Pith. sign in

REVIEW 3 major objections 4 minor 56 references

Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A two-stage role-aware reasoning method — identity anchoring plus scene-matched thinking style — makes role-playing agents answer in character and beats generic chain-of-thought distillation on both main role-play benchmarks.

desk verdict A plausible two-stage recipe for role-playing agents, but the RSO stage's scenario-style mapping is under-validated and the 'significant' claims lack statistical support. read the letter →

arxiv 2506.01748 v1 pith:L3SSU26E submitted 2025-06-02 cs.CL

Yihong Tang , Kehai Chen , Muyun Yang , Zhengyu Niu , Jing Li , Tiejun Zhao , Min Zhang This is my paper
classification cs.CL
keywords role-playingagentsrole-awarereasoningchain-of-thoughtdistillationcharacterconsistencypreferenceoptimizationstyleattentiondiversionlargemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the reason role-playing agents sound shallow or off-character is not missing dialogue data but a missing thinking style: when a reasoning model is asked to play a character, its inner monologue drifts into a formal, task-solving register and it forgets who it is. The proposed method, Role-Aware Reasoning (RAR), attacks both failures directly. Role Identity Activation distills thought traces that are continually anchored to the character's emotions, experiences, standpoints, and motivations, counteracting 'attention diversion'. Reasoning Style Optimization then trains the model on paired preference data so that it adopts fact-focused reasoning for logical-analysis scenes and character-knowledge reasoning for vivid-interaction scenes, counteracting 'style drift'. A careful reader would care because, if the claim holds, RAR is a concrete recipe for making smaller open LLMs think before answering in character, and it explains why naive reasoning distillation can hurt role-play.

What carries the argument

The load-bearing object is the role-aware thought trace, produced by two prompt-controlled generators. The RIA scaffold converts a character profile into a step-by-step thinking template (identity, emotion, experience, standpoint, motivation, plan), so the trace is generated with the character as the grammatical and cognitive subject, and the student is trained on these traces by supervised distillation ($L_{\mathrm{RIA}}$). The RSO stage constructs preference pairs from the same instructions: positive pairs combine the fact-focused prompt $C_{\mathrm{Fact}}$ with logical-analysis scenes $X_{\mathrm{Logic}}$ and the character-knowledge prompt $C_{\mathrm{Know}}$ with vivid-interaction scenes $X_{\mathrm{Story}}$, while negative pairs swap the prompts ($\mathcal{D}^+_S$ and $\mathcal{D}^-_S$); the student then optimizes the contrastive loss $L_{\mathrm{RSO}}=-\mathbb{E}\log\sigma(\pi_{\mathrm{LLM}}(y^+|x)-\pi_{\mathrm{LLM}}(y^-|x))$. The mechanism that carries the argument is the combination: traces that are identity-anchored at every step keep the agent in character, and traces whose register is chosen to fit the scene keep the character's voice believable; the paper shows the two effects separately in ablations and in the hidden-state separation (t-SNE) between the two registers.

What would settle it

Train Reasoning Style Optimization with the scene-to-style pairing inverted — fact-focused prompts on vivid-interaction scenes and character-knowledge prompts on logical-analysis scenes — and compare on CharacterBench. If the paper's account is right, the inverted model should lose the believability, engagement, and style advantages of RAR; if it does not, the gains come from the contrastive preference training itself rather than from the hypothesized scene-to-style mapping. A complementary check is to keep both prompts but feed the model mislabeled scene types at inference time and measure whether the final responses degrade.

Watch

Extended reading notes

Core claim

The paper's claim, stated on its own terms, is that LLMs can be taught to 'think in character' by a two-stage distillation from a large reasoning model into a plain LLM, and that this beats both raw supervised role-play training and direct reasoning distillation. In the first stage, Role Identity Activation, the teacher is prompted with a fixed scaffold — 'I am {character}; first I feel...; second, based on my experience/knowledge/stance...; then I need to consider...; so I'm planning to...' — so every generated thought trace answers to the character's emotions, experiences, standpoints, and motivations rather than to the task. In the second stage, Reasoning Style Optimization, the teacher generates positive traces where the reasoning style matches the scene type (fact-focused thought for logical analysis, character-knowledge thought for vivid interaction) and negative traces where the pairing is mismatched, and the student is trained with a contrastive preference loss to prefer the matched style. On CharacterBench, RAR posts the highest average score (3.69 versus 3.57 for the strongest reasoning baseline), and on SocialBench it posts the highest average (65.4 versus 61.1); ablations show that removing either stage lowers the average, and that each RIA element — emotion, experience, standpoint, motivation — contributes to specific persona dimensions.

