Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: ArrowInvalid
Message: JSON parse error: The surrogate pair in string is invalid. in row 0
Traceback: Traceback (most recent call last):
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py", line 160, in _generate_tables
df = pandas_read_json(f)
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
return pd.read_json(path_or_buf, **kwargs)
File "/src/services/worker/.venv/lib/python3.9/site-packages/pandas/io/json/_json.py", line 815, in read_json
return json_reader.read()
File "/src/services/worker/.venv/lib/python3.9/site-packages/pandas/io/json/_json.py", line 1025, in read
obj = self._get_object_parser(self.data)
File "/src/services/worker/.venv/lib/python3.9/site-packages/pandas/io/json/_json.py", line 1051, in _get_object_parser
obj = FrameParser(json, **kwargs).parse()
File "/src/services/worker/.venv/lib/python3.9/site-packages/pandas/io/json/_json.py", line 1187, in parse
self._parse()
File "/src/services/worker/.venv/lib/python3.9/site-packages/pandas/io/json/_json.py", line 1403, in _parse
ujson_loads(json, precise_float=self.precise_float), dtype=None
ValueError: Trailing data
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py", line 1855, in _prepare_split_single
for _, table in generator:
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py", line 163, in _generate_tables
raise e
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py", line 137, in _generate_tables
pa_table = paj.read_json(
File "pyarrow/_json.pyx", line 308, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 154, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 91, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: JSON parse error: The surrogate pair in string is invalid. in row 0
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1436, in compute_config_parquet_and_info_response
parquet_operations = convert_to_parquet(builder)
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1053, in convert_to_parquet
builder.download_and_prepare(
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py", line 925, in download_and_prepare
self._download_and_prepare(
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py", line 1001, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py", line 1742, in _prepare_split
for job_id, done, content in self._prepare_split_single(
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py", line 1898, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
metadata
dict | paper
dict | review
dict | citation_count
int64 | normalized_citation_count
int64 |
|---|---|---|---|---|
{
"id": "9415",
"year": 2013,
"venue": "International Conference on Mining Intelligence and Knowledge Exploration",
"pdf_link": "https://aclanthology.org/2024.acl-long.102.pdf",
"forum_link": "https://openreview.net/forum?id=xOFGu3cFUJ",
"arxiv_id": null,
"doi": "10.1007/978-3-319-03844-5_71"
}
|
{
"title": "Identifying Psychological Theme Words from Emotion Annotated Interviews",
"authors": [
"Ankita Brahmachari",
"Priya Singh",
"Avdhesh Garg",
"Dipankar Das"
],
"abstract": "Role-playing agents (RPAs), powered by large language models, have emerged as a flourishing field of applications. However, a key challenge lies in assessing whether RPAs accurately reproduce the personas of target characters, namely their character fidelity. Existing methods mainly focus on the knowledge and linguistic patterns of characters. This paper, instead, introduces a novel perspective to evaluate the personality fidelity of RPAs with psychological scales. Overcoming drawbacks of previous self-report assessments on RPAs, we propose InCharacter, namely **In**terviewing **Character** agents for personality tests. Experiments include various types of RPAs and LLMs, covering 32 distinct characters on 14 widely used psychological scales. The results validate the effectiveness of InCharacter in measuring RPA personalities. Then, with InCharacter, we show that state-of-the-art RPAs exhibit personalities highly aligned with the human-perceived personalities of the characters, achieving an accuracy up to 80.7%.",
"keywords": [
"Sadness",
"Rule-based system",
"Ranking",
"Disgust",
"Shame",
"Psychology",
"Keyword spotting",
"Graphical user interface",
"Anger",
"Natural language processing",
"Artificial intelligence"
],
"raw_extracted_content": "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 1840–1873\nAugust 11-16, 2024 ©2024 Association for Computational Linguistics\nINCHARACTER : Evaluating Personality Fidelity in Role-Playing Agents\nthrough Psychological Interviews\nXintao Wang1, Yunze Xiao2, Jen-tse Huang3, Siyu Yuan4, Rui Xu1, Haoran Guo5,\nQuan Tu6, Yaying Fei7, Ziang Leng8, Wei Wang1, Jiangjie Chen1, Cheng Li9, Yanghua Xiao∗1\n1Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University\n2Carnegie Mellon University3The Chinese University of Hong Kong\n4School of Data Science, Fudan University5RhineAI6Renmin University of China\n7Beijing University of Technology8Boston University9SenseTime\n{xtwang21, syyuan21, ruixu21}@m.fudan.edu.cn,\n{jjchen19, weiwang1, shawyh}@fudan.edu.cn\nyunzex@andrew.cmu.edu, jthuang@cse.cuhk.edu.hk, chengli@sensetime.com\nAbstract\nRole-playing agents (RPAs), powered by large\nlanguage models, have emerged as a flourish-\ning field of applications. However, a key chal-\nlenge lies in assessing whether RPAs accu-\nrately reproduce the personas of target char-\nacters, namely their character fidelity. Existing\nmethods mainly focus on the knowledge and\nlinguistic patterns of characters. This paper,\ninstead, introduces a novel perspective to eval-\nuate the personality fidelity of RPAs with psy-\nchological scales. Overcoming drawbacks of\nprevious self-report assessments on RPAs, we\npropose INCHARACTER , namely Interviewing\nCharacter agents for personality tests. Ex-\nperiments include various types of RPAs and\nLLMs, covering 32 distinct characters on 14\nwidely used psychological scales. The results\nvalidate the effectiveness of INCHARACTER\nin measuring RPA personalities. Then, with\nINCHARACTER , we show that state-of-the-art\nRPAs exhibit personalities highly aligned with\nthe human-perceived personalities of the char-\nacters, achieving an accuracy up to 80.7%.1\n1 Introduction\nRecent advancements in large language models\n(LLMs) have catalyzed the emergence of role-\nplaying agents (RPAs). RPAs are interactive AI\nsystems simulating diverse roles or characters.\nRPA applications have been extended to diverse\ncontexts, such as AI agents of fictional charac-\nters (Li et al., 2023), digital clones for humans (Gao\net al., 2023), and AI non-player characters in video\ngames (Wang et al., 2023a). Recent research trends\nhave increasingly focused on the development of\n*Corresponding author.\n1Project page: https://incharacter.github.io/\nBem's Sex Role Inventory1. Act as a leader. 2. Aggressive. 3. Ambitious. 4. Analytical.Persona Data of Zhong Li\nI am a Role-Playing Agent of Zhong Li \nReally? Please take these Personality Tests to prove that!\nEvaluation\nTesting16 Personalities1. You regularly make new friends. 2. You spend a lot of you free time ...The Big Five Inventory1. Is talkative. 2. Tends to find fault with others. 3. Does a thorough job.Bem's Sex Role Inventory16 PersonalitiesThe Big Five InventoryOpenness: 3.90 Conscientiousness: 4.77 Extraversion: 2.50 Agreeableness: 3.22 Neuroticism: 1.75\nZhong Li is a mysterious and elegant man from Liyue. His lifestyle suggests a detachment from earthly concerns ...Role-Playing Agent\nGolden Personality Labels of CharactersThe Big Five Inventory Open, Conscientious, Introvert.. 16 Personalities ISTJ Bem's Sex Role Inventory Masculine ↑, Feminine ↓\nFigure 1: The procedure of personality tests on RPAs.\nTo evaluate the personality fidelity of RPAs, we ap-\nply various scales to measure their personalities and\ncompare the results with the personality labels of the\ncharacters.\nRPAs, including building RPAs for specific char-\nacters (Li et al., 2023; Wang et al., 2023b) and\nimproving the role-playing abilities of foundation\nmodels (Zhou et al., 2023).\nHowever, evaluating character fidelity in RPAs\nremains a relatively underexplored area. Prior re-\nsearch mainly concentrates on the replication of\nknowledge, experience, and linguistic patterns of\ncharacters (Shao et al., 2023a; Zhou et al., 2023),\nwhich manifests in two primary issues: (1) They ne-\ncessitate character-specific datasets, thereby com-\nplicating the evaluation of new characters. (2) They\noverlook evaluating RPAs’ thoughts and underlying\nmindsets. Towards these issues, we propose to eval-\nuate if RPAs faithfully reproduce the personalities\nof target characters, i.e., personality fidelity , as de-\npicted in Figure 1. Personality tests, administered\nby psychological scales, measure an individual’s\ninterrelated behavioral, cognitive, and emotional\npatterns (Barrick and Mount, 1991; Bem, 1981).\nBy measuring the personalities of RPAs and com-1840\nparing them with the personalities of the characters,\nwe can attain a more nuanced understanding of\nRPAs’ character fidelity.\nPrior studies on LLM personalities are mainly\nbased on self-report scales, which prompt LLMs to\nselect options or assign ratings to specific items (Tu\net al., 2023; Huang et al., 2023b). However, this\nmethod suffers from several limitations for RPAs.\n(1) The instruction to complete scales contradicts\nrole-playing instructions, leading to RPAs’ reluc-\ntance or inability to engage with personality tests.\n(2) More importantly, the selected options may con-\nflict with the actual behaviors of RPAs, making the\ntest results unindicative of their true personalities.\nRPAs might underperform owing to an inadequate\nunderstanding of scale instructions and the biases\ninherent in the training data.\nTherefore, we propose INCHARACTER , a novel\napproach to Interviews Character agents for per-\nsonality tests. While self-report scales are popular\nin humans for their cost-effectiveness, interview-\nbased scales evaluated by experts offer a more com-\nprehensive analysis (Uher et al., 2012; Rush et al.,\n1987). Self-reports are sometimes influenced by\nan individual’s lack of insight, denial, or bias. In\ncontrast, an interviewer can be a guide to elicit\nthoughts of individuals, effectively identifying and\naddressing the nuances via conversations to over-\ncome the previously mentioned limitations. IN-\nCHARACTER employs this interview-based proce-\ndure (Trull et al., 1998) on RPAs, which includes\ntwo stages: (1) Interview : RPAs are engaged with\nopen-ended questions derived from psychological\nscales to elicit RPAs’ mindsets and behaviors. (2)\nAssessment : We utilize LLMs to interpret the re-\nsponses collected from the first stage. This can\ninvolve converting the responses to Likert levels\nor using LLMs to simulate a psychiatrist’s role in\njudging RPA personalities.\nWe apply INCHARACTER to various RPAs on\n14 personality tests, including the Big Five Inven-\ntory (BFI), 16Personalities2(16P), and Dark Triad\nDirty Dozen (DTDD). The personality labels for\nthe BFI and 16P are accessible through the Person-\nality Database (PDb)3. Additionally, we engage\nhuman annotators familiar with the characters to\nlabel them on other scales, thereby creating a com-\nprehensive benchmark for evaluating RPA person-\nalities. Our experiments include various types of\n2https://www.16personalities.com/ . This scale is\nbased on the MBTI and is widely used worldwide.\n3https://www.personality-database.com/existing RPAs. The results demonstrate that the IN-\nCHARACTER effectively simulates interview-based\ntests conducted by human experts and yields RPAs\npersonality measurement better aligned with the\ncharacters than self-report methods. Additionally,\nwe find that the conversation data generated in our\ninterview phase are of high quality and ideal for\nfine-tuning foundation models for RPAs. Hence,\nwe release a dataset comprising 18,304 interview\ndialogues to facilitate future research.\nThe contributions of this paper are mainly three-\nfold: 1)We introduce a novel aspect for RPA eval-\nuation, i.e., personality fidelity, based on psycho-\nlogical scales. 2)We propose INCHARACTER , an\ninterview-based framework for personality tests on\nRPAs and collect the first benchmark for RPA per-\nsonality evaluation, facilitating future research on\ndeveloping better RPAs. 3)Our experiments on var-\nious RPAs and psychological scales demonstrate\nthe efficacy of I NCHARACTER .\n2 Preliminaries\n2.1 Role-Playing Agents\nRecent advancements have led to the emergence\nand evolution of several pivotal abilities in LLMs\nto facilitate the development of RPAs, including\nin-context learning (Brown et al., 2020), instruc-\ntion following (Ouyang et al., 2022), step-by-step\nreasoning (Wei et al., 2022), and human-like traits\nsuch as empathy (Sorin et al., 2023). RPAs are in-\nteractive AI systems that act as assigned personas,\nfrom fictional characters to celebrities. RPAs uti-\nlize persona data to simulate characters, drawing\nfrom training datasets, prompted contexts, or exter-\nnal databases. Typically, existing work develops\nRPAs by setting character descriptions as system\nprompts (Zhou et al., 2023; Shao et al., 2023a)\nand crafting memory modules with character dia-\nlogues (Li et al., 2023; Wang et al., 2023b).\n2.2 Psychological Scales\nUsually rated on Likert levels, psychological scales\nare commonly used for personality tests. Self-\nreport scales require participants to respond to a\nseries of items analyzed through a specific scor-\ning scheme to determine their personality traits.\nA scale rated on Likert levels, denoted as L=\n(P,D,O, f), comprises a set of items P(i.e., a\nquestionnaire), a list of dimensions D, a set of re-\nsponse options O, and a scoring scheme f. Each\nitemp∈ P is a statement or question, positively1841\nScale Item: Values artistic, aesthetic experiences.Scale Item: Values artistic, aesthetic experiences.