general psychology , volume=. 1998 , publisher=

    1998

  16. [29]

    , author=

    Building a practically useful theory of goal setting and task motivation: A 35-year odyssey. , author=. American psychologist , volume=. 2002 , publisher=

    2002

  17. [30]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  18. [31]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  19. [32]

    arXiv preprint arXiv:2507.18071 , year=

    Group sequence policy optimization , author=. arXiv preprint arXiv:2507.18071 , year=

    Pith/arXiv arXiv

  20. [33]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    Pith/arXiv arXiv

  21. [35]

    Proceedings of the

    Zhou, Jinfeng and Huang, Yongkang and Wen, Bosi and Bi, Guanqun and Chen, Yuxuan and Ke, Pei and Chen, Zhuang and Xiao, Xiyao and Peng, Libiao and Tang, Kuntian and Zhang, Rongsheng and Zhang, Le and Lv, Tangjie and Hu, Zhipeng and Wang, Hongning and Huang, Minlie , year = 2025, eprint =. Proceedings of the. doi:10.1609/AAAI.V39I24.34806 , urldate =

    work page doi:10.1609/aaai.v39i24.34806 2025

  22. [36]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=

  23. [37]

    2025 , month =

    GPT-5 System Card , author =. 2025 , month =

    2025

  24. [38]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  25. [39]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  26. [40]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  27. [41]

    2024 , howpublished =

    Doubao Large Language Models , author =. 2024 , howpublished =

    2024

  28. [42]

    Towards a

    Chen, Chaoran and Yao, Bingsheng and Zou, Ruishi and Hua, Wenyue and Lyu, Weimin and Li, Toby Jia-Jun and Wang, Dakuo , editor =. Towards a. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.938 , eventtitle =

    work page doi:10.18653/v1/2025.findings-acl.938 2025

  29. [43]

    Chen, Nuo and Wang, Yan and Deng, Yang and Li, Jia , date =. The. 2025 , eprint =. doi:10.48550/arXiv.2407.11484 , pubstate =

    work page doi:10.48550/arxiv.2407.11484 2025

  30. [44]

    Chen, Jiangjie and Wang, Xintao and Xu, Rui and Yuan, Siyu and Zhang, Yikai and Shi, Wei and Xie, Jian and Li, Shuang and Yang, Ruihan and Zhu, Tinghui and Chen, Aili and Li, Nianqi and Chen, Lida and Hu, Caiyu and Wu, Siye and Ren, Scott and Fu, Ziquan and Xiao, Yanghua , date =. From. 2024 , journaltitle =. 2404.18231 , eprinttype =

    arXiv 2024

  31. [45]

    Shanahan, Murray and McDonell, Kyle and Reynolds, Laria , date =. Role-. 2023 , eprint =. doi:10.48550/arXiv.2305.16367 , pubstate =

    work page doi:10.48550/arxiv.2305.16367 2023

  32. [46]

    Tseng, Yu-Min and Huang, Yu-Chao and Hsiao, Teng-Yun and Chen, Wei-Lin and Huang, Chao-Wei and Meng, Yu and Chen, Yun-Nung , editor =. Two. Findings of the. 2024 , pages =. doi:10.18653/v1/2024.findings-emnlp.969 , eventtitle =

    work page doi:10.18653/v1/2024.findings-emnlp.969 2024

  33. [47]

    Tseng, Yu-Min and Huang, Yu-Chao and Hsiao, Teng-Yun and Chen, Wei-Lin and Huang, Chao-Wei and Meng, Yu and Chen, Yun-Nung , date =. Two. 2024 , eprint =. doi:10.48550/arXiv.2406.01171 , pubstate =

    work page doi:10.48550/arxiv.2406.01171 2024

  34. [48]

    2025 , eprint =

    Bai, Ting and Kang, Jiazheng and Fan, Jiayang , date =. 2025 , eprint =. doi:10.48550/arXiv.2412.20024 , pubstate =

    work page doi:10.48550/arxiv.2412.20024 2025

  35. [49]