Load-bearing premise

The second stage presupposes that every role-playing scene falls into exactly one of two categories — logical analysis, which should be thought through with fact-focused reasoning, or vivid interaction, which should be thought through with character-knowledge reasoning — and the paper (Section 3.2 and Appendix A) never describes how scenes were assigned to these categories in the training data, nor checks that the same assignment is sensible across the 3,956 characters in evaluation.

Editorial extensions

If this is right

  • Reasoning helps role-play only when it is role-anchored: the Distill and MoreThink baselines show that generic or unguided long thought traces plateau or regress, especially on memory and persona consistency, while RAR's anchored traces improve them.
  • Each of the four identity elements is doing real work: removing emotion lowers emotion self-regulation and empathy, removing experience lowers memory consistency, and removing standpoint or motivation lowers behavior and attribute consistency (Figure 2).
  • The model genuinely learns two reasoning registers and a switching rule: t-SNE shows hidden states for fact-focused and character-knowledge thought cleanly separated in RAR but intermingled in the Vanilla model (Figure 3).
  • RIA and RSO are complementary: without RIA the model loses memory consistency, behavior consistency, human-likeness, and engagement; without RSO the reasoning loses coherence and effectiveness while staying role-relevant (Tables 3-4).
  • The recipe transfers from a teacher reasoning model to an 8-billion-parameter base LLM via LoRA, so the gains are not contingent on running a large reasoning model at inference time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because RSO fixes a single binary style axis and assumes it fits every character, a natural extension the authors do not pursue is to learn the scene-to-style assignment per character, for instance deriving the preferred reasoning register from the character's own profile, and testing whether that beats the fixed mapping.
  • The RIA scaffold is task-independent, so the same 'I am ..., I feel ..., I plan to ...' template could plausibly steady other consistency-critical generation settings, such as long-horizon dialogue agents or narrative generation, where models drift into a generic assistant voice; the paper only tests the role-play setting.
  • The benchmark gains are produced by automated judges, and RAR's traces are deliberately less concise; whether the vividness that wins those judgments also sustains human engagement over long conversations is an open question that a human-preference study would settle.
  • The paper only trains the distilled student; if RIA works by keeping the model's attention on the role, the same prompt scaffold applied directly to a frozen reasoning model at inference time is a cheap test that would isolate the mechanism from the distillation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Role-Aware Reasoning (RAR), a two-stage training method for role-playing agents. Stage one, Role Identity Activation (RIA), uses LRM-generated reasoning traces prompted with character elements (emotion, experience, standpoint, motivation) and distills them into a LLaMA-3-8B student. Stage two, Reasoning Style Optimization (RSO), constructs preference pairs by prompting the LRM with two hand-designed styles (fact-focused and character-knowledge-focused) over two scenario types (logical analysis and vivid interaction), then applies contrastive preference optimization. The authors evaluate on CharacterBench and SocialBench, reporting that RAR outperforms several baselines including Distill, Thinking Modes, Neeko, and Character-GLM. Ablations and additional analyses (reason-trace quality via GPT-4o, t-SNE, case studies) are used to attribute the gains to RIA and RSO.

Significance. If the claims hold, RAR is a practical and clearly specified recipe for adding character-consistent reasoning to open-weight LLMs, with meaningful gains over several established baselines on two public benchmarks. The evaluation is largely non-circular: CharacterBench and SocialBench are not used to fit any parameters, and the method description is detailed enough to reimplement. The main risks are that the RSO component rests on an undisclosed and unvalidated scenario taxonomy, that all reported results come from single runs without variance or significance statistics, and that RAR trains on strictly more synthetic data than the Distill baseline, making the isolated contribution of each component unclear. The central idea—adaptive internal reasoning style conditioned on role and context—is interesting and worth pursuing, but the current evidence is not yet sufficient to support the abstract's claim of significant improvement.