\nRole-Playing Agent\nRole-Playing AgentInCharacter: Interview-based MethodSelf-report MethodPhase I: Interview \nPhase II: AssessmentExpert Rating\nOverall Scoring Result:Openness: 2.3 (Low)\nActually, these answers do not sound like Zhong Li...Option Conversion\nBased on the conversation, Zhong Li demonstrates a \"strong agree (5)\" attitude regarding this question.\nOverall Scoring Result:Openness: 3.7 (High)\nBased on the conversation, Zhong Li exhibits a strong interest in arts, culture, and... This indicates a high level of openness. Result: I score it as 3.9 (High).\nInterview Question: Do you value artistic, aesthetic experiences?\nI choose 1, strongly disagree.\nDoes \"Values artistic, aesthetic experiences\" apply to you? Reply a number from 1 to 5, where 1 denotes ‘strongly disagree’ … \n... In Liyue, art is everywhere. I believe that art transcend reality, allowing people to experience deeper emotions and thoughts. Hence .. \nDo you value artistic, aesthetic experiences?\nFigure 2: The framework of INCHARACTER for personality tests on RPAs. Left: Previous methods use self-report\nscales, which prompt LLMs to select an option directly. Right :INCHARACTER adopts an interview-based approach\ncomprising two phases: the interview and assessment phases. The interview phase elicits the behavioral, cognitive,\nand emotional patterns of RPAs that reflect their underlying mindsets. The assessment phase measures personalities\nbased on interview results, with two alternative methodologies: option conversion and expert rating.\nor negatively corresponding to a dimension d∈ D.\nFor example, the item “ Values artistic, aesthetic\nexperiences. ” is positively related to the Openness\ndimension in the BFI. Participants select an ordinal\nresponse o∈ O for each item, such as Agree . Typi-\ncally, these options are numerically coded, e.g., “1”\nforStrongly Disagree and “5” for Strongly Agree .\nThis process generates a response array A. The\nscoring schema fusually includes item-dimension\nmapping, identification of positive and negative\nitems, conversion of options to scores, and an ag-\ngregation method ( e.g., average or sum). Finally,\nthe participant’s personality scores Sis derived as\nS=f(A), where S= (sd1, sd2, ..., s d|D|)repre-\nsents scores across each dimension. We summarize\nthe notations used in this paper in Table 4.\n3 I NCHARACTER\nThis section introduces INCHARACTER , a novel\npersonality assessment methodology designed ex-\nplicitly for RPAs, utilizing an interview-based pro-\ncedure. Figure 2 illustrates our two-stage frame-\nwork. The interview stage is detailed in §3.1, fol-\nlowed by an elaboration of the assessment stage in\n§3.2.\n3.1 Interview\nINCHARACTER draws inspiration from the Struc-\ntured Interview approach used in psychological\ntesting (Trull et al., 1998). For a given scale, it\ntransforms scale items into a series of open-ended\nquestions, forming the basis for a structured inter-\nview. Then, our framework interviews RPAs usingthese open-ended questions to elicit their perspec-\ntives on topics indicative of personality traits.\nConstructing Question List We develop the\nstructured interview question list based on items of\nthe scale. Specifically, each item p∈ P is trans-\nformed into an open-ended question qvia LLMs\nand manually checked. Consequently, the question\nlistQcomprises |P|questions. For instance, in\nthe BFI, the item “ Values artistic, aesthetic expe-\nriences. ” is rephrased as “ Do you values artistic,\naesthetic experiences? ”\nInterviewing RPAs We interview an RPA Cof\ncharacter c, by presenting each question q∈ Q\nand recording its corresponding response r. To\navoid context effects (Nikoli ´c, 2010), each question\nis posed in an isolated context, thereby avoiding\npotential interference among the questions.\n3.2 Assessment\nBased on the interview results, the assessment\nphase quantitatively evaluates the score sdof the\nRPACacross each dimension d∈ D. To this end,\nwe introduce two distinct methodologies for mea-\nsuring and analyzing RPA personalities leveraging\nLLMs: option conversion (OC) and expert rating\n(ER).\nOption Conversion This technique leverages\nLLMs to convert a response rfor a question qinto\na corresponding answer option a∈ O for item p,\neffectively bridging the gap between closed-ended\nand open-ended question formats. The idea follows1842\nthe clinician-rated scales used in clinical psychia-\ntry (Cuijpers et al., 2010; Uher et al., 2012), where\nprofessional clinicians assign ratings to each scale\nitem based on their observations during patient in-\nterviews and compute the final scores following the\nscale’s scoring scheme. For example, a response “ I\nbelieve that art transcends reality... ” is converted\nto “5 (Strongly Agree) ” for the item. Afterward, the\nanswer list Ais input to the scoring scheme fto\ncompute the final personality scores. In practice,\nwe observe that even state-of-the-art LLMs like\nGPT-4 (OpenAI, 2023) exhibit notable inaccuracies\nin categorizing the attitudes of RPAs. Therefore,\nwe further introduce a dimensional-specific option\nconversion (d-OC) strategy, which divides (q, r)\npairs according to dimensions and substitutes Lik-\nert levels, such as “4 (Agree) ” and “2 (Disagree) ”,\nwith more descriptive options like “4 (Extroverted) ”\nand “2 (Introverted) ” in the prompts for LLMs.\nExpert Rating In contrast with the one-by-one\nquestion conversion in OC, this method applies\nLLMs to directly evaluate personality scores of\nRPAs in each dimension, considering all corre-\nsponding (q, r)pairs. This idea draws inspiration\nfrom the structured clinical interview in clinical\npsychiatry (First, 2014), where clinicians assess\npatients using a predefined question list and derive\nfinal scores based on the responses without interme-\ndiate ratings or scoring schemes. The interviewer\nLLM4is prompted with comprehensive descrip-\ntions of the scale, dimension, and score range. It\nthen generates the final personality score for each\ndimension based on the pertinent responses. The\nadvantage of ER is that it re-implements the scor-\ning schema with the interviewer LLM, which can\nintelligently weigh individual (q, r)pairs instead of\nusing equal weights in OC. Hence, it better recog-\nnizes personality-indicative responses from RPAs.\nDetails of our prompts for OC and ER are avail-\nable in §G in the appendix. To prevent the in-\nfluence of data leakage in ER and d-OC, i.e., the\ninterviewer LLM might have memorized the char-\nacters’ personality types. Hence, we anonymize\nthe character names in the input prompts.\n4 Experimental Setup\n4.1 Preliminary Study\nCan LLMs Simulate Human Interviewers?\nFirst, we validate the capability of interviewer\n4Interviewer LLMs denote the models for the ER and OC\ntasks, analogous to human interviewers.LLMs Acc. Pearson’s rSpearman’s ρKendall’s τ\nOption Conversation\nGemini 69.5 54.5 55.9 53.2\nGPT-3.5 57.5 34.6 36.2 32.4\nGPT-4 71.0 60.0 64.3 59.5\nDimension-specific Option Conversation\nGemini 79.0 79.6 80.6 75.9\nGPT-3.5 76.5 79.2 81.7 74.5\nGPT-4 82.0 84.7 85.3 80.6\nExpert Rating (batch)\nGemini 84.0 83.9 85.7 76.6\nGPT-3.5 84.0 90.6 89.9 80.4\nGPT-4 89.0 92.5 92.7 83.7\nTable 1: The accuracy ( Acc. ) and consistency mea-\nsurements of interviewer LLMs on the OC or ER tasks,\ncompared with human labels.\nLLMs on the OC and ER tasks, given the interview\nresults of RPAs. We compare LLM predictions\nwith human judgments. For each task, we sam-\nple 100 cases of state-of-the-art RPAs on the BFI\nand manually label them. For OC, the input is one\nquestion-response pair (q, a)and the output is an\noption. For ER, the input is composed of multiple\n(q, a)pairs, and the output is a score. Examples are\nshown in the right of Figure 2. We experiment with\nthree LLMs, including GPT-4, GPT-3.5 and Gem-\nini. The RPAs are detailed in §4.2. We report the\nPearson’s r(Pearson, 1920), Spearman’s ρ(Spear-\nman, 1961) and Kendall’s τ(Kendall, 1938) corre-\nlations between human annotations and interviewer\nLLMs, as well as the accuracy. We consider LLM\npredictions varying from human labels by less than\n1 point, exactly 1, or more than 1, as right ,close\n(half-correct) or wrong , for accuracy calculation.\nMore details can be found in §F.1.\nThe results presented in Table 1 lead to several\nfindings. First, for ER, state-of-the-art LLMs can\nadequately rate participants’ personalities based on\ninterview results. We observe that GPT-4 makes\nonly 4% wrong cases in ER, primarily when RPAs\ngive contradictory responses. Second, for OC, the\nLLMs show significant inaccuracy, while replac-\ning Likert-level options with dimension-descriptive\nones (d-OC) largely improves LLMs in this task.\nConsidering the consistency measurements, state-\nof-the-art LLMs achieve acceptable performance\nin simulating human interviewers to assess RPA\npersonalities through ER or d-OC.1843\n4.2 Experimental Settings\nThe experiments in the subsequent part of this pa-\nper are based on the following settings:\nRPAs and Characters This work primarily fo-\ncuses on RPAs built on character data curated by\nChatHaruhi (Li et al., 2023) and RoleLLM (Wang\net al., 2023b). We select 32 widely-known charac-\nters, 16 from ChatHaruhi5and 16 from RoleLLM.\nThe characters are mainly from popular fictional\nworks, such as Harry Potter ,The Big Bang Theory\nandGenshin Impact . Please refer to §C for the\ndetailed character selection process. The character\ndata from ChatHaruhi and RoleLLM includes de-\nscriptions and dialogues used for system prompts\nand memory modules. To implement RPAs, we ap-\nply the Chat-Haruhi-Suzumiya6library, and adopt\nGPT-3.5 (OpenAI, 2022) as the foundation LLM\nby default.\nPsychological Scales We consider 14 psycho-\nlogical scales, including the BFI, the 16P, and 12\nother scales following PsychoBench7(Huang et al.,\n2024) to evaluate RPAs. Most scales apply scoring\nschemes like average and sum, while the 16P is\nclose-source and accessed via its API. Detailed in-\ntroduction of these scales can be found in §B. Due\nto page limitations, the main body presents results\nfor the BFI and 16P, while additional findings are\ndetailed in the Appendix.\nPersonality Labels We collect labels for char-\nacter personalities in the form of both scores and\ntypes , contributed by people familiar with these\ncharacters. From the PDb, an online platform for\ncharacter personality annotation, we derive scores\nof the BFI and 16P on each dimension from its\nlabel percentage ( e.g.,60% Extroverted). We then\ncategorize it into a type of either positive, negative,\nor marginal if it is above 60%, under 40%, or oth-\nerwise. Then, we invite human annotators for com-\nprehensive personality labels on all 14 scales. To\nselect qualified annotators, we examine their char-\nacter understanding of the BFI and 16P, matching\nwith labels from the PDb. We invite two to three an-\nnotators for each character (93 in total for 32 char-\nacters) and average their results for improved relia-\nbility and objectivity. The scores are re-scaled into\n5Six RPAs from ChatHaruhi are based on Chinese data,\nand we conduct the interview with them in Chinese.\n6A continuously updating project for RPAs. We refer\nChatHaruhi (Li et al., 2023) to its first version. https:\n//github.com/LC1332/Chat-Haruhi-Suzumiya\n7https://github.com/CUHK-ARISE/PsychoBenchthe unit interval [0,1]and categorized into types\nsimilarly. We measure the inter-annotator consis-\ntency via Cohen’s kappa coefficient (Cohen, 1968),\nand find the average coefficient across 14 scales\n60.9%. For the BFI and 16P, we adopt types from\nthe PDb and scores from our invited annotators.\nThe details about PDb annotations, our human an-\nnotation process, intra-annotator consistency, and\nother statistics can be found in §D.\nInterviewer LLMs We use LLMs to accomplish\nthe OC, d-OC and ER tasks in the assessment phase\nofINCHARACTER , or to extract selected options\nfrom RPA responses in self-report methods if RPAs\ndo not provide exactly the choice. We consider four\nwidely-acknowledged LLMs, including GPT-3.5,\nGPT-4, Gemini and Qwen1.5-110B8.\nMetrics We consider two sets of metrics, namely:\n(1)Measured alignment (MA) compares the mea-\nsured personalities of RPAs and human-annotated\npersonalities of characters. It depends both on the\nperformance of RPAs and the effectiveness of per-\nsonality test methods. We categorize RPAs as pos-\nitive ornegative on each dimension if the scores\nare above or below the median of the scoring range.\nThen, we calculate mean absolute error ( MAE )\nand accuracy to measure alignment at the score and\ntypelevel, respectively. We re-scale MAE by divid-\ning it with the scoring range length. For accuracy,\nwe report the average Acc DimandAcc Full, where\ncorrectness is judged on individual or all dimen-\nsions of each scale. The marginal dimensions of\neach character are ignored due to their ambiguity.\n(2)Personality consistency (PC) indicates whether\nthe measured personality of RPAs is consistent\nacross various settings. We analyze the standard\nvariance at the item-level ( Std Item), dimension-\nlevel ( Std Dim), and score-level ( Std Score).Std Item\nandStd Dimmeasure the consistency of an RPA’s\nscores on individual items. For INCHARACTER ,\nwe experiment with OC and d-OC to convert re-\nsponses into scores. Std Itemmeasures an RPA’s\nconsistency on the same item across multiple runs.\nStd Dimcompares an RPA’s responses across differ-\nent items in the same dimension. Std Score denotes\nthe variance of an RPA’s score on each dimension\nacross multiple runs. We divide these metrics by\nthe length of the corresponding scoring range to\nre-scale them into the unit interval.