    2025 , eprint =

    Cheng, Xilong and Qin, Yunxiao and Tan, Yuting and Li, Zhengnan and Wang, Ye and Xiao, Hongjiang and Zhang, Yuan , date =. 2025 , eprint =. doi:10.48550/arXiv.2505.12814 , pubstate =

    work page doi:10.48550/arxiv.2505.12814 2025

  36. [50]

    Chen, Nuo and Wang, Yan and Jiang, Haiyun and Cai, Deng and Li, Yuhan and Chen, Ziyang and Wang, Longyue and Li, Jia , editor =. Large. Findings of the. 2023 , pages =. doi:10.18653/v1/2023.findings-emnlp.570 , eventtitle =

    work page doi:10.18653/v1/2023.findings-emnlp.570 2023

  37. [51]

    2025 , eprint =

    Chen, Weishu and Tang, Jinyi and Hou, Zhouhui and Han, Shihao and Zhan, Mingjie and Huang, Zhiyuan and Liu, Delong and Guo, Jiawei and Zhao, Zhicheng and Su, Fei , date =. 2025 , eprint =. doi:10.48550/arXiv.2509.11860 , pubstate =

    work page doi:10.48550/arxiv.2509.11860 2025

  38. [52]

    Chen, Siyuan and Si, Qingyi and Yang, Chenxu and Liang, Yunzhi and Lin, Zheng and Liu, Huan and Wang, Weiping , date =. A. 2024 , eprint =. doi:10.48550/arXiv.2411.02457 , pubstate =

    work page doi:10.48550/arxiv.2411.02457 2024

  39. [53]

    Thespian:

    Cui, Christopher and Peng, Xiangyu and Riedl, Mark , date =. Thespian:. 2023 , eprint =. doi:10.48550/arXiv.2308.01872 , pubstate =

    work page doi:10.48550/arxiv.2308.01872 2023

  40. [54]

    2025 , doi =

    Fang, Feiteng and Lin, Ting-En and Wu, Yuchuan and Liu, Xiong and Huang, Xiang and Chen, Dingwei and Ye, Jing and Zhang, Haonan and Zhu, Liang and Alinejad-Rokny, Hamid and Yang, Min and Huang, Fei and Li, Yongbin , date =. 2025 , doi =

    2025

  41. [55]

    He, Kai and Huang, Yucheng and Wang, Wenqing and Ran, Delong and Sheng, Dongming and Huang, Junxuan and Lin, Qika and Xu, Jiaxing and Liu, Wenqiang and Feng, Mengling , editor =. Crab:. Proceedings of the 63rd. 2025 , pages =. doi:10.18653/v1/2025.acl-long.731 , eventtitle =

    work page doi:10.18653/v1/2025.acl-long.731 2025

  42. [56]

    Emotional

    Huang, Le and Lan, Hengzhi and Sun, Zijun and Shi, Chuan and Bai, Ting , date =. Emotional. Proceedings of the \. 2024 , eprint =. doi:10.1109/ICKG63256.2024.00023 , eventtitle =

    work page doi:10.1109/ickg63256.2024.00023 2024

  43. [57]

    Huang, Yuxuan , date =. Orca:. 2024 , eprint =. doi:10.48550/arXiv.2411.10006 , pubstate =

    work page doi:10.48550/arxiv.2411.10006 2024

  44. [58]

    Enhancing

    Ji, Ke and Lian, Yixin and Li, Linxu and Gao, Jingsheng and Li, Weiyuan and Dai, Bin , editor =. Enhancing. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.1344 , eventtitle =

    work page doi:10.18653/v1/2025.findings-acl.1344 2025

  45. [59]

    Kong, Aobo and Zhao, Shiwan and Chen, Hao and Li, Qicheng and Qin, Yong and Sun, Ruiqi and Zhou, Xin and Zhou, Jiaming and Sun, Haoqin , date =. Self-. 2024 , eprint =. doi:10.48550/arXiv.2407.08995 , pubstate =

    work page doi:10.48550/arxiv.2407.08995 2024

  46. [60]