major comments (3)
  1. [Section 3.2, Eqs. (3)-(5); Appendix A Figures 5-6; Appendix C] The RSO stage is load-bearing, but its preference construction is not validated. The assignment of RoleBench dialogues to X_Logic and X_Story is never described, and Appendix C only acknowledges that the scenario taxonomy is 'predefined' without justifying it. If the labels are arbitrary, Eq. (3) builds preference pairs on a spurious ground truth. More importantly, the student model at inference receives only the dialogue context, not the scenario label; for RSO to work, the scenario must be inferable from x. The t-SNE in Figure 3 is generated by prompting the model with explicit style prompts, which demonstrates steerability by a prompt, not automatic style selection from context. The authors should specify how X_Logic and X_Story were labeled (including inter-annotator agreement, if any), and provide a direct test of whether the trained model produces the intended style when the scenario label is withheld, e.g., by comparing style-selection accuracy or by evaluating on held-out dialogues with known scenario labels.
  2. [Tables 1-4; abstract] Every quantitative claim of improvement rests on single-run numbers without standard deviations, confidence intervals, or significance tests. For example, in Table 1 the CharacterBench average is 3.69 for RAR versus 3.57 for Distill, and in Table 2 the SocialBench average is 65.4 versus 61.1. These differences could be within run-to-run noise, especially with LoRA fine-tuning and GPT-4o-based evaluation (Table 4), which is itself stochastic. The abstract's word 'significantly' is therefore unsupported. The authors should report multiple seeds with means and variances, and apply an appropriate significance test or, if that is infeasible, temper the claim.
  3. [Section 4.1 and Appendix E.1] The RAR training data is a superset of the Distill baseline's data: RAR uses DR obtained by prompting the LRM with the RIA instruction CR (Eq. 1) and then adds RSO preference data, while Distill is trained only on the original RoleBench-Train instructions. Any extra data can improve downstream performance regardless of the method's conceptual contribution. The paper should control for data volume and compute, e.g., by training a Distill model on the same DR set (without RSO) or on an equal number of generic reasoning traces, to show that the gains come from role-aware reasoning rather than from more training examples.
minor comments (4)
  1. [Table 1 caption] The caption says 'Performance comparison of different methods on the SocialBench', but the columns (M C, F A, BCK, ACb, etc.) are CharacterBench dimensions; the text correctly describes Table 1 as CharacterBench in Section 4.2. The caption should be corrected.
  2. [Section 4.4, Table 4 and Appendix E.2] The main text says RAR 'is less concise than Distill and its ablations' and that 'this trade-off appears beneficial', while Appendix E.2 states that 'RAR consistently achieves the highest scores across all four metrics: Coherence, Role Relevance, Effectiveness, and Conciseness.' Table 4 shows RAR with the lowest Conciseness score (1.81 vs. 2.06 for Distill), so the two passages are contradictory; the appendix text should be aligned with the table.
  3. [Figure 3] The caption and the 'Eliminate the Style Drift' paragraph do not state which data points are plotted or how the 'different reasoning style prompts' are applied. The text says 'responses generated from 500 in different scenarios randomly sampled data points', but the figure caption says the hidden states are for 'responses generated under different reasoning style prompts'. Clarify the protocol: are these prompts the style prompts from RSO, and are the scenarios labeled or inferred?
  4. [References and reproducibility] No URL or release information is provided for the code or the constructed datasets (DR, D+S, D-S); given that the paper's contribution is a training method, releasing these artifacts would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RAR's central claim is tested on external benchmarks not used to fit any parameter; the RSO/RIA training signals are construction steps, not predictions.