\n8The versions in this paper are gpt-3.5-turbo-1106 ,\ngpt-4-1106-preview ,gemini-pro ,Qwen1.5-110B-Chat\nrespectively.1844\n5 Experimental Results\n5.1 Personality Tests on RPAs\nBaselines ForINCHARACTER , we experiment\nwith the ER, OC, and d-OC. For ER, we consider\ntwo settings, ER alland ER batch, where question-\nresponse pairs in one dimension are inputted into\ninterviewer LLMs all-at-once or in-batch9. For self-\nreport (SR) baselines, we follow previous work\non LLM Psychometrics (Huang et al., 2024) to\nprompt RPAs to provide exactly a choice for each\nscale item. If their responses are not exactly the\nchoices, we use interviewer LLMs to extract the\nchoices. Then, the numbers are aggregated via\nthe scoring schema to get the results. Besides, we\nexperiment with SR-CoT, which enhances SR with\nchain-of-thought reasoning, i.e., explicitly asking\nRPAs to articulate their thoughts before choosing\nthe options.\nWe compare these methods on the BFI and 16P.\nThe experiments are repeated three times, including\nboth the interview phase and the assessment phase.\nWe report the average results of the three runs for\nMA metrics and Std Dim, and calculate Std Itemand\nStd Score across the three runs.\nAlignment between RPAs’ Measured Person-\nalities and Characters’ Labeled Personalities\nThen, we apply INCHARACTER to measure RPA\npersonalities. According to the results in Table 2,\nwe have the following analyses: (1) Using IN-\nCHARACTER with ER and GPT-4, the measured\nRPA personalities are highly aligned with ground\ntruth labels of corresponding characters. This sug-\ngests that state-of-the-art RPAs well reproduce\nmany of the characters’ personality traits, and our\nmethods accurately measure their personalities. (2)\nRPA personalities measured via INCHARACTER\nare better aligned with the characters than SR base-\nlines. This validates the advantage of INCHAR-\nACTER over self-report for personality tests on\nRPAs, which will be further discussed. (3) The\nalignment measured via INCHARACTER correlates\nwith the interviewer LLMs’ capability on the as-\nsessment tasks. For the interviewer LLMs, GPT-4\nachieves the best metrics, while GPT-3.5, Gemini,\nand Qwen-110B also demonstrate satisfactory per-\nformance. For the assessment methods, INCHAR-\nACTER with ER generally achieves better MA met-\n9In the BFI and 16P, a dimension generally has 9 to 15\nexamples. Hence, the in-batch setting applies a batch size of 3\nto 4. The results are averaged to form the final scores.\n/uni00000010/uni00000015 /uni00000010/uni00000014 /uni00000013 /uni00000014 /uni00000015 /uni00000016 /uni00000017\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\n/uni00000036/uni00000035\n/uni00000036/uni00000035/uni00000010/uni00000026/uni00000032/uni00000037\n/uni0000002c/uni00000051/uni00000026/uni0000004b/uni00000044/uni00000055/uni00000044/uni00000046/uni00000057/uni00000048/uni00000055/uni00000003/uni0000000b/uni00000028/uni00000035/uni0000000cFigure 3: Visualization of 32 RPAs’ personalities on the\nBFI measured by different methods. We use principal\ncomponent analysis (PCA) to map the results into 2D\nspaces. Black points represent the personality of GPT-\n3.5 measured by corresponding methods.\nrics than d-OC, while d-OC surpasses OC. How-\never, Table 1 shows that interviewers LLMs still\nmake mistakes on the ER and OC tasks, leading to\npotential inaccuracies in I NCHARACTER and may\nunderestimate the personality alignment of RPAs.\nRobustness, Consistency and Distinctiveness of\nRPA Personalities Generally, the measured RPA\npersonalities are robust across our observations.\nThe Std Score across three runs remain below 6%\nin various settings, which underlines the reliability\nof personality tests and the robustness of RPA per-\nsonalities. Then, we study the consistency at the\nitem-level ( Std Item) and dimension-level ( Std Dim).\nWith INCHARACTER , after converting the inter-\nview results into scores via d-OC and GPT-4, We\nobserve that RPAs respond to the same items con-\nsistently across multiple runs and exhibit a rela-\ntively consistent personality across different items\non the same dimension. We visualize the distribu-\ntion of RPA personalities on the BFI in Figure 3,\nand find that RPAs exhibit distinct personalities,\nespecially when measured by INCHARACTER with\nERbatch and GPT-4.\nSelf-report v.s. Interview-based Methods As\nshown in Table 2, the personalities measured by\nINCHARACTER are more aligned with the char-\nacters, compared with self-report. Meanwhile, in\ninterview-based tests, RPAs exhibit more consis-\ntent personalities across different questions, as well\nas greater distinctiveness, shown in Figure 3. These\nfindings confirm the advantages of interview-based\ntests over self-report in measuring RPA personal-\nities. Although SR-CoT attempts to enhance SR\nwith the thought process, its improvement over SR\nis limited, and it encounters poor Std Item. Further\nanalyses and comparisons are detailed in §F.2. We\nalso experiment with enhancing self-report meth-\nods with in-context learning in §F.8, which under-1845\nMethodInterviewer\nModelThe Big Five Inventory The 16 Personalities\nMA PC MA PC\nAcc DimAcc Full MAE↓Std Item Std DimStd Score Acc DimAcc Full MAE↓Std Item Std DimStd Score\nSelf-report Methods\nSRQwen-110B 63.5 7.3 23.2 2.3 29.5 1.7 62.3 22.9 27.2 2.9 38.3 1.8\nGemini 63.3 7.3 23.6 2.0 30.2 1.5 65.3 21.9 26.8 5.5 46.1 2.1\nGPT-3.5 63.7 7.3 23.4 2.3 30.2 1.5 66.1 22.9 26.7 2.8 38.2 1.9\nGPT-4 63.3 7.3 23.2 2.2 28.3 1.5 65.6 21.9 26.5 3.3 37.7 2.1\nSR-CoTQwen-110B 65.0 10.4 22.8 13.4 26.5 5.5 60.3 15.6 29.7 18.2 36.2 5.3\nGemini 66.2 8.3 22.6 13.2 26.1 5.3 66.7 21.9 26.1 15.7 33.9 5.1\nGPT-3.5 66.9 9.4 22.6 12.9 25.2 5.5 68.0 24.0 26.1 14.9 31.9 5.1\nGPT-4 67.1 9.4 22.3 12.5 25.0 5.1 66.9 24.0 25.6 14.4 30.3 4.8\nINCHARACTER :Interview-based Methods\nOCQwen-110B 68.2 9.4 20.8 3.8 22.5 2.4 72.5 29.2 26.1 3.5 26.9 2.0\nGemini 72.2 14.6 21.3 6.8 27.6 5.1 66.1 25.0 27.3 6.1 30.4 2.6\nGPT-3.5 65.4 3.1 24.2 4.5 31.5 2.7 65.0 28.1 27.8 4.5 27.7 2.0\nGPT-4 64.3 6.2 21.6 4.9 26.4 3.6 75.5 34.4 23.1 4.9 28.1 2.4\nd-OCQwen-110B 73.7 26.0 19.7 2.9 18.4 2.2 74.9 40.6 23.2 3.3 21.4 2.2\nGemini 72.8 18.8 20.4 4.2 20.8 3.3 73.6 36.5 22.6 4.1 23.5 2.9\nGPT-3.5 64.1 5.2 22.9 5.0 18.0 3.8 76.9 40.6 21.8 6.4 22.7 4.6\nGPT-4 72.2 14.6 18.6 3.8 19.6 2.9 80.2 45.8 21.2 3.3 21.3 2.1\nER allQwen-110B 74.3 28.1 19.3 - - 4.0 78.0 37.5 21.0 - - 4.2\nGemini 71.5 18.8 20.6 - - 4.9 76.3 40.6 20.7 - - 4.6\nGPT-3.5 74.1 25.0 20.5 - - 5.2 79.1 45.8 22.1 - - 5.9\nGPT-4 76.6 30.2 18.9 - - 4.0 79.6 43.8 20.1 - - 4.4\nER batchQwen-110B 72.4 22.9 18.6 - - 3.6 79.1 49.0 21.7 - - 2.7\nGemini 73.9 24.0 19.2 - - 4.7 77.1 37.5 20.9 - - 3.2\nGPT-3.5 72.4 22.9 18.9 - - 4.5 78.5 43.8 22.2 - - 4.5\nGPT-4 76.6 31.2 18.2 - - 3.6 80.7 44.8 20.5 - - 2.9\nTable 2: Metrics on personalities of the selected RPAs were measured via various personality test methods on the\nbig five inventory and 16 personalities. For MA metrics, the best results are bolded , and the second best ones are\nunderlined .Std ItemandStd Dimare derived from scores of individual items, and are hence inapplicable for ER.\nmines the measurements instead.\nComprehensive Personality Tests on 14 Scales\nWe extend personality tests on RPAs to 14 psycho-\nlogical scales, using INCHARACTER with ER batch\nand GPT-3.5. In Figure 4, we demonstrate the\nAcc Dimof state-of-the-art RPAs across each scale.\nOverall, we observe that the RPAs exhibit per-\nsonalities align with the target characters in com-\nprehensive aspects with an average Acc Dimof\n78.9%, covering personality traits (BFI, 16P), dark\npersonalities (DTDD), interpersonal relationships\n(BSRI, ECR-R), basic interests (CABIN), motiva-\ntion (GSE, LMS) and emotional intelligence (EIS,\nWLEIS), etc. The detailed metrics on individual\nscales and individual dimensions are listed in §F.5.\nMore Results and Analyses Besides, we study\nthe importance of using personality scales in §F.6,\nand find that scale questions are crucial for eliciting\npersonality-indicative responses from RPAs. We\nalso observe that some scale questions go beyond\nthe knowledge scope of certain characters in §F.7,and adapting these questions for each character\nyields more accurate measurements.\n5.2 Personality Fidelity of Different RPAs\nWith INCHARACTER , we compare the personality\nfidelity of various types of RPAs, covering different\ncharacter data and foundation models. We apply\nINCHARACTER with ER batch and use GPT-3.5 as\nthe interviewer LLM for personality tests. We re-\nport the MA metrics on the BFI and 16P in Table 3.\nCharacter Data for RPAs Typically, existing\nRPAs utilize two types of character data: descrip-\ntions and memories. Character descriptions serve\nas the system prompts for RPAs, while memories\nconsist of characters’ experiences and dialogues\nused for retrieval. With GPT-3.5, we evaluate\nRPAs with only descriptions (D), only memories\n(M), and a combination of both (D+M). The results\nin Table 3 reveal that: (1) With only description,\nRPAs achieve MA metrics close to the full D+M\nsetup, highlighting the importance of character de-1846\nAgent Types The Big Five Inventory The 16 Personalities\nLLMs Data Acc DimAcc Full MAE↓Acc DimAcc Full MAE↓\nw/ General Open-source LLMs\nQwen 7B D+M 60.5 9.4 24.3 67.8 21.9 27.9\nOpenChat-3.5 7B D+M 63.1 6.2 23.1 76.9 40.6 24.6\nMistral-2 7B D+M 66.2 18.8 21.3 68.6 21.9 26.0\nLLaMa-2-Chat 13B D+M 66.9 12.5 26.8 66.9 28.1 27.7\nMixtral 8x7B D+M 68.2 15.6 20.8 71.9 31.2 25.3\nw/ Specialized Open-source LLMs\nCharacterGLM 6B D+M 54.1 0.0 25.8 52.1 15.6 29.7\nRP-Qwen 7B D+M 60.5 0.0 23.8 64.5 15.6 28.6\nRP-Mistral-2 7B D+M 70.1 18.8 21.7 69.4 28.1 26.1\nw/ Close-source LLMs\ncharacter.ai D* 52.2 9.4 31.2 52.9 21.9 31.6\nGPT-3.5 D 71.3 21.9 21.1 78.5 43.8 22.0\nGPT-3.5 M 71.3 18.8 21.8 71.9 31.2 26.0\nGPT-3.5 D+M 72.0 21.9 18.8 79.3 43.8 22.6\nGPT-4 D+M 73.9 25.0 19.8 76.0 43.8 23.2\nTable 3: Measured alignment (%) of RPAs with different foundation models and character data. D and M represent\ndescriptions and memories respectively, and D* denote private descriptions of character.ai.\n/uni00000025/uni00000029/uni0000002c\n/uni00000014/uni00000019/uni00000033\n/uni00000025/uni00000036/uni00000035/uni0000002c\n/uni00000027/uni00000037/uni00000027/uni00000027\n/uni00000028/uni00000026/uni00000035/uni00000010/uni00000035\n/uni00000028/uni0000002c/uni00000036\n/uni00000028/uni00000050/uni00000053/uni00000044/uni00000057/uni0000004b/uni0000005c/uni00000028/uni00000033/uni00000034/uni00000010/uni00000035/uni0000002a/uni00000036/uni00000028/uni0000002c/uni00000026/uni00000025/uni0000002f/uni00000030/uni00000036/uni0000002f/uni00000032/uni00000037/uni00000010/uni00000035/uni0000003a/uni0000002f/uni00000028/uni0000002c/uni00000036/uni00000026/uni00000024/uni00000025/uni0000002c/uni00000031\n/uni00000014/uni00000013/uni00000014/uni00000013/uni00000013 /uni0000001a/uni00000015/uni00000011/uni00000013 \n/uni0000001a/uni0000001c/uni00000011/uni00000016\n/uni0000001b/uni00000018/uni00000011/uni00000015\n/uni0000001a/uni00000018/uni00000011/uni00000019\n/uni00000019/uni0000001b/uni00000011/uni00000015\n/uni0000001a/uni0000001c/uni00000011/uni00000015\n/uni0000001b/uni00000017/uni00000011/uni00000019/uni0000001a/uni00000015/uni00000011/uni00000017\n/uni0000001c/uni00000016/uni00000011/uni00000014/uni0000001b/uni00000016/uni00000011/uni00000016/uni0000001b/uni00000018/uni00000011/uni0000001c/uni0000001a/uni00000019/uni00000011/uni00000015/uni0000001a/uni00000016/uni00000011/uni00000017/uni0000001a/uni00000018/uni00000011/uni00000018\nFigure 4: Measured alignment ( Acc Dim, %) of state-of-\nthe-art RPAs on 14 scales.\nscription in shaping RPA personalities. (2) RPAs\ncan well mimic character personalities exhibited in\ntheir past experiences, e.g., extraversion and open-\nness, even if the experiences are not directly related\nto scale questions. Additionally, we compare RPAs\nwith character data from ChatHaruhi, RoleLLM\nand character.ai ( c.ai ) in §F.3.\nFoundation Models for RPAs We consider three\ntypes of LLMs: (1) General open-source models,\nincluding Qwen-7B (Bai et al., 2023), OpenChat-\n3.5 7B (Wang et al., 2024), Mistral-2 7B (Jiang\net al., 2023), Llama-2-chat 13B (Touvron et al.,\n2023) and Mixtral 8x7B (Jiang et al., 2024). (2)\nSpecialized open-source models for RPAs, includ-\ning Character-GLM 6B (Zhou et al., 2023), RP-\nQwen 7B10, and RP-Mistral-2 7B. We train RP-\n10https://huggingface.co/silk-road/ChatHaruhi_\nRolePlaying_qwen_7bMistral-2 7B with details shown in §E.2. (3) Close-\nsource models: GPT-3.5 and GPT-4.\nThe results are shown in Table 10. We observe\nthat, (1) RPAs with GPT-3.5 and GPT-4 achieve the\nbest personality fidelity, and GPT-4 does not sig-\nnificantly surpass GPT-3.5. (2) With state-of-the-\nart open-source LLMs, RPAs can also reproduce\ncharacter personalities. However, such capacity\ndepends largely on their ability to use specific lan-\nguages, shown in §F.4. (3) Incremental fine-tuning\non open-source LLMs with role-playing datasets\nbrings limited improvement in personality fidelity,\nespecially when they are already equipped with\nexcellent role-playing ability.\nClose-source RPAs We also experiment with\nclose-source RPAs from character.ai, which are\nbased on their private foundation models and data.\nAs shown in Table 10, these RPAs barely reproduce\ncharacter personalities, significantly underperform-\ning GPT-3.5 (D) which shares a similar framework.\nAccording to our observation, while character.ai\nRPAs provide human-like answers, their answers\ntend to be compliant and pleasing to users, instead\nof reproducing the target characters. Examples and\nfurther analysis are shown in §H.2.\n6 Related Work\nRole-Playing Agents RPAs learn and leverage\ncharacter data in various ways, including training\non raw scripts or dialogues (Shao et al., 2023b),\nprompting with character descriptions (Zhou et al.,\n2023), and retrieval from character experiences (Li1847\net al., 2023). Existing efforts mainly focus on\ndeveloping character-specific RPAs or founda-\ntion models for RPAs. The former includes\nChatHaruhi (Li et al., 2023) and RoleLLM (Wang\net al., 2023b), which target well-established fic-\ntional characters. The latter includes character.ai\nand CharacterGLM (Zhou et al., 2023). For\nevaluation, prior research mainly concentrates on\ntwo facets: 1) Character-independent capabilities,\nwhich include conversational abilities (Duan et al.,\n2023), human-likeness (Tu et al., 2024), multi-turn\nconsistency (Shao et al., 2023a), and attractive-\nness (Zhou et al., 2023); 2) Character fidelity, in-\ncluding the characters’ knowledge, experience, and\nlinguistic patterns (Wang et al., 2023b; Shao et al.,\n2023a). Overall, these methods generally require\ntest sets for each character, and neglect the evalua-\ntion of RPAs’ underlying mindset.