    Tachikuma:

    Liang, Yuanzhi and Zhu, Linchao and Yang, Yi , date =. Tachikuma:. 2023 , eprint =. doi:10.48550/arXiv.2307.12573 , pubstate =

    work page doi:10.48550/arxiv.2307.12573 2023

  47. [61]

    2023 , url =

    Li, Cheng and Leng, Ziang and Yan, Chenxi and Shen, Junyi and Wang, Hao and Mi, Weishi and Fei, Yaying and Feng, Xiaoyang and Yan, Song and Wang, HaoSheng and Zhan, Linkang and Jia, Yaokai and Wu, Pingyu and Sun, Haozhen , date =. 2023 , url =

    2023

  48. [62]

    Proceedings of the 2025

    Liu, Cheng and Lu, Yifei and Ye, Fanghua and Li, Jian and Chen, Xingyu and Ren, Feiliang and Tu, Zhaopeng and Li, Xiaolong , editor =. Proceedings of the 2025. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.1389 , eventtitle =

    work page doi:10.18653/v1/2025.emnlp-main.1389 2025

  49. [63]

    Liu, Wenhao and An, Siyu and Lu, Junru and Wu, Muling and Li, Tianlong and Wang, Xiaohua and Lv, Changze and Zheng, Xiaoqing and Yin, Di and Sun, Xing and Huang, Xuanjing , editor =. Tell. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.311 , eventtitle =

    work page doi:10.18653/v1/2025.findings-acl.311 2025

  50. [64]

    Long, Weifan and Wen, Wen and Zhai, Peng and Zhang, Lihua , date =. Role. 2024 , eprint =. doi:10.48550/arXiv.2411.01166 , pubstate =

    work page doi:10.48550/arxiv.2411.01166 2024

  51. [65]

    Lu, Keming and Yu, Bowen and Zhou, Chang and Zhou, Jingren , date =. Large. Proceedings of the 62nd. 2024 , pages =. doi:10.18653/v1/2024.acl-long.423 , eventtitle =

    work page doi:10.18653/v1/2024.acl-long.423 2024

  52. [66]

    Proceedings of the 2025

    Park, Jeiyoon and Park, Chanjun and Lim, Heuiseok , editor =. Proceedings of the 2025. 2025 , pages =. doi:10.18653/v1/2025.naacl-industry.24 , eventtitle =

    work page doi:10.18653/v1/2025.naacl-industry.24 2025

  53. [67]

    Park, Jeiyoon and Han, Yongshin and Kim, Minseop and Yang, Kisu , date =. Dynamic. 2025 , eprint =. doi:10.48550/arXiv.2508.02016 , pubstate =

    work page doi:10.48550/arxiv.2508.02016 2025

  54. [68]

    Codifying

    Peng, Letian and Shang, Jingbo , date =. Codifying. 2025 , url =

    2025

  55. [69]

    Qin, Haiming and Zhang, Jiwei and Zhang, Wei and Lu, KeZhong and Zhou, Mingyang and Liao, Hao and Mao, Rui , editor =. R-. Proceedings of the 2025. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.1372 , eventtitle =

    work page doi:10.18653/v1/2025.emnlp-main.1372 2025

  56. [70]

    Capturing

    Ran, Yiting and Wang, Xintao and Xu, Rui and Yuan, Xinfeng and Liang, Jiaqing and Xiao, Yanghua and Yang, Deqing , editor =. Capturing. Findings of the. 2024 , pages =. doi:10.18653/v1/2024.findings-emnlp.853 , eventtitle =

    work page doi:10.18653/v1/2024.findings-emnlp.853 2024

  57. [71]

    Ruangtanusak, Saksorn and Taveekitworachai, Pittawat and Pipatanakul, Kunat , date =. Talk. 2025 , eprint =. doi:10.48550/arXiv.2509.00482 , pubstate =

    work page doi:10.48550/arxiv.2509.00482 2025

  58. [72]