full rationale

The paper's central quantitative claims (Tables 1 and 2) evaluate RAR on CharacterBench and SocialBench, which are public benchmark suites and are not used to fit any parameter of the model. RIA and RSO are training procedures defined on RoleBench-Train: RIA supervises the model to follow character-activation prompts, and RSO constructs preference pairs according to the style-scenario mapping in Eqs. 3-4. The reported benchmark gains can therefore falsify the method even though the style targets in Eqs. 3-5 are by construction aligned with the desired reasoning styles. The t-SNE visualization and GPT-4o trace evaluation in Section 4.4 are self-referential checks: the t-SNE is computed under explicit style prompts, and the rubric resembles the training objectives, but these analyses are not the load-bearing evidence for the main contribution. No load-bearing self-citation is present: the sole first-author prior work [11] is cited only as an example of early in-context-learning role-playing, and the co-authored related-work citations [42,44] do not provide the method's justification. Appendix C explicitly acknowledges the predefined RSO taxonomy as a limitation rather than presenting it as an external, forced result. The undisclosed X_Logic/X_Story labeling is a reproducibility and validity gap, but it is not a circular reduction: the central comparison against external benchmarks remains an independent test of the trained system.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method rests on the teacher model's ability to generate character-faithful thoughts, on a manually fixed binary mapping from scene type to reasoning style, and on the validity of two automated benchmarks. No free parameters are fitted to the evaluation, but several training hyperparameters and prompt templates are hand-chosen and not searched.

free parameters (4)
  • Learning rate for RIA stage = 1e-4
    Chosen by hand in Appendix E.1; no search reported; affects distillation quality.
  • Learning rate for RSO stage = 5e-5
    Chosen by hand in Appendix E.1; no search reported; affects preference optimization.
  • LoRA rank / alpha / dropout = rank 64, alpha 16, dropout 0.1
    Fixed across all models for fair comparison; standard choice, not derived from data.
  • Maximum sequence length (reasoning vs non-reasoning) = 7096 / 1024
    Set per model type in Appendix E.1; could affect trace length and style characteristics.
assumptions (5)
  • domain assumption The teacher LRM Qwen2-32B produces high-quality, character-faithful reasoning traces when prompted with the RIA instruction.
    The entire RIA distillation relies on teacher outputs being role-consistent; no human validation of trace quality is reported before training (Section 3.1).
  • ad hoc to paper The binary scenario taxonomy (X_Logic vs X_Story) and the fact-vs-knowledge style mapping are correct and sufficient for all role-playing contexts.
    Section 3.2, Eqs. 3-4 defines the mapping; the paper gives no procedure for classifying RoleBench scenes and no justification beyond two illustrative dialogue cases.
  • standard math The preference loss in Eq. 5 (sigmoid of log-likelihood difference) trains style adaptation without a reference policy or reward model.
    This is a known contrastive objective, though the missing reference term is not discussed; it is a standard form of preference loss.
  • domain assumption CharacterBench and SocialBench automated judges measure role-playing quality accurately enough to rank methods.
    Section 4.1 describes the benchmarks; the paper relies on benchmark protocols and does not run its own human evaluation of the final models.
  • domain assumption Base LLaMA-3-8B with LoRA can absorb the distilled reasoning and style skills at 8B scale.
    The experiments assume the student model is capable; the limitations section notes only one base model was tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning." pith.science (2026). https://pith.science/paper/L3SSU26E

@misc{pith2026250601748,
  author       = {Pith},
  title        = {Pith review of: Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3SSU26E}},
  note         = {Machine review of arXiv:2506.01748}
}
read the original abstract

The advancement of Large Language Models (LLMs) has spurred significant interest in Role-Playing Agents (RPAs) for applications such as emotional companionship and virtual interaction. However, recent RPAs are often built on explicit dialogue data, lacking deep, human-like internal thought processes, resulting in superficial knowledge and style expression. While Large Reasoning Models (LRMs) can be employed to simulate character thought, their direct application is hindered by attention diversion (i.e., RPAs forget their role) and style drift (i.e., overly formal and rigid reasoning rather than character-consistent reasoning). To address these challenges, this paper introduces a novel Role-Aware Reasoning (RAR) method, which consists of two important stages: Role Identity Activation (RIA) and Reasoning Style Optimization (RSO). RIA explicitly guides the model with character profiles during reasoning to counteract attention diversion, and then RSO aligns reasoning style with the character and scene via LRM distillation to mitigate style drift. Extensive experiments demonstrate that the proposed RAR significantly enhances the performance of RPAs by effectively addressing attention diversion and style drift.

Figures

Figures reproduced from arXiv: 2506.01748 by the authors.