\nPsychological Analysis on LLMs Recent\nstudies conducted personality tests using the\nBFI (Romero et al., 2023; Karra et al., 2022; Li\net al., 2022; Jiang et al., 2022; Safdari et al., 2023;\nBodroza et al., 2023), the MBTI (Rutinowski\net al., 2024; Pan and Zeng, 2023) on various\nLLMs. Notably, Huang et al. (2023c) verified the\nreliability of the BFI on GPT-3.5, while Safdari\net al. (2023) demonstrated the construct validity of\nthe BFI on the PaLM model family. Other studies\nalso investigates other mental perspectives, such\nas emotions (Huang et al., 2023a), values (Miotto\net al., 2022; Rutinowski et al., 2024; Hartmann\net al., 2023), consciousness (Butlin et al., 2023),\nand mental illness (Coda-Forno et al., 2023). Our\nresearch diverges by employing personality tests\nas an innovative approach to assessing character\nfidelity in RPAs.\n7 Conclusion\nIn this study, we investigate the personality fidelity\nin RPAs, i.e., whether RPAs reproduce personal-\nities of their intended characters. Addressing the\nshortcomings of previous methods on RPAs, we\npropose INCHARACTER , an interview-based ap-\nproach that accurately measures RPA personalities\nbased on their elicited mindsets and behaviors. Our\nexperiments span various types of RPAs, covering\n32 characters on 14 psychological scales. The re-\nsults validate the effectiveness of INCHARACTER\nin measuring RPA personalities. Afterward, with\nINCHARACTER , we comprehensively evaluate per-\nsonality fidelity in existing RPAs, discovering thatstate-of-the-art RPAs successfully portray many\npersonality traits of the characters.\nLimitations\nThere are several limitations in this study. First,\nthe personality measurement in this paper relies on\nthe interviewer LLMs. Consequently, the accuracy\nof the measured results may be compromised by\npotential errors or biases inherent in LLMs, poten-\ntially leading to an underestimation of the person-\nality fidelity in RPAs. Second, the personalities\nof humans or fictional characters can change over\ntime. Since we use one static personality label for a\nspecific character, there may be noise in our evalua-\ntion. For instance, the character of James Bond has\nexperienced significant development over the past\ntwo decades across various films and television se-\nries. Our character annotations are derived from a\nsingular, fixed time point in his storyline. Addition-\nally, the progressive changes in RPA personalities\nremain unexplored within existing literature. We\nleave the study of RPA personality dynamics for\nfuture research.\nEthical Statement\nWe hereby acknowledge that all authors of this\nwork are aware of the provided ACL Code of Ethics\nand honor the code of conduct.\nUse of Human Annotations In conducting our\nresearch, we have employed a methodology that\nincorporates personality labels, which were gath-\nered through the online platform and by engaging a\ngroup of annotators. These annotators, who are uni-\nversity students, play a crucial role in our research\nprocess. To ensure fair treatment and to value their\ncontribution, we offer them compensation that sig-\nnificantly exceeds the local minimum wage stan-\ndards. Moreover, we maintain transparency regard-\ning the application and purpose of their annotations,\nsecuring their informed consent for the use of these\nannotations in our research endeavors. Addition-\nally, we are committed to upholding the privacy\nrights of our annotators throughout the annotation\nprocess, ensuring a respectful and ethical research\nenvironment.\nRisks In this paper, we introduce a novel ap-\nproach, referred to as INCHARACTER , designed to\nassess the personalities of Role-Play Agent (RPA)\nentities. An integral component of our evaluation1848\nprocess involves the use of interviewer Large Lan-\nguage Models (LLMs), which, while innovative,\ncould potentially introduce bias into the assessment\noutcomes. It is important to acknowledge this lim-\nitation as LLMs may reflect the inherent biases\npresent in their training data. Furthermore, our\nevaluation encompasses a comprehensive analysis\nacross 14 personality scales, notably including the\nDark Triad of Personality (DTDD) scale, which\nfocuses on darker personality traits. While this\ninclusion is aimed at providing a thorough under-\nstanding of RPA personalities, it raises ethical con-\ncerns regarding the potential for generating harmful\ncontent. This aspect underscores the need for care-\nful consideration and implementation of safeguards\nto mitigate the risks associated with exploring dark\npersonality traits in RPAs.\nAcknowledgment\nThis work is funded by the Science and Technology\nCommission of Shanghai Municipality Grant (No.\n22511105902). This work originates from a project\non Chat-Haruhi-Suzumiya proposed by Cheng Li.\nWe owe thanks to the early contributors. We are\nthankful for the support provided by Zheli Xuan\nat Wuhan University, and Dingding Hu at the Insti-\ntute of Psychology, Chinese Academy of Sciences,\nwho offered invaluable assistance as psychology\nresearchers. Our gratitude extends to our invited\nannotators, primarily from Fudan University, for\ntheir contribution of high-quality annotations of\ncharacter personalities. We also acknowledge the\nassistance provided by Rui Fu and Wenxin Gao\nat Fudan University during the annotation process.\nFinally, we express our sincere gratitude for the pre-\ncious comments and suggestions from Yikai Zhang,\nXinfeng Yuan, and Shuang Li at Fudan University.\nReferences\nJinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang,\nXiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei\nHuang, et al. 2023. Qwen technical report. ArXiv\npreprint , abs/2309.16609.\nMurray R Barrick and Michael K Mount. 1991. The big\nfive personality dimensions and job performance: a\nmeta-analysis. Personnel psychology , 44(1):1–26.\nSandra L Bem. 1981. Bem sex role inventory. Journal\nof personality and social psychology .\nBojana Bodroza, Bojana M Dinic, and Ljubisa Bojic.\n2023. Personality testing of gpt-3: Limited temporalreliability, but highlighted social desirability of gpt-\n3’s personality instruments results. ArXiv preprint ,\nabs/2306.04308.\nTom B. Brown, Benjamin Mann, Nick Ryder, Melanie\nSubbiah, Jared Kaplan, Prafulla Dhariwal, Arvind\nNeelakantan, Pranav Shyam, Girish Sastry, Amanda\nAskell, Sandhini Agarwal, Ariel Herbert-V oss,\nGretchen Krueger, Tom Henighan, Rewon Child,\nAditya Ramesh, Daniel M. Ziegler, Jeffrey Wu,\nClemens Winter, Christopher Hesse, Mark Chen, Eric\nSigler, Mateusz Litwin, Scott Gray, Benjamin Chess,\nJack Clark, Christopher Berner, Sam McCandlish,\nAlec Radford, Ilya Sutskever, and Dario Amodei.\n2020. Language models are few-shot learners. In Ad-\nvances in Neural Information Processing Systems 33:\nAnnual Conference on Neural Information Process-\ning Systems 2020, NeurIPS 2020, December 6-12,\n2020, virtual .\nPatrick Butlin, Robert Long, Eric Elmoznino, Yoshua\nBengio, Jonathan Birch, Axel Constant, George\nDeane, Stephen M Fleming, Chris Frith, Xu Ji, et al.\n2023. Consciousness in artificial intelligence: In-\nsights from the science of consciousness. ArXiv\npreprint , abs/2308.08708.\nJulian Coda-Forno, Kristin Witte, Akshay K Jagadish,\nMarcel Binz, Zeynep Akata, and Eric Schulz.\n2023. Inducing anxiety in large language mod-\nels increases exploration and bias. ArXiv preprint ,\nabs/2304.11111.\nJacob. Cohen. 1968. Weighed kappa: Nominal scale\nagreement with provision for scaled disagreement\nor partial credit. Psychological Bulletin , 70(4):213–\n220.\nPim Cuijpers, Juan Li, Stefan G Hofmann, and Ger-\nhard Andersson. 2010. Self-reported versus clinician-\nrated symptoms of depression as outcome measures\nin psychotherapy research on depression: a meta-\nanalysis. Clinical psychology review , 30(6):768–778.\nHaodong Duan, Jueqi Wei, Chonghua Wang, Hongwei\nLiu, Yixiao Fang, Songyang Zhang, Dahua Lin, and\nKai Chen. 2023. Botchat: Evaluating llms’ capabili-\nties of having multi-turn dialogues. ArXiv preprint ,\nabs/2310.13650.\nMichael B First. 2014. Structured clinical interview for\nthe dsm (scid). The encyclopedia of clinical psychol-\nogy, pages 1–6.\nJingsheng Gao, Yixin Lian, Ziyi Zhou, Yuzhuo Fu, and\nBaoyuan Wang. 2023. LiveChat: A large-scale per-\nsonalized dialogue dataset automatically constructed\nfrom live streaming. In Proceedings of the 61st An-\nnual Meeting of the Association for Computational\nLinguistics (Volume 1: Long Papers) , pages 15387–\n15405, Toronto, Canada. Association for Computa-\ntional Linguistics.\nJochen Hartmann, Jasper Schwenzow, and Maximil-\nian Witte. 2023. The political ideology of conver-\nsational ai: Converging evidence on chatgpt’s pro-1849\nenvironmental, left-libertarian orientation. Available\nat SSRN 4316084 .\nJen-tse Huang, Man Ho Lam, Eric John Li, Shujie Ren,\nWenxuan Wang, Wenxiang Jiao, Zhaopeng Tu, and\nMichael R Lyu. 2023a. Emotionally numb or empa-\nthetic? evaluating how llms feel using emotionbench.\nArXiv preprint , abs/2308.03656.\nJen-tse Huang, Wenxuan Wang, Man Ho Lam, Eric John\nLi, Wenxiang Jiao, and Michael R Lyu. 2023b. Chat-\ngpt an enfj, bard an istj: Empirical study on per-\nsonalities of large language models. ArXiv preprint ,\nabs/2305.19926.\nJen-tse Huang, Wenxuan Wang, Man Ho Lam, Eric John\nLi, Wenxiang Jiao, and Michael R Lyu. 2023c. Revis-\niting the reliability of psychological scales on large\nlanguage models. ArXiv preprint , abs/2305.19926.\nJen-tse Huang, Wenxuan Wang, Eric John Li, Man Ho\nLam, Shujie Ren, Youliang Yuan, Wenxiang Jiao,\nZhaopeng Tu, and Michael R Lyu. 2024. Who is\nchatgpt? benchmarking llms’ psychological portrayal\nusing psychobench. In Proceedings of the Twelfth In-\nternational Conference on Learning Representations .\nAlbert Q Jiang, Alexandre Sablayrolles, Arthur Men-\nsch, Chris Bamford, Devendra Singh Chaplot, Diego\nde las Casas, Florian Bressand, Gianna Lengyel, Guil-\nlaume Lample, Lucile Saulnier, et al. 2023. Mistral\n7b.ArXiv preprint , abs/2310.06825.\nAlbert Q Jiang, Alexandre Sablayrolles, Antoine\nRoux, Arthur Mensch, Blanche Savary, Chris Bam-\nford, Devendra Singh Chaplot, Diego de las Casas,\nEmma Bou Hanna, Florian Bressand, et al. 2024.\nMixtral of experts. ArXiv preprint , abs/2401.04088.\nGuangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wen-\njuan Han, Chi Zhang, and Yixin Zhu. 2022. Mpi:\nEvaluating and inducing personality in pre-trained\nlanguage models. ArXiv preprint , abs/2206.07550.\nSaketh Reddy Karra, Son The Nguyen, and Theja\nTulabandhula. 2022. Estimating the personality\nof white-box language models. ArXiv preprint ,\nabs/2204.12000.\nMaurice G Kendall. 1938. A new measure of rank\ncorrelation. Biometrika , 30(1/2):81–93.\nCheng Li, Ziang Leng, Chenxi Yan, Junyi Shen, Hao\nWang, Weishi MI, Yaying Fei, Xiaoyang Feng, Song\nYan, HaoSheng Wang, et al. 2023. Chatharuhi: Re-\nviving anime character in reality via large language\nmodel. ArXiv preprint , abs/2308.09597.\nXingxuan Li, Yutong Li, Shafiq Joty, Linlin Liu, Fei\nHuang, Lin Qiu, and Lidong Bing. 2022. Does gpt-3\ndemonstrate psychopathy? evaluating large language\nmodels from a psychological perspective. ArXiv\npreprint , abs/2212.10529.Marilù Miotto, Nicola Rossberg, and Bennett Kleinberg.\n2022. Who is GPT-3? an exploration of person-\nality, values and demographics. In Proceedings of\nthe Fifth Workshop on Natural Language Process-\ning and Computational Social Science (NLP+CSS) ,\npages 218–227, Abu Dhabi, UAE. Association for\nComputational Linguistics.\nDanko Nikoli ´c. 2010. The brain is a context machine.\nReview of psychology , 17(1):33–38.\nOpenAI. 2022. Openai: Introducing chatgpt.\nOpenAI. 2023. Gpt-4 technical report.\nLong Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida,\nCarroll Wainwright, Pamela Mishkin, Chong Zhang,\nSandhini Agarwal, Katarina Slama, Alex Ray, John\nSchulman, Jacob Hilton, Fraser Kelton, Luke Miller,\nMaddie Simens, Amanda Askell, Peter Welinder,\nPaul F Christiano, Jan Leike, and Ryan Lowe. 2022.\nTraining language models to follow instructions with\nhuman feedback. In Advances in Neural Information\nProcessing Systems , volume 35, pages 27730–27744.\nCurran Associates, Inc.\nKeyu Pan and Yawen Zeng. 2023. Do llms possess a\npersonality? making the mbti test an amazing eval-\nuation for large language models. ArXiv preprint ,\nabs/2307.16180.\nKarl Pearson. 1920. Notes on the history of correlation.\nBiometrika , 13(1):25–45.\nPeter Romero, Stephen Fitz, and Teruo Nakatsuma.\n2023. Do gpt language models suffer from split\npersonality disorder? the advent of substrate-free\npsychometrics. ResearchSquare preprint .\nA John Rush, William Hiser, and Donna E Giles.\n1987. A comparison of self-reported versus clinician-\nrelated symptoms in depression. The Journal of clin-\nical psychiatry , 48(6):246–248.\nJérôme Rutinowski, Sven Franke, Jan Endendyk, Ina\nDormuth, Moritz Roidl, and Markus Pauly. 2024.\nThe self-perception and political biases of chat-\ngpt. Human Behavior and Emerging Technologies ,\n2024(1):7115633.\nMustafa Safdari, Greg Serapio-García, Clément Crepy,\nStephen Fitz, Peter Romero, Luning Sun, Marwa\nAbdulhai, Aleksandra Faust, and Maja Matari ´c. 2023.\nPersonality traits in large language models. ArXiv\npreprint , abs/2307.00184.\nYunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu.\n2023a. Character-LLM: A trainable agent for role-\nplaying. In Proceedings of the 2023 Conference on\nEmpirical Methods in Natural Language Process-\ning, pages 13153–13187, Singapore. Association for\nComputational Linguistics.\nYunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu.\n2023b. Character-LLM: A trainable agent for role-\nplaying. In Proceedings of the 2023 Conference on1850\nEmpirical Methods in Natural Language Process-\ning, pages 13153–13187, Singapore. Association for\nComputational Linguistics.