    2025 , eprint =

    Rupprecht, Timothy and Nan, Enfu and Akbari, Arash and Akbari, Arman and Lu, Lei and Maan, Priyanka and Duffy, Sean and Zhao, Pu and He, Yumei and Kaeli, David and Wang, Yanzhi , date =. 2025 , eprint =. doi:10.48550/arXiv.2509.12168 , pubstate =

    work page doi:10.48550/arxiv.2509.12168 2025

  59. [73]

    Character-

    Shao, Yunfan and Li, Linyang and Dai, Junqi and Qiu, Xipeng , editor =. Character-. Proceedings of the 2023. 2023 , pages =

    2023

  60. [74]

    Identity-

    Sun, Libo and Wang, Siyuan and Huang, Xuanjing and Wei, Zhongyu , date =. Identity-. 2024 , eprint =. doi:10.48550/arXiv.2407.19412 , pubstate =

    work page doi:10.48550/arxiv.2407.19412 2024

  61. [75]

    2024 , eprint =

    Tang, Yihong and Ou, Jiao and Liu, Che and Zhang, Fuzheng and Zhang, Di and Gai, Kun , date =. 2024 , eprint =. doi:10.48550/arXiv.2409.14710 , pubstate =

    work page doi:10.48550/arxiv.2409.14710 2024

  62. [76]

    Thinking in

    Tang, Yihong and Chen, Kehai and Yang, Muyun and Niu, Zhengyu and Li, Jing and Zhao, Tiejun and Zhang, Min , date =. Thinking in. 2025 , eprint =. doi:10.48550/arXiv.2506.01748 , pubstate =

    work page doi:10.48550/arxiv.2506.01748 2025

  63. [77]

    Proceedings of the 1st

    Tao, Meiling and Xuechen, Liang and Shi, Tianyu and Yu, Lei and Xie, Yiting , editor =. Proceedings of the 1st. 2024 , pages =

    2024

  64. [78]

    2023 , eprint =

    Tu, Quan and Chen, Chuanqi and Li, Jinpeng and Li, Yanran and Shang, Shuo and Zhao, Dongyan and Wang, Ran and Yan, Rui , date =. 2023 , eprint =. doi:10.48550/arXiv.2308.10278 , pubstate =

    work page doi:10.48550/arxiv.2308.10278 2023

  65. [79]

    2025 , eprint =

    Wang, Xintao and Wang, Heng and Zhang, Yifei and Yuan, Xinfeng and Xu, Rui and Huang, Jen-tse and Yuan, Siyu and Guo, Haoran and Chen, Jiangjie and Zhou, Shuchang and Wang, Wei and Xiao, Yanghua , date =. 2025 , eprint =. doi:10.48550/arXiv.2502.09082 , pubstate =

    work page doi:10.48550/arxiv.2502.09082 2025

  66. [80]

    2025 , eprint =

    Wang, Xiaoyang and Zhang, Hongming and Ge, Tao and Yu, Wenhao and Yu, Dian and Yu, Dong , date =. 2025 , eprint =. doi:10.48550/arXiv.2501.15427 , pubstate =

    work page doi:10.48550/arxiv.2501.15427 2025

  67. [81]

    2025 , url =

    Wang, Zongsheng and Sun, Kaili and Wu, Bowen and Yu, Qun and Li, Ying and Wang, Baoxun , date =. 2025 , url =

    2025

  68. [82]

    Wang, Noah and Peng, Z.y. and Que, Haoran and Liu, Jiaheng and Zhou, Wangchunshu and Wu, Yuhan and Guo, Hongcheng and Gan, Ruitong and Ni, Zehao and Yang, Jian and Zhang, Man and Zhang, Zhaoxiang and Ouyang, Wanli and Xu, Ke and Huang, Wenhao and Fu, Jie and Peng, Junran , editor =. Findings of the. 2024 , pages =. doi:10.18653/v1/2024.findings-acl.878 , ...

    work page doi:10.18653/v1/2024.findings-acl.878 2024

  69. [83]