Figure 1
Figure 1. Overview of the proposed RAR. Given a user query, a LRM generates structured thoughts. However, traditional reasoning may suffer from attention diversion and style drift, leading to generic, out-of-character responses. To address this, our method incorporates RIA and RSO. RIA activates key role traits (e.g., emotions, motivations) to distill role-consistent thoughts. RSO guides the model to generate reasoning traces… view at source ↗
Figure 2
Figure 2. Analysis of RIA components’ impact on CharacterBench persona metrics. The figure [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The t-SNE visualization of hidden states from different layers of the RAR model for [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The prompt for RIA. The thought process generated this time must conform to the following requirements to match the character and the atmosphere of the current Context Type. Style Core: Vivid and imaginative / Rigorous and logical / Intuition-driven and as￾sociative Fo…
Figure 5
Figure 5. Figure 5: The prompt for logical scenarios of the RSO. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: The prompt for vivid scenarios of the RSO. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: The prompt for reason trace evaluation on coherence. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: The prompt for reason trace evaluation on relevance. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The prompt for reason trace evaluation on effectiveness. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: The prompt for reason trace evaluation on conciseness. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 16 canonical work pages

  1. [1]

    The oscars of ai theater: A survey on role-playing with language models.arXiv preprint arXiv:2407.11484, 2024

    Nuo Chen, Yan Wang, Yang Deng, and Jia Li. The oscars of ai theater: A survey on role-playing with language models.arXiv preprint arXiv:2407.11484, 2024

    arXiv 2024

  2. [2]

    Speak from heart: An emotion-guided llm-based multimodal method for emotional dialogue generation

    Chenxiao Liu, Zheyong Xie, Sirui Zhao, Jin Zhou, Tong Xu, Minglei Li, and Enhong Chen. Speak from heart: An emotion-guided llm-based multimodal method for emotional dialogue generation. InProceedings of the 2024 International Conference on Multimedia Retrieval, ICMR ’24, page 533–542, New York, NY , USA, 2024. Association for Computing Machinery. ISBN 979...

    doi:10.1145/3652583 2024

  3. [3]

    Bernstein

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, New York, NY , USA, 2023. Association for Computing Machinery. ISBN 9798400701320. doi...

    arXiv 2023

  4. [4]

    CharacterGLM: Customizing social characters with large language models

    Jinfeng Zhou, Zhuang Chen, Dazhen Wan, Bosi Wen, Yi Song, Jifan Yu, Yongkang Huang, Pei Ke, Guanqun Bi, Libiao Peng, JiaMing Yang, Xiyao Xiao, Sahand Sabour, Xiaohan Zhang, Wenjing Hou, Yijia Zhang, Yuxiao Dong, Hongning Wang, Jie Tang, and Minlie Huang. CharacterGLM: Customizing social characters with large language models. In Franck Dernoncourt, Daniel ...

    2024

  5. [5]

    Neeko: Leveraging dynamic LoRA for efficient multi-character role-playing agent

    Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, and Liehuang Zhu. Neeko: Leveraging dynamic LoRA for efficient multi-character role-playing agent. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12540–12557, Miami, Florida, US...

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

  6. [6]

    Character-LLM: A trainable agent for role-playing

    Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-LLM: A trainable agent for role-playing. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13153–13187, Singapore, December 2023. Association for Computational Linguistics

    work page 2023

  7. [7]

    SocialBench: Sociality evaluation of role-playing conversational agents

    Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Gao Xing, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, and Fei Huang. SocialBench: Sociality evaluation of role-playing conversational agents. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of the Association for Computational Linguistics: ACL 2024, pages 2108–2126, Bangkok, Thai...

    doi:10.18653/v1/2024.findings-acl.125 2024

  8. [8]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    arXiv 2024

Show all 56 references
  1. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  2. [10]

    Reasoning does not necessarily improve role-playing ability.arXiv preprint arXiv:2502.16940, 2025

    Xiachong Feng, Longxu Dou, and Lingpeng Kong. Reasoning does not necessarily improve role-playing ability.arXiv preprint arXiv:2502.16940, 2025. 10

    2025 arXiv

  3. [11]

    Enhancing personalized dialogue generation with contrastive latent variables: Combining sparse and dense persona

    Yihong Tang, Bo Wang, Miao Fang, Dongming Zhao, Kun Huang, Ruifang He, and Yuexian Hou. Enhancing personalized dialogue generation with contrastive latent variables: Combining sparse and dense persona. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors,Proceedings...