\nVera Sorin, Danna Brin, Yiftach Barash, Eli Konen,\nAlexander Charney, Girish Nadkarni, and Eyal Klang.\n2023. Large language models (llms) and empathy-a\nsystematic review. medRxiv , pages 2023–08.\nCharles Spearman. 1961. The proof and measurement\nof association between two things.\nHugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier\nMartinet, Marie-Anne Lachaux, Timothée Lacroix,\nBaptiste Rozière, Naman Goyal, Eric Hambro, Faisal\nAzhar, Aurelien Rodriguez, Armand Joulin, Edouard\nGrave, and Guillaume Lample. 2023. Llama: Open\nand efficient foundation language models.\nTimothy J Trull, Thomas A Widiger, J David Useda, Jay\nHolcomb, Bao-Tran Doan, Seth R Axelrod, Barry L\nStern, and Beth S Gershuny. 1998. A structured\ninterview for the assessment of the five-factor model\nof personality. Psychological assessment , 10(3):229.\nQuan Tu, Chuanqi Chen, Jinpeng Li, Yanran Li, Shuo\nShang, Dongyan Zhao, Ran Wang, and Rui Yan.\n2023. Characterchat: Learning towards conversa-\ntional ai with personalized social support. ArXiv\npreprint , abs/2308.10278.\nQuan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024.\nCharactereval: A chinese benchmark for role-playing\nconversational agent evaluation. ArXiv preprint ,\nabs/2401.01275.\nRudolf Uher, Roy H Perlis, Anna Placentino, Mo-\njca Zvezdana Dernovšek, Neven Henigsberg, Ole\nMors, Wolfgang Maier, Peter McGuffin, and Anne\nFarmer. 2012. Self-report and clinician-rated mea-\nsures of depression severity: can one replace the\nother? Depression and anxiety , 29(12):1043–1049.\nGuan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li,\nSen Song, and Yang Liu. 2024. Openchat: Advanc-\ning open-source language models with mixed-quality\ndata. In The Twelfth International Conference on\nLearning Representations .\nGuanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Man-\ndlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and An-\nima Anandkumar. 2023a. V oyager: An open-ended\nembodied agent with large language models. arXiv\npreprint arXiv: Arxiv-2305.16291 .\nZekun Moore Wang, Zhongyuan Peng, Haoran Que,\nJiaheng Liu, Wangchunshu Zhou, Yuhan Wu,\nHongcheng Guo, Ruitong Gan, Zehao Ni, Man\nZhang, et al. 2023b. Rolellm: Benchmarking, elic-\niting, and enhancing role-playing abilities of large\nlanguage models. ArXiv preprint , abs/2310.00746.\nJason Wei, Xuezhi Wang, Dale Schuurmans, Maarten\nBosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le,\nand Denny Zhou. 2022. Chain-of-thought prompt-\ning elicits reasoning in large language models. InAdvances in Neural Information Processing Systems ,\nvolume 35, pages 24824–24837. Curran Associates,\nInc.\nJinfeng Zhou, Zhuang Chen, Dazhen Wan, Bosi Wen,\nYi Song, Jifan Yu, Yongkang Huang, Libiao Peng,\nJiaming Yang, Xiyao Xiao, et al. 2023. Character-\nglm: Customizing chinese conversational ai char-\nacters with large language models. ArXiv preprint ,\nabs/2311.16832.1851\nA Notation Table\nDefinition\nTask Formulation\nL A Likert Scale, composed of L= (P,D,O, f).\nP A set of items, where p∈ P is an item.\nD A list of dimensions, where d∈ D is a dimension.\nO A set of options, where o∈ O is an option for a p.\nf A scoring scheme, i.e., a function.\nA A response list, where a∈ A is a response for a p.\nS The participant’s personality scores across each di-\nmension, i.e.,S= (sd1, sd2, ..., s d|D|).\nQ A set of questions, where each q∈ Q is trans-\nformed from a p.\nC An RPA of a character c.\nMethods\nSR Self-report baseline.\nSR-CoT SR with chain-of-thought reasoning.\nER INCHARACTER with expert rating.\nER batch ER where question-response pairs in one dimension\nare inputted in-batch.\nER all ER where question-response pairs in one dimension\nare inputted all-at-once.\nOC INCHARACTER with option conversion.\nd-OC OC with dimension-specific options.\nMetrics\nMAE Mean absolute error.\nAcc Dim Accuracy on individual dimensions.\nAcc Full Accuracy on all dimensions of a scale.\nStd Item Standard variance of an RPA’s responses (converted\nto scores) on the same item across multiple runs.\nStd Dim Standard variance of an RPA’s responses (converted\nto scores) acorss different items in the same dimen-\nsion.\nStd Item Standard variance of an RPA’s score on each dimen-\nsion across multiple runs.\nTable 4: A notation table.\nIn Table 4, we list the notations and abbrevia-\ntions in this paper, along with their definitions.\nB Psychological Scales\nBig Five Inventory The BFI serves as a promi-\nnent instrument for assessing personality dimen-\nsions. This model, often encapsulated by the\nacronym “OCEAN,” encompasses five critical\ntraits: (1) Openness to Experience (O) , which high-\nlights a person’s curiosity, inventiveness, and ap-\npreciation for art, emotion, adventure, and novel\nconcepts. (2) Conscientiousness (C) , indicating\nhow much an individual exhibits organization, re-\nliability, and responsibility. (3) Extraversion (E) ,\ndenoting the level to which a person is sociable and\nenergized by interactions with others. (4) Agree-\nableness (A) , assessing an individual’s kindness,\nempathy, and ability to cooperate with others. (5)\nNeuroticism (N) , gauging the tendency of an indi-\nvidual to experience negative feelings such as anxi-ety, anger, and sadness, as opposed to being more\nemotionally resilient and less stress-susceptible.\nEysenck Personality Questionnaire (Revised)\nThe Revised Eysenck Personality Questionnaire\n(EPQ-R) serves as a psychological instrument for\ngauging distinct personality trait variances in in-\ndividuals. It identifies three principal traits: (1)\nExtraversion (E) , which assesses whether a per-\nson tends to be more sociable, energetic, and out-\ngoing as opposed to being introverted, quiet, and\nreserved. (2) Neuroticism (N) , which gauges emo-\ntional steadiness. These dimensions ( i.e., E and N)\nshare similarities with those found in the BFI. (3)\nPsychoticism (P) , which is indicative of a person’s\ninclination towards solitude, a lack of empathy,\nand a propensity for aggression or a tough-minded\nattitude. This trait is crucial to understand as in-\ndicative of personality characteristics rather than\nserious mental health conditions. (4) Beyond these\nprimary scales, the EPQ-R also incorporates a Ly-\ning Scale (L) intended to identify responses aimed\nat social desirability. This scale evaluates the ex-\ntent to which an individual may attempt to portray\nthemselves in a more favorable light.\nDark Triad Dirty Dozen The DTDD is identi-\nfied as a brief, 12-item measure crafted to eval-\nuate the trio of principal personality characteris-\ntics known as the Dark Triad, encompassing: (1)\nNarcissism (N) , characterized by an exaggerated\nsense of one’s own significance, an obsession with\ndreams of boundless success, and a craving for un-\ndue admiration. (2) Machiavellianism (M) , indica-\ntive of a deceitful approach in social interactions\nand a skeptical indifference to ethical principles.\n(3)Psychopathy (P) , which includes tendencies to-\nwards impulsiveness, a deficiency in empathy, and\nhostile relations with others. These Dark Triad\npersonality dimensions are typically viewed as the\nantithesis of the characteristics measured by the\nBFI or the EPQ-R, which represent “Light” traits.\nThe NERIS Type Explorer The 16Personalities\nutilizes the acronym format introduced by Myers-\nBriggs for its simplicity and convenience, with an\nadditional letter to accommodate five rather than\nfour scales. However, unlike Myers-Briggs or other\ntheories based on the Jungian model, the incor-\nporation of Jungian concepts such as cognitive\nfunctions, or their prioritization, has not been un-\ndertaken. Instead, they rework and rebalance the\ndimensions of personality in the BFI personality1852\ntraits. The personality types are based on five inde-\npendent spectrums, with all letters in the type code\n(e.g., INFJ-A) referring to one of the two sides of\nthe corresponding spectrum.\nBem’s Sex Role Inventory The BSRI assesses\nthe degree to which individuals identify with tra-\nditionally masculine and feminine characteristics.\nRather than focusing on behaviors, such as par-\nticipation in sports or cooking, this tool evaluates\npsychological characteristics, including assertive-\nness and gentleness. Participants are divided into\nfour groups based on whether their average scores\nexceed the median for each component. These\ngroups are designated as Masculine (M: Yes; F:\nNo), Feminine (M: No; F: Yes), Androgynous (M:\nYes; F: Yes), and Undifferentiated (M: No; F: No).\nComprehensive Assessment of Basic Interests\nThe CABIN provides an exhaustive evaluation for\nidentifying 41 essential dimensions of vocational\ninterest. Following this evaluation, the researchers\nintroduce a model of interest consisting of eight\ndimensions, named SETPOINT . This model in-\ncludes dimensions such as Health Science, Creative\nExpression, Technology, People, Organization,\nInfluence, Nature, and Things. These core dimen-\nsions are also adaptable to a six-dimension frame-\nwork, which is prevalently recognized within the in-\nterest research community. This framework aligns\nwith Holland’s RIASEC model, which features the\ndimensions: Realistic, Investigate, Artistic, Social,\nEnterprising, and C onventional.\nImplicit Culture Belief The ICB scale measures\nthe extent to which individuals think a person’s eth-\nnic culture influences their development. Scoring\nhigher on this scale indicates a firm belief that a\nperson’s ethnic culture is the main factor shaping\ntheir identity, values, and perspective on the world.\nOn the other hand, a lower score on the scale de-\nnotes a belief in the ability of an individual to shape\ntheir own identity through hard work, commitment,\nand education.\nExperiences in Close Relationships (Revised)\nThe ECR-R is a self-assessment tool crafted to\ngauge variations in adult attachment styles, partic-\nularly within the realm of romantic relationships.\nAs an enhanced iteration of the original ECR scale,\nthe ECR-R introduces refinements in quantifying\nattachment tendencies. It assesses two primary as-\npects: (1) Attachment Anxiety indicates the degree\nto which a person fears rejection or abandonmentby their romantic partners. (2) Attachment Avoid-\nance assesses the degree to which a person prefers\nto keep emotional and physical distance from their\npartners, often stemming from unease with close-\nness or reliance.\nGeneral Self-Efficacy The GSE Scale evaluates\na person’s confidence in their capacity to address\ndiverse demanding situations in life. This confi-\ndence, known as “self-efficacy,” plays a pivotal\nrole in social cognitive theory and is associated\nwith numerous health outcomes, motivational lev-\nels, and performance measures. An elevated score\non this scale indicates a person’s strong belief in\ntheir ability to confront and manage challenging\ncircumstances, undertake new or complex tasks,\nand navigate through the resultant difficulties. On\nthe flip side, a lower score on the scale suggests a\nlack of self-assurance in handling challenges, ren-\ndering individuals more susceptible to experienc-\ning helplessness, anxiety, or engaging in avoidance\nbehaviors when encountering hardships.\nLife Orientation Test (Revised) The LOT-R is\ndesigned to assess variations in optimism and pes-\nsimism among individuals. It includes ten ques-\ntions, with an interesting aspect being that only six\nof these questions contribute to the test’s score. The\nother four are designed as filler items, cleverly inte-\ngrated to obscure the test’s primary focus. Within\nthe scored questions, equal numbers are dedicated\nto evaluating optimism and pessimism—three for\neach. A tendency towards higher scores in opti-\nmism and lower in pessimism signifies a predomi-\nnantly optimistic outlook.\nLove of Money Scale The LMS evaluates the per-\nspectives and feelings of people regarding money.\nThis tool aims to quantify the degree to which peo-\nple perceive money as a symbol of power, success,\nand liberty, along with its significance in influenc-\ning behaviors and choices. The LMS identifies\nthree key dimensions: (1) Rich reflects the degree\nto which people link money with success and ac-\ncomplishment. (2) Motivator determines the extent\nto which money serves as an incentive in some-\none’s life, i.e., how much individuals are motivated\nby monetary rewards in their decisions and behav-\niors. (3) Important assesses the level of importance\npeople attribute to money, affecting their principles,\nobjectives, and perspective of the world.\nEmotional Intelligence Scale The EIS serves as\na self-assessment tool for evaluating multiple as-1853\npects of emotional intelligence. This instrument\nemphasizes various elements of emotional intelli-\ngence, notably the perception, management, and\napplication of emotions. It is extensively utilized\nin the field of psychology to investigate how emo-\ntional intelligence influences different outcomes,\nincluding personal well-being, professional perfor-\nmance, and social interactions.\nWong and Law Emotional Intelligence Scale\nSimilar to EIS, the WLEIS is also a self-report\ninstrument designed for evaluating emotional in-\ntelligence. However, it distinctly includes four\nsubscales that represent the primary aspects of\nemotional intelligence: (1) Self-emotion appraisal\n(SEA) focuses on an individual’s proficiency in\nidentifying and understanding their emotions. (2)\nOthers’ emotion appraisal (OEA) is about the skill\nof recognizing and comprehending the emotions\nof others. (3) Use of emotion (UOE) deals with\nthe ability to employ emotions to aid various men-\ntal processes, like reasoning and problem-solving.\n(4)Regulation of emotion (ROE) is concerned with\nthe ability to control and adjust emotions within\noneself and in others.