    2025 , eprint =

    Wang, Yongjie and Leung, Jonathan and Shen, Zhiqi , date =. 2025 , eprint =. doi:10.48550/arXiv.2505.18541 , pubstate =

    work page doi:10.48550/arxiv.2505.18541 2025

  70. [84]

    MECoT: Markov emotional chain-of-thought for personality-consistent role-playing,

    Wei, Yangbo and Huang, Zhen and Zhao, Fangzhou and Feng, Qi and Xing, Wei W. , editor =. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.435 , eventtitle =

    work page doi:10.18653/v1/2025.findings-acl.435 2025

  71. [85]

    Ontologically

    Weir, Nathaniel and Thomas, Ryan and family=Amore, given=Randolph, prefix=d', useprefix=true and Hill, Kellie and Van Durme, Benjamin and Jhamtani, Harsh , editor =. Ontologically. Proceedings of the 2024. 2024 , pages =. doi:10.18653/v1/2024.emnlp-main.520 , eventtitle =

    work page doi:10.18653/v1/2024.emnlp-main.520 2024

  72. [86]

    Crafting

    Yang, Bohao and Liu, Dong and Xiao, Chenghao and Zhao, Kun and Tang, Chen and Li, Chao and Yuan, Lin and Guang, Yang and Lin, Chenghua , editor =. Crafting. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-emnlp.1100 , eventtitle =

    work page doi:10.18653/v1/2025.findings-emnlp.1100 2025

  73. [87]

    2025 , eprint =

    Yang, Shihao and Lu, Zhicong and Yang, Yong and Lv, Bo and Shen, Yang and Liu, Nayu , date =. 2025 , eprint =. doi:10.48550/arXiv.2511.08017 , pubstate =

    work page doi:10.48550/arxiv.2511.08017 2025

  74. [88]

    2025 , eprint =

    Yang, Tao and Zhu, Yuhua and Quan, Xiaojun and Liu, Cong and Wang, Qifan , date =. 2025 , eprint =. doi:10.48550/arXiv.2502.03821 , pubstate =

    work page doi:10.48550/arxiv.2502.03821 2025

  75. [89]

    2025 , url =

    Yao, Bingsheng and Sun, Bo and Dong, Yuanzhe and Lu, Yuxuan and Wang, Dakuo , date =. 2025 , url =

    2025

  76. [90]

    Findings of the

    Ye, Jing and Wang, Rui and Wu, Yuchuan and Ma, Victor and Fang, Feiteng and Huang, Fei and Li, Yongbin , editor =. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-emnlp.18 , eventtitle =

    work page doi:10.18653/v1/2025.findings-emnlp.18 2025

  77. [91]

    Findings of the

    Yin, Xuyan and Yang, Xinran and Li, Zihao and Zou, Lixin and Li, Chenliang , editor =. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.findings-emnlp.323 , eventtitle =

    work page doi:10.18653/v1/2025.findings-emnlp.323 2025

  78. [92]

    Yu, Yeyong and Yu, Runsheng and Wei, Haojie and Zhang, Zhanqiu and Qian, Quan , editor =. Beyond. Proceedings of the 63rd. 2025 , pages =. doi:10.18653/v1/2025.acl-long.586 , eventtitle =

    work page doi:10.18653/v1/2025.acl-long.586 2025

  79. [93]

    Yu, Xiaoyan and Luo, Tongxu and Wei, Yifan and Lei, Fangyu and Huang, Yiming and Peng, Hao and Zhu, Liehuang , editor =. Neeko:. Proceedings of the 2024. 2024 , pages =. doi:10.18653/v1/2024.emnlp-main.697 , eventtitle =

    work page doi:10.18653/v1/2024.emnlp-main.697 2024

  80. [94]

    Proceedings of the 18th

    Zhao, Runcong and Zhang, Wenjia and Li, Jiazheng and Zhu, Lixing and Li, Yanran and He, Yulan and Gui, Lin , editor =. Proceedings of the 18th. 2024 , pages =

    2024

Showing first 80 references.