    2023 doi

  4. [12]

    Multi-party chat: Conversational agents in group settings with humans and models.arXiv preprint arXiv:2304.13835, 2023

    Jimmy Wei, Kurt Shuster, Arthur Szlam, Jason Weston, Jack Urbanek, and Mojtaba Komeili. Multi-party chat: Conversational agents in group settings with humans and models.arXiv preprint arXiv:2304.13835, 2023

    2023 arXiv

  5. [13]

    Editing personality for llms.arXiv preprint arXiv:2310.02168, 2023

    Shengyu Mao, Ningyu Zhang, Xiaohan Wang, Mengru Wang, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. Editing personality for llms.arXiv preprint arXiv:2310.02168, 2023

    2023 arXiv

  6. [14]

    Does role-playing chatbots capture the character personalities? assessing personality traits for role-playing chatbots.arXiv preprint arXiv:2310.17976, 2023

    Xintao Wang, Yaying Fei, Ziang Leng, and Cheng Li. Does role-playing chatbots capture the character personalities? assessing personality traits for role-playing chatbots.arXiv preprint arXiv:2310.17976, 2023

    2023 arXiv

  7. [15]

    InCharacter: Evaluating personality fidelity in role-playing agents through psychological interviews

    Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, Jiangjie Chen, Cheng Li, and Yanghua Xiao. InCharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. In Lun-Wei Ku, And...

    2024

  8. [16]

    A survey on in-context learning.arXiv preprint arXiv:2301.00234, 2022

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning.arXiv preprint arXiv:2301.00234, 2022

    2022 arXiv

  9. [17]

    Character is destiny: Can large language models simulate persona-driven decisions in role-playing?arXiv preprint arXiv:2404.12138, 2024

    Rui Xu, Xintao Wang, Jiangjie Chen, Siyu Yuan, Xinfeng Yuan, Jiaqing Liang, Zulong Chen, Xiaoqing Dong, and Yanghua Xiao. Character is destiny: Can large language models simulate persona-driven decisions in role-playing?arXiv preprint arXiv:2404.12138, 2024

    2024 arXiv

  10. [18]

    Llms + persona-plug = personalized llms.arXiv preprint arXiv:2409.11901, 2024

    Jiongnan Liu, Yutao Zhu, Shuting Wang, Xiaochi Wei, Erxue Min, Yu Lu, Shuaiqiang Wang, Dawei Yin, and Zhicheng Dou. Llms + persona-plug = personalized llms.arXiv preprint arXiv:2409.11901, 2024

    2024 arXiv

  11. [19]

    Identity-driven hierarchical role-playing agents.arXiv preprint arXiv:2407.19412, 2024

    Libo Sun, Siyuan Wang, Xuanjing Huang, and Zhongyu Wei. Identity-driven hierarchical role-playing agents.arXiv preprint arXiv:2407.19412, 2024

    2024 arXiv

  12. [20]

    IBSEN: Director-actor agent collab- oration for controllable and interactive drama script generation

    Senyu Han, Lu Chen, Li-Min Lin, Zhengshan Xu, and Kai Yu. IBSEN: Director-actor agent collab- oration for controllable and interactive drama script generation. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association f...

    2024 doi

  13. [21]

    From role-play to drama- interaction: An LLM solution

    Weiqi Wu, Hongqiu Wu, Lai Jiang, Xingyuan Liu, Hai Zhao, and Min Zhang. From role-play to drama- interaction: An LLM solution. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of the Association for Computational Linguistics: ACL 2024, pages 3271–3290, Bangko...

    2024

  14. [22]

    A multi- task role-playing agent capable of imitating character linguistic styles.arXiv preprint arXiv:2411.02457, 2024

    Siyuan Chen, Qingyi Si, Chenxu Yang, Yunzhi Liang, Zheng Lin, Huan Liu, and Weiping Wang. A multi- task role-playing agent capable of imitating character linguistic styles.arXiv preprint arXiv:2411.02457, 2024

    2024 arXiv

  15. [23]

    Guess what i am thinking: A benchmark for inner thought reasoning of role-playing language agents.arXiv preprint arXiv:2503.08193, 2025