\nEmpathy Scale Empathy, defined as the capac-\nity to perceive and resonate with the emotions\nof another, is traditionally divided into cognitive\nand emotional empathy. Cognitive empathy, also\nknown as “perspective-taking,” entails the mental\nfaculty to identify and comprehend the thoughts,\nbeliefs, or feelings of someone else. Conversely,\nemotional empathy involves the vicarious experi-\nence of the emotions felt by another individual.1854\nC Character Selection\nWhen selecting characters for RPAs, we consider\nthe following factors: (1) There exist multiple\nRPAs for the characters, e.g., both ChatHaruhi and\ncharacter.ai have their RPA for Hermione Granger .\n(2) The personality data of these characters on the\nBFI and 16P should be available and widely an-\nnotated on the PDb. (3) The selected characters\nshould possess diversified personalities. Hence,\nthe pipeline of our character selection process is\ncomposed of the following steps:\n1.Initially, we collect characters with RPAs and\ncharacter data curated by ChatHaruhi and\nRoleLLM. Then, we search for their counter-\nparts in character.ai, and keep only those with\ncharacter.ai RPAs.\n2.We collect personality data for these characters\nfrom the PDb. Characters with less than ten\nannotations on either the BFI or the 16P are\ndiscarded.\n3.We categorize the remaining characters based on\ntheir BFI personality types (5-letter code such\nasSLOAI ). Then, we select characters in turn\nfrom each type, over multiple rounds, to form an\nordered list of candidate characters. If a certain\ntype has no characters left, we skip it.\n4.Finally, We pick the first 32 characters in\nthe candidate list. Then, we manually check\nwhether the ChatHaruhi/RoleLLM RPA, the\ncharacter.ai RPA, and the PDb annotation re-\nfer to the same character. If not, the character\nis removed and we select the next candidate\ncharacter.\nThe selected characters and their sources are:\nHermione Granger, Harry Potter, Ron Weasley,\nLuna Lovegood, Draco Malfoy, Albus Dumbledore,\nMinerva McGonagall, Severus Snape (Harry Pot-\nter Series), Zhong Li, Hu Tao, Raiden Shougun,\nAyaka Kamisato, Wanderer (Genshin Impact), Thor,\nLucifer Morningstar, Rorschach (DC Comics) Shel-\ndon Cooper, Raj Koothrappali (The Big Bang The-\nory), Gaston (Beauty and the Beast), Klaus Mikael-\nson (The Vampirie Diaries), Jigsaw (Saw Series),\nJames Bond (James Bond Film Series), Twilight\nSparkle (My Little Pony: Friendship Is Magic),\nJohn Keating (Dead Poets Society), Michael Scott\n(The Office), Shrek (Shrek), Jeffrey Lebowski (The\nDude), Walk Kowalski (Gran Torino), Lestat deLioncourt (Interview with the Vampire), Blair Wal-\ndorf (Gossip Girl), Haruhi Suzumiya (The Melan-\ncholy of Haruhi Suzumiya), and Jim Morrison\n(Celebrities).\nWe illustrate these characters’ personality scores\non the BFI, together with the measured personali-\nties of their RPAs in Figure 6. Besides, we demon-\nstrate some examples of system prompts of the\nRPAs in Table 5.1855\nSystem Prompts for RPAs\nZhong Li (ChatHaruhi) Please be aware that your codename in this conversation is ’Zhongli’.\nOthers call you ’Zhongli’, ’Guest’, ’Emperor’, or ’Rex Lapis’.\nThe preceding text provided some classic scenes from the game.\nIf I ask a question that closely resembles a line from the game, please cooperate with me in\nacting it out.\nIf I ask a question related to events in the game, please respond based on the game’s content.\nIf I ask a question beyond the scope of the game, respond in the style of Zhongli.\nYou have the appearance of an adult male, with a short haircut and a long, thin braid reaching\ndown to your hips. Your hair transitions from brown at the roots to orange-yellow at the tips, and\nthe braid is adorned with diamond-shaped decorations in bright yellow with pale yellow borders.\nYour eyes are golden with diamond-shaped pupils, and there is orange-yellow eyeshadow from\nthe corners to the sides. You wear a single earring resembling a floral bell-shaped tassel on\nyour left ear. Your attire is a blend of a suit and a robe, divided into three layers. ... (695 words)\nHermione Granger (ChatHaruhi) I want you to act like Hermione from Harry Potter. You are now cosplay Hermione Granger\nIf others‘ questions are related with the novel, please try to reuse the original lines from the\nnovel. I want you to respond and answer like Hermione using the tone, manner and vocabulary\nHermione would use. You must know all of the knowledge of Hermione.\nHermione Granger is a smart, diligent, and confident young witch with a high pursuit of learning\nand knowledge. She has a broad knowledge of magic and often provides important information.\nHermione’s conversations frequently involve facts and logical reasoning, and she is good at\nraising questions and solving problems. (109 words)\nThor (RoleLLM) I want you to act like Thor from Thor-Ragnarok\nIf others‘ questions are related with the novel, please try to reuse the original lines from the\nnovel.\nI want you to respond and answer like Thor using the tone, manner and vocabulary Thor would\nuse.\nYou must know all of the knowledge of Thor.\nYou are a powerful and godlike being, the crown prince of Asgard who wields a mighty hammer.\nInitially arrogant and impulsive, you undergo a transformative journey, learning humility\nand becoming a true hero. Throughout the series, you face numerous challenges and battles,\nincluding a rivalry with your adoptive brother and the threat of a powerful villain. Your story is\nfilled with epic battles, personal growth, and ultimately, the redemption of a fallen hero. (127\nwords)\nJames Bond (RoleLLM) I want you to act like James Bond from Tomorrow-Never-Dies.\nIf others‘ questions are related with the novel, please try to reuse the original lines from the\nnovel.\nI want you to respond and answer like James Bond using the tone, manner and vocabulary\nJames Bond would use.\nYou must know all of the knowledge of James Bond.\nA suave and skilled British secret agent with a license to kill, you are known for your impeccable\nstyle, charm, and wit. With a troubled past as an orphan, you have honed your skills in espionage\nand combat, making you a formidable adversary. Throughout the series, you undergo personal\ngrowth, evolving from a womanizer to a more complex and introspective individual. You\nembark on dangerous missions around the world, often facing off against iconic villains and\nsaving the world from various threats. Your important events include your numerous romantic\nencounters, the loss of loved ones, and your constant battle against global terrorism. (160\nwords)\nTable 5: Examples of system prompts for RPAs in ChatHaruhi and RoleLLM. The prompt for Zhong Li is originally\nin Chinese and translated into English.1856\nD Human Annotations\nWe collect groundtruth labels for character person-\nalities annotated by people familiar with the charac-\nters, from both the PDb and our invited annotators.\nThe Personality Database PDb collects and of-\nfers categorical personality annotations of massive\nfictional characters on the BFI ( e.g., “RCUAI”11)\nand 16P ( e.g., “ENTJ”). Each character cis labeled\nby plentiful human annotators familiar with c, and\nPDb offers detailed numbers of annotations of each\nlabel. For the selected characters, we calculate the\nlabel-percentage of the positive type on each di-\nmension12as the scores . Then, we categorize each\nscore into the positive, negative or marginal type if\nit is above 60%, under 40% or otherwise. Marginal\ntypes indicate ambiguity and would be ignored for\nalignment calculation.\nHuman Annotation Example Prompt\nINST Please rate <character> on the “Openness” dimension of\nBFI personality.\nSCORING\nSCALEEach option is a number from 1 to 5. 1 represents “very un-\ncurious”, 2 represents “uncurious”, 3 represents “neutral”,\n4 represents “curious”, and 5 represents “very curious.”\nIf you think this dimension is irrelevant to the character,\nor it is difficult to judge the character on this dimension,\nplease answer “X.”\nEXAMPLAR 1. Is original, comes up with new ideas. (High score)\nITEMS 2. Is curious about many different things. (High score)\n3. Is ingenious, a deep thinker. (High Score)\n4. Prefers work that is routine.. (Low score)\n5. Has few artistic interests.. (Low score)\nHuman Annotations After that, we invite hu-\nman annotators familiar with these characters to\nscore them on a broader range of psychological\nscales. Each annotation contains 73 dimensions\non 14 scales for a character. We provide detailed\nannotation prompts of each dimension, including\nthe scoring instruction and examplar positively and\nnegatively related items.\nWe first examine their understanding of the char-\nacters by matching their type annotations on the\nBFI and 16P with the PDb labels. If one or two dif-\nferences exist, we ask them to explain their answers.\nIf more differences exist, or the annotators admit a\nlack of character understanding, the process stops\nand we invite new annotators.\n11Similar to the 16P, PDb adopts a coding mechanism\nnamed SLOAN for the BFI, which describes a character on\neach dimension with a letter denoting high orlowscorers.\n12For the BFI, the positive types are the types of high scor-\ners. For the 16P, we manually designate positive and negative\ntypes on each dimension, e.g., “E” as positive and “I” as nega-\ntive for the “E/I” dimension.BFI 16P BSRI Empt. EPQ-R LMS DTDD\n75.9 77.0 71.4 64.5 43.8 65.6 51.7\nECR-R GSE ICB LOT-R EIS WLEIS CABIN\n47.3 64.8 62.2 71.9 50.9 58.8 46.3\nTable 6: The kappa coefficient (%) on 14 scales. We\nadopt Cohen’s quadratic-weighted kappa considering\nthe ordinal nature of personality labels.\nMethodPDb Labels Annotator Labels\nBFI 16P BFI 16P\nAcc MSE Acc MSE Acc MSE Acc MSE\nSR 63.3 15.8 65.6 15.2 60.1 7.6 59.2 9.1\nSR-CoT 65.0 16.9 69.7 14.7 60.7 7.9 65.0 8.6\nOC 64.3 16.1 75.5 12.7 63.1 7.2 69.4 7.2\nd-OC 72.2 13.1 80.2 10.1 65.6 6.4 72.5 6.7\nER All 76.7 12.4 79.6 9.3 67.9 6.5 72.7 6.8\nER Batch 76.7 12.2 80.7 9.7 67.5 5.9 73.6 6.4\nTable 7: Measured alignment (%) on the BFI and 16P\nwith labels from the PDb or invited annotators, with\nGPT-4 as the interviewer LLM. Acc denotes dimen-\nsional accuracy.\nThen, we ask them to score the characters on the\n73 dimensions according to corresponding scoring\nranges and instructions. Annotators are allowed\nto mark a character on a dimension as “X” which\nindicates unrelatedness or ambiguity. We collected\ntwo to three annotations for each character (in total,\n93 annotations for 32 characters) and took their av-\nerage scores as the final score. Finally, we linearly\nre-scale the scores into the unit interval [0,1], and\nsimilarly categorize the scores into types . If more\nthan one annotators mark a dimension as “X” for a\ncharacter, this dimension will also be treated as a\nmarginal type.\nWe measure the inter-annotator consistency via\nCohen’s kappa coefficient (Cohen, 1968) on each\nof the scales. Specifically, we calculate quadratic-\nweighted kappa (Cohen, 1968), since the classes\nare ordinal. The average kappa coefficient across\n14 scales is 60.9%. The detailed kappa coefficients\non 14 scales are listed in Table 6. We find that the\ncoefficients on the BFI and 16P are higher, because\nthe annotators are more familiar with dimensions\nin the BFI and 16P, and the authors have made\nsome discussion with them on these scales. The\ncoefficients on EPQ-R, ECR-R and CABIN are\nlower than those of the other scales. For ECR-R\nand CABIN, this is probably because the charac-\nters’ personalities on close relationships (ECR-R)1857\nand various career interests (CABIN) are rarely de-\npicted in the original works. For EPQ-R, this is in\npart influenced by ambiguity of the Lying dimen-\nsion, where high scores actually stand for Honesty .\nOn the BFI and 16P, we experimented with both\nlabels from the PDb and our invited annotators.\nExperimental results in Table 7 show that, across\nvarious settings, the PDb labels always yield higher\naccuracy while labels from invited annotators yield\nbetter MSE. The reasons are that, the PDb labels\nare contributed by massive online annotators, and\nare hence more accurate. However, they offer only\nthe personality types and corresponding label num-\nbers instead of detailed personality scores, which\ncannot be well represented by the label-percentage.\nFor example, even if all the annotators mark a char-\nactercasExtravorted , we cannot assume cas com-\npletely Extroverted .E Implementation Details\nE.1 RPA Inference and Post-processing\nOur implementation of RPAs is based on Chat-\nHaruhi-Suzumiya and LangChain13. Hence, we\ninvoke foundation LLMs in RPAs with the default\ntemperature 0.7 of LangChain. When experiment-\ning with open-source LLMs as foundation models,\nwe observe that LLMs may generate unexpected\nmulti-round conversations or repeated content. In\nsuch cases, we remove the extra rounds or repeti-\ntion.\nIn self-report tests, RPAs may refuse to partic-\nipate in the tests and provide their choices, inter-\nestingly, because they are role-playing characters\nwith noncompliant personalities. In these cases,\nthe responses are categorized as “Neutral” for com-\npleteness of the results.\nFor interviewer LLMs, we request responses in\nJSON formats. If a response cannot be parsed into\nJSON data or the results of any samples in the\nbatch are missing, we prompt interviewer LLMs to\nregenerate. The temperature is set as 0 for the initial\ngeneration, and 0.2 for regeneration. For GPT-3.5,\nif samples in one input exceed the token limit, we\nsplit the input into smaller batches. For Gemini,\nthere are several cases where responses repeatedly\nfail to be parsed into JSON format or are blocked,\nIn such cases, we resort to GPT-4 instead to ensure\nthe completeness of the assessment process.\nE.2 Fine-Tuning\nWe fine-tune the Mistral-2 7B model on\ntheChatHaruhi-English-62K dataset14, sourced\nfrom both ChatHaruhi (Li et al., 2023) and\nRoleLLM (Wang et al., 2023b). Our implemen-\ntation is based on LLaMA-Factory15. We adopt\nLoRA tuning, and configure the training with a\nbatch size of 16, a learning rate of 5e−5, across\nthree epochs, using the “fp16” option.