    Rui Xu, MingYu Wang, XinTao Wang, Dakuan Lu, Xiaoyu Tan, Wei Chu, and Yinghui Xu. Guess what i am thinking: A benchmark for inner thought reasoning of role-playing language agents.arXiv preprint arXiv:2503.08193, 2025

    2025 arXiv

  16. [24]

    Coser: Coordinating llm-based persona simulation of established roles.arXiv preprint arXiv:2502.09082, 2025

    Xintao Wang, Heng Wang, Yifei Zhang, Xinfeng Yuan, Rui Xu, Jen tse Huang, Siyu Yuan, Haoran Guo, Jiangjie Chen, Wei Wang, Yanghua Xiao, and Shuchang Zhou. Coser: Coordinating llm-based persona simulation of established roles.arXiv preprint arXiv:2502.09082, 2025

    2025

  17. [25]

    A survey on post-training of large language models.arXiv preprint arXiv:2503.06072, 2025

    Guiyao Tie, Zeli Zhao, Dingjie Song, Fuyang Wei, Rong Zhou, Yurou Dai, Wen Yin, Zhejian Yang, Jiangyue Yan, Yao Su, et al. A survey on post-training of large language models.arXiv preprint arXiv:2503.06072, 2025

    2025 arXiv

  18. [26]

    Llm post-training: A deep dive into reasoning large language models.arXiv preprint arXiv:2502.21321, 2025

    Komal Kumar, Tajamul Ashraf, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, Phillip HS Torr, Salman Khan, and Fahad Shahbaz Khan. Llm post-training: A deep dive into reasoning large language models.arXiv preprint arXiv:2502.21321, 2025. 11

    2025 arXiv

  19. [27]

    From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419, 2025

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, et al. From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419, 2025

    2025 arXiv

  20. [28]

    Towards large reasoning models: A survey of reinforced reasoning with large language models.arXiv preprint arXiv:2501.09686, 2025

    Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, et al. Towards large reasoning models: A survey of reinforced reasoning with large language models.arXiv preprint arXiv:2501.09686, 2025

    2025 arXiv

  21. [29]

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567, 2025

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wangxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567, 2025

    2025 arXiv

  22. [30]

    Thinking machines: A survey of llm based reasoning strategies.arXiv preprint arXiv:2503.10814, 2025

    Dibyanayan Bandyopadhyay, Soham Bhattacharjee, and Asif Ekbal. Thinking machines: A survey of llm based reasoning strategies.arXiv preprint arXiv:2503.10814, 2025

    2025 arXiv

  23. [31]

    Trade-offs in large reasoning models: An empirical analysis of deliberative and adaptive reasoning over foundational capabilities.arXiv preprint arXiv:2503.17979, 2025

    Weixiang Zhao, Xingyu Sui, Jiahe Guo, Yulin Hu, Yang Deng, Yanyan Zhao, Bing Qin, Wanxiang Che, Tat-Seng Chua, and Ting Liu. Trade-offs in large reasoning models: An empirical analysis of deliberative and adaptive reasoning over foundational capabilities.arXiv preprint arXiv:2...

    2025

  24. [32]

    Gemini 2.0 flash thinking experimental model 01-21.Google DeepMind’s Blog, 2025

    Google DeepMind. Gemini 2.0 flash thinking experimental model 01-21.Google DeepMind’s Blog, 2025. URLhttps://deepmind.google/technologies/gemini/

    2025

  25. [33]

    Team Kimi, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599, 2025

    2025 arXiv

  26. [34]

    Qwq-32b: Embracing the power of reinforcement learning.Qwen’s Blog, 2025

    Team Qwen. Qwq-32b: Embracing the power of reinforcement learning.Qwen’s Blog, 2025. URL https://qwenlm.github.io/blog/qwq-32b

    2025

  27. [35]

    Openai o3-mini system card.OpenAI’s Blog, 2025

    OpenAI. Openai o3-mini system card.OpenAI’s Blog, 2025. URL https://openai.com/index/ o3-mini-system-card

    2025

  28. [36]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  29. [37]

    Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    2025 arXiv

  30. [38]

    Small models struggle to learn from strong reasoners.arXiv preprint arXiv:2502.12143, 2025

    Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubra- manian, and Radha Poovendran. Small models struggle to learn from strong reasoners.arXiv preprint arXiv:2502.12143, 2025

    2025

  31. [39]

    Llms can easily learn to reason from demonstrations structure, not content, is what matters!arXiv preprint arXiv:2502.07374, 2025

    Dacheng Li, Shiyi Cao, Tyler Griggs, Shu Liu, Xiangxi Mo, Shishir G Patil, Matei Zaharia, Joseph E Gonzalez, and Ion Stoica. Llms can easily learn to reason from demonstrations structure, not content, is what matters!arXiv preprint arXiv:2502.07374, 2025

    2025 arXiv

  32. [40]

    Open Thoughts

    OpenThoughts Team. Open Thoughts. https://open-thoughts.ai, January 2025

    2025

  33. [41]

    Let’s verify step by step.arXiv preprint arXiv:2305.20050, 2023

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step.arXiv preprint arXiv:2305.20050, 2023

    2023 arXiv

  34. [42]

    Benchmarking and improving large vision-language models for fundamental visual graph understanding and reasoning.arXiv preprint arXiv:2412.13540, 2024

    Yingjie Zhu, Xuefeng Bai, Kehai Chen, Yang Xiang, Jun Yu, and Min Zhang. Benchmarking and improving large vision-language models for fundamental visual graph understanding and reasoning.arXiv preprint arXiv:2412.13540, 2024

    2024 arXiv

  35. [43]

    Alphazero-like tree-search can guide large language model decoding and training

    Ziyu Wan, Xidong Feng, Muning Wen, Stephen Marcus McAleer, Ying Wen, Weinan Zhang, and Jun Wang. Alphazero-like tree-search can guide large language model decoding and training. InForty-first International Conference on Machine Learning, 2024

    2024

  36. [44]

    Memory-augmented query reconstruction for llm-based knowledge graph reasoning.arXiv preprint arXiv:2503.05193, 2025

    Mufan Xu, Gewen Liang, Kehai Chen, Wei Wang, Xun Zhou, Muyun Yang, Tiejun Zhao, and Min Zhang. Memory-augmented query reconstruction for llm-based knowledge graph reasoning.arXiv preprint arXiv:2503.05193, 2025

    2025 arXiv

  37. [45]

    Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022. 12

    2022

  38. [46]

    Noah Wang, Z.y. Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, Man Zhang, Zhaoxiang Zhang, Wanli Ouyang, Ke Xu, Wenhao Huang, Jie Fu, and Junran Peng. RoleLLM: Benchmarking, eliciting, and enhancing role-playing abil...

    2024

  39. [47]

    Characterbench: Benchmarking character customization of large language models

    Jinfeng Zhou, Yongkang Huang, Bosi Wen, Guanqun Bi, Yuxuan Chen, Pei Ke, Zhuang Chen, Xiyao Xiao, Libiao Peng, Kuntian Tang, et al. Characterbench: Benchmarking character customization of large language models. InProceedings of the AAAI Conference on Artificial Intelligence, v...

    2025

  40. [48]

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

    Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-acl.878

    2024 doi

  41. [49]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  42. [50]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  43. [51]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System D...

    2024

  44. [52]

    Safechain: Safety of language models with long chain-of-thought reasoning capabilities

    Fengqing Jiang, Zhangchen Xu, Yuetai Li, Luyao Niu, Zhen Xiang, Bo Li, Bill Yuchen Lin, and Radha Poovendran. Safechain: Safety of language models with long chain-of-thought reasoning capabilities. In ICLR 2025 Workshop on Bidirectional Human-AI Alignment

    2025

  45. [53]

    Gpt-4o system card.OpenAI, 2024

    OpenAI. Gpt-4o system card.OpenAI, 2024. URL https://cdn.openai.com/gpt-4o-system-card. pdf

    2024

  46. [54]

    Qlora: efficient finetuning of quantized llms

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY , USA, 2024. Curran Associates Inc

    2024

  47. [56]

    Only through intelligence and wisdom can you control the whole situation and achieve ultimate victory and success

    Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9:2579–2605, 2008. 13 A Method Details Detailed RIA prompt can be found in Figure 4. Detailed RSO prompt can be found in Figure 5-6. B Benchmark To comprehensively...

    2008

  48. [2024]

    doi: 10.18653/v1/2024.findings-acl.196

    Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-acl.196

    2024 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.