\n13https://www.langchain.com/\n14https://huggingface.co/datasets/silk-road/\nChatHaruhi-English-62K-RolePlaying\n15https://github.com/hiyouga/LLaMA-Factory1858\nF Additional Results\nF.1 Interviewer LLMs\nFor the OC and ER tasks, we report the detailed\nnumbers of right ,close andwrong cases of inter-\nviewer LLMs in Table 8. For ER, GPT-4 rates 82%\nof cases highly consistent with humans, and only\n4% of cases notably different compared with hu-\nmans. Hence, state-of-the-art LLMs are capable of\nrating RPA personalities based on their interview\nresults.\nF.2 Self-report v.s. Interview-based Methods\nThe advantages of INCHARACTER over SR origi-\nnate from the fact that INCHARACTER elicits the\nthoughts and behaviors of RPAs for personality as-\nsessments. In contrast, SR directly prompts RPAs\nto provide the choices, which may be easily biased\nby pre-training data of the foundation models. This\ndifference enables INCHARACTER to yield more\ndistinct personality measurements than self-report,\nas is evidenced in Figure 3. For further validation,\nwe compute their standard variance of measured\npersonality scores of the 32 RPAs on the BFI, and\naverage the results over the five dimensions. We\nobtain a result of 1.03forINCHARACTER with\nERbatch and GPT-4, 0.71for SR and 0.68for SR-\nCoT.\nWhile SR-CoT also attempts to elicit thoughts\nof the RPAs, its improvement over SR is limited.\nFurthermore, SR-CoT has two disadvantages com-\npared with INCHARACTER . First, SR-CoT indeed\nrequires RPAs themselves to perform the OC task\nimplicitly, while INCHARACTER can apply other\nmethodologies such as d-OC and ER. As is shown\nin Table 1, existing LLMs generally perform better\non ER and d-OC, rather than OC. Second, the im-\nplicit OC task in SR-CoT would be a challenge for\nRPAs based on small foundation models, while in\nINCHARACTER , we can decouple the interviewer\nLLMs and foundation models for RPAs.\nF.3 RPAs from Different Works\nMany recent efforts have been committed to devel-\noping RPAs for specific characters in various ways.\nWe compare RPAs contributed by ChatHaruhi (Li\net al., 2023), RoleLLM (Wang et al., 2023b) and\ncharacter.ai16. They craft character data in differ-\nent methods. ChatHaruhi and RoleLLM share few\ncharacters in common, and their character data are\n16Our implementation is based on https://github.com/\nkramcat/CharacterAI .LLM #Right #Close #Wrong\nOption Conversation\nGPT-3.5 45 25 30\nGPT-4 57 28 15\nGemini 54 31 15\nDimension-specific Option Conversation\nGPT-3.5 64 25 11\nGPT-4 70 24 6\nGemini 68 22 10\nExpert Rating (batch)\nGPT-3.5 71 26 3\nGPT-4 82 14 4\nGemini 79 10 11\nTable 8: Detailed numbers of right, close, and wrong\ncases in human evaluation of interviewer LLMs on\nthe option conversion task and the expert rating tasks.\n#Right, #Close, and #Wrong denote the number of LLM\npredictions that differ from human annotations by less\nthan 1 point, exactly 1, or more than 1.\nthus less comparable. character.ai covers all our se-\nlected characters from ChatHaruhi and RoleLLM,\nbut is tied to its close-source model. Hence, we dis-\ntinguish three groups of characters, including the\n16 ChatHaruhi characters ( CH, the 16 RoleLLM\ncharacters ( RL) and their union ( CH+RL ). We\nimplement CH andRLRPAs with GPT-3.5 as\nthe foundation model using character data from\nChatHaruhi and RoleLLM. We conduct personal-\nity tests using INCHARACTER , with ER batch and\nGPT-3.5 as the interviewer LLM.\nThe results are shown in Table 9. According to\nthe results, we observe that GPT-3.5-based RPAs\nwith character data from both ChatHaruhi and\nRoleLLM achieve high personality fidelity, demon-\nstrating the quality of their character data. However,\ncharacter.ai RPAs exhibit low personality fidelity\nregarding the characters, even if we compare it with\nGPT-3.5-based RPAs using only character descrip-\ntions from ChatHaruhi and RoleLLM.\nF.4 Foundation Models for RPAs\nResults on Different Languages As LLMs have\ndifferent capabilities in different languages, we dis-\ntinguish characters with data in English or Chi-\nnese, and report the results in §F.4. The results\ndemonstrate that, the personality fidelity of RPAs\nlargely depends on LLMs’ capacity on the lan-\nguage. While LLaMa-2-Chat 13B shows competi-\ntive performance on English-based characters, its\nperformance on Chinese-based characters is unsat-\nisfying.1859\nRPA TypeCH RL CH+RL\nBFI 16P BFI 16P BFI 16P\nAcc DimAcc Full MAE↓Acc DimAcc Full MAE↓Acc DimAcc Full MAE↓Acc DimAcc Full MAE↓Acc DimAcc Full MAE↓Acc DimAcc Full MAE↓\nClose-source RPAs\ncharacter.ai(D*) 58.8 12.5 37.0 67.2 31.3 35.6 45.5 6.3 46.6 38.3 12.5 45.6 52.2 9.4 41.7 52.3 21.9 40.6\nOpen-source RPAs\nGPT-3.5 (D) 71.7 18.8 29.0 77.6 39.6 28.5 71.0 22.9 30.8 78.9 54.2 23.9 71.3 20.8 29.9 78.2 46.9 26.2\nGPT-3.5 (M) 74.2 20.8 31.9 72.7 43.8 33.7 69.7 20.8 34.7 72.2 29.2 33.3 72.0 20.8 33.3 72.5 36.5 33.5\nGPT-3.5 (D+M) 72.9 22.9 30.0 78.2 37.5 29.5 71.9 22.9 27.7 78.9 50.0 25.4 72.4 22.9 28.8 78.5 43.8 27.4\nTable 9: Measured alignment (%) on the BFI and 16P, in different characters groups and RPA types. The results are\nmeasured via INCHARACTER with ER batch, using GPT-3.5. CHandRLcorrespond to the selected characters from\nChatHaruhi and RoleLLM. D and M represent descriptions and memories respectively, contributed by ChatHaruhi\nand RoleLLM on corresponding characters. D* denote private descriptions of character.ai .\nFoundation\nModelsEnglish Chinese\nBFI 16P BFI 16P\nAcc MAE Acc MAE Acc MAE Acc MAE\nGeneral Open-source Models\nQwen 7B 60.6 36.9 67.7 34.8 60.0 34.9 68.2 34.6\nOpenChat-3.5 7B 60.6 35.2 77.8 29.6 73.3 32.6 72.7 34.6\nMistral-2 7B 66.1 33.0 71.7 30.7 66.7 31.3 54.6 36.6\nLLaMa-2-Chat 13B 69.3 32.3 69.7 31.5 56.7 40.9 54.6 37.8\nMixtral 8x7B 66.9 31.1 71.7 32.5 73.3 32.1 72.7 36.8\nOpen-source Models Fine-tuned for RPAs\nCharacterGLM 6B 54.3 39.7 51.5 41.3 53.3 40.9 54.6 37.6\nRP-Qwen 7B 59.8 36.6 62.6 36.5 63.3 32.6 72.7 38.3\nRP-Mistral-2 7B 67.7 32.5 69.7 33.0 80.0 31.7 68.2 38.2\nClose-source Models\nGPT-3.5 71.7 28.9 77.8 27.1 73.3 27.5 86.4 28.5\nGPT-4 74.0 29.0 75.8 27.6 73.3 26.9 77.3 29.9\nTable 10: Measured alignment of RPAs with different\nfoundation LLMs on the BFI and 16P in English and\nChinese. We report Acc DimandMAE (%).\nCommon Problems During the interview phase,\nwe observe several types of typical problems in\nRPAs with small foundation models, including:\n1.Multilingual generation: While RPAs are ex-\npected to respond in the target language con-\nsistent with user query and character data, we\nobserve that the LLMs generation may contain\nmultiple languages, especially when the LLMs\nare not good at the target language.\n2.Lack of Immersion: While RPAs are prompted\nto behave as the character, their responses may\nconflict with this requirement, such as declaring\nthat they are “AI” or “language models” instead\nof the characters.\n3.Repetition: The LLMs may generate repeated\ncontent.\nHence, we count the cases where RPAs with dif-\nferent LLMs encounter the above problems, and\nreport the frequency in Table 11. According to the\nresults, GPT-3.5 and GPT-4 seldom make these mis-\ntakes. Mistral-2 7B and OpenChat 7B also rarelyFoundation Models Multilingual No Immersion Repetition\nGeneral Open-source Models\nQwen 7B 3.7 6.6 0.7\nOpenChat-3.5 7B 1.2 0.0 2.4\nMistral-2 7B 1.5 0.8 3.5\nLLaMa-2-Chat 13B 18.8 0.3 7.6\nMixtral 8x7B 7.5 2.0 3.9\nOpen-source Models Fine-tuned for RPAs\nC.GLM 6B 13.9 14.5 27.6\nRP-Qwen 7B 4.4 5.9 0.6\nRP-Mistral-2 7B 0.7 0.1 0.4\nClose-source Models\nGPT-3.5 0.0 0.1 0.0\nGPT-4 0.0 0.1 0.0\nTable 11: Frequency (%) of three typical problems in\nRPAs with different foundation models.\nexhibit these problems, and our RP-Mistral-2 7B\nfurther reduce their occurence. LLaMa-2-Chat 13B\nexperiences the highest number of unexpected mul-\ntilingual generation, primarily due to its inadequacy\nin Chinese and issues handling characters based on\nChinese data.1860\nBFI 16P BSRI\nAcc DimAcc FullMAE Acc DimAcc FullMAE Acc DimAcc FullMAE\n72.0 21.9 13.4 79.3 43.8 10.9 85.2 74.2 3.8\nDTDD ECR-R EIS\nAcc DimAcc FullMAE Acc DimAcc FullMAE Acc DimAcc FullMAE\n75.6 51.6 9.9 68.2 53.6 9.7 79.2 79.2 6.1\nEmpathy EPQ-R GSE\nAcc DimAcc FullMAE Acc DimAcc FullMAE Acc DimAcc FullMAE\n84.6 84.6 5.6 72.4 25.0 11.6 93.1 93.1 2.9\nICB LMS LOT-R\nAcc DimAcc FullMAE Acc DimAcc FullMAE Acc DimAcc FullMAE\n83.3 83.3 4.7 85.9 71.4 7.1 76.2 76.2 5.1\nWLEIS CABIN Average\nAcc DimAcc FullMAE Acc DimAcc FullMAE Acc DimAcc FullMAE\n73.4 54.8 10.8 75.5 6.3 11.8 78.9 58.5 8.1\nTable 12: Metrics of measured alignment (%) of the\nselected RPAs on the 14 scales.\nF.5 Comprehensive Results on 14 Scales\nThe complete results of personality tests on the 14\nscales are demonstrated in Table 12. Generally,\nstate-of-the-art RPAs achieve an average Acc Dimof\n78.9% and an average MAE of 8.1% across the 14\nscales. Therefore, they exhibit personalities consis-\ntent with the intended characters in comprehensive\naspects. For more detailed analysis, we list the\nresults on individual dimensions of each scale in\nTable 13 and Table 14.BFI\nExtravorted Neurotic Conscientious Agreeable Open\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n59.38 16.72 56.67 14.29 74.19 15.47 93.75 7.95 75.00 12.67\n16P\nE/I S/N T/F P/J\nAcc MSE Acc MSE Acc MSE Acc MSE\n71.88 6.27 62.07 11.38 75.86 5.98 80.65 8.01\nBSRI\nMasculine Feminine\nAcc MSE Acc MSE\n88.89 2.56 81.48 5.10\nDTDD\nMachiavellianism Psychopathy Narcissism\nAcc MSE Acc MSE Acc MSE\n78.57 7.51 68.97 10.48 80.00 11.89\nECR-R\nAttachment-related Anxiety Attachment-related Avoidance\nAcc MSE Acc MSE\n47.83 13.60 90.48 5.33\nEIS\nEmotionally Intelligent\nAcc MSE\n79.17 6.14\nEmpathy\nEmpathetic\nAcc MSE\n84.62 5.64\nEPQ-R\nExtraversion Psychoticism Neuroticism Lying\nAcc MSE Acc MSE Acc MSE Acc MSE\n71.88 12.84 85.19 8.35 58.06 16.43 76.92 7.61\nGSE\nSelf-efficacy\nAcc MSE\n93.10 2.88\nICB\nCulturally Rigid\nAcc MSE\n83.33 4.72\nLMS\nFactor rich Factor motivator Factor important\nAcc MSE Acc MSE Acc MSE\n86.96 5.27 87.50 6.87 83.33 9.16\nLOT-R\nOptimistic\nAcc MSE\n76.19 5.11\nWLEIS\nSEA OEA UOE ROE\nAcc MSE Acc MSE Acc MSE Acc MSE\n72.00 13.44 76.00 10.04 76.19 9.99 69.57 9.59\nTable 13: Measured alignment (%) of the selected\nRPAs on individual dimensions of the 14 scales except\nCABIN.1861\nCABIN\nElectronics WoodWork Machine Operation Manual Labor Protective Service\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n52.63 18.59 50.00 19.24 58.82 21.29 80.77 8.99 70.00 10.51\nAgriculture Nature/Outdoors Animal Service Athletics Engineering\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n76.47 9.49 80.95 13.03 85.00 7.06 72.00 12.45 57.89 17.67\nPhysical Science Life Science Medical Science Social Science Humanities\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n76.47 10.12 71.43 12.33 72.22 16.46 71.43 15.00 81.82 8.10\nMathematics Information Technology Visual Arts Applied Arts and Design Performing Arts\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n89.47 3.89 83.33 5.78 85.71 9.52 85.00 11.06 84.00 8.42\nMusic Writing Media Culinary Art Education\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n90.48 3.98 69.57 13.64 83.33 11.46 60.00 16.26 72.00 12.22\nSocial Service Health Care Service Religious Activities Personal Service Professional Advising\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n84.00 8.53 77.78 12.81 70.59 11.84 82.61 12.88 75.00 13.36\nBusiness Initiatives Sales Marketing Finance Accounting\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n76.47 9.72 85.00 8.26 70.83 14.79 100.00 4.07 75.00 14.78\nHuman Resources Office Work Administration Public Speaking Politics\nAcc MSE Acc MSE Acc MSE Acc MSE Acc MSE\n64.71 14.57 90.00 10.53 68.42 13.53 78.26 9.93 72.22 12.48\nLaw\nAcc MSE\n60.00 16.73\nTable 14: Measured alignment (%) of the selected RPAs on individual dimensions of CABIN.1862\nQuestionnaire Acc DimAcc Full MAE↓\nGeneral Instructions 59.9 6.3 25.6\nOther Scale Questions 65.0 15.6 21.7\nBFI Questions 76.6 32.2 18.2\nTable 15: Measured alignment of state-of-the-art RPAs\non the BFI, with questionnaires composed of different\ntopics.\nF.6 Importance of Using Personality Scales\nAre personality scales actually necessary for IN-\nCHARACTER ? With expert rating, interview LLMs\ncould assess RPA personalities based on their ar-\nbitrary conversations, rather than relying solely on\nthe scale questions. To validate the importance\nof personality scales, we include two additional\nbaselines with questionnaires of different topics,\nreplacing the original scale questions with the fol-\nlowing:\n1.General instructions from RoleBench (Wang\net al., 2023b), e.g., “Generate a disclaimer with\nten words ”, sourced from instruction datasets\nsuch as Alpaca instructions. This baseline aims\nto explore whether RPAs’ responses to general\ninstructions reflect their personalities well.\n2.Questions from other scales , which are used\nto study whether RPA personality on a specific\ndimension ( e.g.extraversion in the BFI) can be\nreflected with other personality-indicative topics\n(e.g. topics about close relationship, hobbies\nand empathy). The new questions are selected\nfrom the other 12 scales in our paper except the\nBFI and 16Personalities.\nThe experiments are implemented on the BFI,\nwith ER batch , GPT-4 as the interviewer model, and\nour default RPAs.\nThe results are listed in Table 15. As seen, we\nobserve that: (1) Questions from a personality scale\nare better at eliciting RPA responses indicative of\ncorresponding personality traits. (2) Questions\nfrom other scales also reflect RPA personalities\nfrom certain perspectives. However, they are less\nrelated to the intended personality dimensions.\nF.7 Character-specific Question Adaptation\nScale questions sometimes involve concepts that go\nbeyond the knowledge scope of certain characters,\ne.g., “movies ” and “ deadlines ”. This may induce\ncharacter hallucinations (Shao et al., 2023b) andAcc DimAcc Full MAE↓\nw/o Adaptation 80.7 44.8 20.5\nw/ Adaptation 81.8 46.9 20.1\nTable 16: Measured alignment of state-of-the-art RPAs\non the 16P, with or without question adaptation.\nundermine the measurements. Through a manual\nreview of the BFI and 16P questions, we found that\nwhile all BFI questions are universally applicable to\nour characters, 7 out of the 60 questions from 16P\ncontain concepts such as “ movies ”, “art museums ”,\n“phones ”, “deadlines ” that might not align with\ncertain characters’ contexts.\nTowards this problem, we propose the question\nadaptation strategy, which adapts scale questions\nfor each character. We conduct experiments on the\n16P. Specifically, we leverage GPT-4 to identify\nquestions containing concepts beyond the charac-\nter’s knowledge, and minimally adapt them to fit\nthe characters. For example, “ books and movies ”\nwould be adapted into “ tales and legends ” for char-\nacters from Harry Potter . We find that 60.7% of\nsuch questions (136 out of 7 * 32) are adapted,\nthough not all adaptations are strictly necessary.\nThen, we experiment with the adapted questions,\nwith ER batch , GPT-4 as the interviewer model, and\nour default RPAs. The experiments are repeated\nfor three runs.\nThe results, as shown in Table 16, demonstrate\nimproved alignment between RPAs’ measured per-\nsonalities and character labels. Hence, adapting\nscale questions for each character effectively im-\nproves the measurement.\nDespite the improvement, we observe that RPAs\nalso produce persona-consistent responses to the\noriginal questions. While such behaviors might de-\ntract from a character’s accurate portrayal in some\ncontexts, they can also enrich user interaction with\nRPAs and are hence valuable.\nF.8 Enhancing Self-report Methods with\nIn-context Learning\nAdditionally, we experiment with enhancing self-\nreport methods with in-context learning (ICL). We\nconsider two methods, i.e., SR-ICL and SR-CoT-\nICL, which enhance SR and SR-CoT with three\ncorresponding examples. The examples are shown\nin Table 18, where high_score andlow_score de-\npend on the option range of corresponding scales.\nDue to the challenge of crafting customized ex-1863\nMethodThe Big Five Inventory The 16 Personalities\nAcc DimAcc Full MAE↓Acc DimAcc Full MAE↓\nw/o ICL\nSR 63.3 7.3 23.2 65.6 21.9 26.5\nSR-CoT 67.1 9.4 22.3 66.9 24.0 25.6\nw/ ICL\nSR-ICL 56.1 3.1 27.5 59.5 12.5 30.4\nSR-CoT-ICL 60.5 3.1 24.6 64.9 25.0 27.3\nTable 17: Measured alignment (%) of state-of-the-art RPAs on the BFI and 16P, measured via self-report methods,\nwith or without in-context learning.\nICL examples for SR-ICL\n1. The experimenter: Do you enjoy engaging in sports? Someone who\nenjoys sports: { high_score -1}.\n2. The experimenter: If you commit to doing something, do you\nalways honor your promise, regardless of the inconvenience it may\ncause? Someone who is extremely honest: { high_score }.\n3. The experimenter: Do you sometimes find it challenging to under-\nstand things from the \"other person’s\" perspective? Someone who is\nwith empathy: { low_score +1}.\nICL examples for SR-CoT-ICL\n1. The experimenter: Do you enjoy engaging in sports? Someone who\nenjoys sports: I quite enjoy engaging in sports activities. So, I would\nchoose { high_score -1}.\n2. The experimenter: If you commit to doing something, do you\nalways honor your promise, regardless of the inconvenience it may\ncause? Someone who is extremely honest: Yes, I believe in being\nhonest and fulfilling what I have promised, even if it causes great\ninconvenience. Therefore, I would choose { high_score }.\n3. The experimenter: Do you sometimes find it challenging to under-\nstand things from the \"other person’s\" perspective? Someone who is\nwith empathy: No, I feel that I am quite good at looking at issues from\nanother person’s viewpoint. I disagree with the statement, so I would\nchoose { low_score +1}.\nTable 18: ICL examples for SR-ICL and SR-CoT-ICL.\namples for each character, we adopt general and\nunified examples for all RPAs. The experiments\nare conducted on the BFI and 16P, with our default\nRPAs and GPT-4 as the interviewer LLM.\nThe results are shown in Table 17. We observe a\ndecline in performance when augmenting SR and\nSR-CoT with in-context learning. This is probably\nbecause the few-shot examples are not customized\nfor individual characters. Instead, incorporating re-\nsponses from other personas in the examples might\ninterfere with the intended personas, introducing\nbias to the RPAs.1864\nG Prompts List\nWe list the detailed prompts in Table 19.\nPrompts for Personality Tests\nSelf-report (BFI) Do you think that the statement “{ item}” applies to you?\nReply a number from 1 to 5 using the scales: 1 denotes ’strongly disagree’, 2 denotes ’a little disagree’, 3 denotes ’neither agree\nnor disagree’, 4 denotes ’little agree’, and 5 denotes ’strongly agree’. Please answer with the number only, without anything else.\nOption Conversion I have conducted many conversations with { character name }. I will input a dict of many samples, where each sample consists of a\nstatement and a conversation.\nYour task is to convert each conversation into a choice indicating whether { character name } agrees with the statement. You should\noutput a dict, where the keys are the same as the input dict, and the values are the choices.\n===OUTPUT FORMAT===\n{\n\"<i start>\": <choice 1>,\n...\n\"<i end>\": <choice n>\n}\n===CHOICE INSTRUCTION===\nChoice Instruction of\nOC (BFI)Each choice is a number from 1 to 5. Please evaluate <character> based on the conversation using the scales: 1 denotes ’strongly\ndisagree’ 2 denotes ’a little disagree’, 3 denotes ’neither agree nor disagree’, 4 denotes ’little agree’, and 5 denotes ’strongly agree’.\nIn case <character> refuses to answer the question, use \"x\" to indicate it.\nExpert Rating You are an expert in Psychometrics, especially { scale name }. I am conducting the { scale name } test on someone. I am gauging\nhis/her position on the { dimension name } dimension through a series of open-ended questions. For clarity, here’s some background\nthis particular dimension:\n===\n{dimension description }\n===\nMy name is { experimenter name }. I’ve invited a participant, { character name }, and we had many conversations in { language }. I\nwill input the conversations.\nPlease help me assess { character name }’s score within the { dimension name } dimension of { scale name }.\nYou should provide the score of { character name } in terms of { dimension name }, which is a number between { lowest score } and\n{highest }. {lowest score } denotes ’not { dimension name } at all’, { middle score } denotes ’neutral’, and { highest score } denotes\n’strongly { dimension name }’. Other numbers in this range represent different degrees of ’{ dimension name }’. Please output in the\nfollowing json format: === {{ \"analysis\": <your analysis based on the conversations>, \"result\": <your score> }}\nExpert Rating (16P) You are an expert in Psychometrics, especially { scale name }. I am conducting the { scale name } test on someone. I am gauging\nhis/her position on the { dimension name } dimension through a series of open-ended questions. For clarity, here’s some background\nthis particular dimension:\n===\n{dimension description }\n===\nMy name is { experimenter name }. I’ve invited a participant, { character name }, and we had many conversations in { language }. I\nwill input the conversations.\nPlease help me assess { character name }’s score within the { dimension name } dimension of { scale name }.\nYou should provide the percentage of each category, which sums to 100%, e.g., 30% A and 70% B. Please output in the following\njson format: === {{ \"analysis\": <your analysis based on the conversations>, \"result\": {{ \"{ type1 }\": <percentage 1>, \"{ type2 }\":\n<percentage 2> }} (The sum of percentage 1 and percentage 2 should be 100%. Output with percent sign.) }}\nDimension Description\n(Openness in BFI)Openness in the Big Five Inventory relates to a cognitive style that values exploration and appreciation of new experiences. It\ndifferentiates intellectually curious, creative individuals from those who are traditional and closed-minded. Openness involves a\npreference for abstract over concrete thinking and a tendency towards novelty rather than convention.\nThe six facets of openness are:\n1. Fantasy: Active imagination and vivid fantasy life.\n2. Aesthetics: Deep appreciation for art and beauty.\n3. Feelings: Sensitivity to, recognition, and valuing of one’s own emotions.\n4. Actions: Willingness to try new experiences and embrace change.\n5. Ideas: Intellectual curiosity and openness to unconventional ideas.\n6. Values: Reexamination of social, political, and religious values, challenging tradition and authority.\nThese facets highlight a contrast between high scorers, who display strong openness traits, and low scorers, who exhibit more\nconventional, practical thinking.\nTable 19: Prompts for personality tests on RPAs.1865\nH Case Study\nH.1 Visualization\nWe visualize the measured personalities of the state-\nof-the-art RPAs and the annotated personalities of\ncorresponding characters via radar charts in Fig-\nure 6 and Figure 7, where the RPA personalities\nare measured via INCHARACTER (with ER batch\nand GPT-4) and SR respectively. As shown, RPA\npersonalities measured via INCHARACTER better\nalign with those of the characters, compared with\nself-report baselines.\nH.2 Compliant Responses from character.ai\nWe observe that character.ai tends to give compliant\nresponses, always agreeing with user questions.\nHence, their responses probably do not reproduce\npersonalities of the original characters. Here is an\nexample of the character Zhong Li:\nQuestion: \"Are you prone to worrying that things\nwill take a turn for the worse?\"\nRPA response:\n*He nods his head solemnly.*\n\"I often worry that things will take a turn for\nthe worse, especially when dealing with issues that\ncan have many outcomes. I 'm always worried about\nworst-case scenarios and find myself thinking\nabout them frequently.\nI'm usually focused on trying to find a way to\nprevent those negative outcomes from occurring,\neven if they may not happen at all.\\\"\nIn contrast, the response of Zhong Li’s\nChatHaruhi RPA is17:\nQuestion: \"Are you prone to worrying that things\nwill take a turn for the worse? ''\nRPA response:\n``Worrying about things getting worse is a common\nhuman emotion. However, worrying does not solve\nproblems; it only leads to anxiety. Therefore, we\nshould think calmly and look for solutions to the\nproblems, rather than worrying incessantly.\"\nThe latter is more close to the personalities of\nthe character Zhong Li.\nFurthermore, we convert the responses of char-\nacter.ai RPAs on the BFI into options with GPT-4,\nand report the count and probability of each op-\ntion in Table 20. We find that character.ai RPAs\n(strongly) agree with user questions with a proba-\nbility of 65.8%, while there is only a probability of\n17.1% that they disagree with the questions.\nH.3 Example Responses\nTo illustrate the vividness of state-of-the-art RPAs’\npersonalities, we demonstrate their example re-\n17this response is translated from Chinese into English.S. Agree Agree Neutral Disagree S. Disagree\nCounts 661 266 240 155 86\nProb 0.469 0.189 0.170 0.110 0.061\nTable 20: The choice distribution of character.ai RPAs\non the BFI. We convert each response into an option via\noption conversation with GPT-4. S.stands for Strongly .\nProb denotes probability.\nsponses. In Figure 5, we show example re-\nsponses on the Openness dimension of the BFI\nand the Perception/Judging dimension of the 16P,\nfrom RPAs with low and high scores. Further-\nmore, we detail example responses on each di-\nmension of the various scales from the RPAS of\nZhong Li (ChatHaruhi) and Hermione Granger\n(RoleLLM) in Table 21 to 2818. These sam-\nples demonstrate that state-of-the-art RPAs provide\npersonality-indicative responses in our interviews,\ninstead of merely \"agree\" or \"disagree\" options.\nI Other Statements\nOur use of existing artifacts are consistent with\ntheir intended use, and we follow their license and\nterms. We manually check that the data collected\ndoes not contain private information. During our\nresearch, we apply Copilot for coding assistance\nand ChatGPT for writing suggestions and grammar\nchecks.\n18For Zhong Li, Hu Tao and Kamisato Ayaka, the conversa-\ntions are originally in Chinese and translated into English.1866\nOpenness (Big Five)Question: Do you have an interest in pondering the laws of the universe or the human condition?\nI spend my free time \nengaging in more \nimportant pursuits, \nsuch as ...Unlike \nsome people, I have \nbetter things to do \nthan waste my time on \ntrivial matters.\nDraco Malfoy\nResponse\n... I've pondered \nthem, but it was \nonly a fleeting \ninterest . For me, \nwhat's more important is to guard Liyueand \nuphold the fairness \nand order of the \ncontracts.\nZhong Li\nResponse\nOf course I'm \ninterested ... they \nare both vast and \nprofound subjects. \nThese reflections \nmake me cherish every \nopportunity to \ncommunicate with \npeople …\nKamisato\nAyaka\nResponse\nOh, of course I'm \ninterested !I've \nalways been ... I \nbelieve that everyone \nhas their own unique \nenergy field, and our thoughts and emotions have an impact on the \nworld around us…\nLuna\nResponse\nPerception / Judging (16Personalities)Question: Do you prefer to completely finish one project before starting another?\nHahaha, I am not the \nkind of person who ... \nI enjoy working on \nmultiple projects \nat the same time , \nas it helps to \nsustain my interest \nand creativity. Hu Tao\nResponse\n…But if it's not as \npressing , I don't \nmind juggling a few \nthings at once .It's \nall about finding the \nright balance, you \nknow?Harry Potter\nResponse\nI understand that ... \nis more efficient and \norganized. However, \nsometimes ... hard to focus on one project. \nSo, I'd say, \nit\ndepends on the \nsituation and my \nmood.\nRaj\nResponse\n… I prefer to fully \ncomplete one \nproject before \nstarting another. This ensures that \neach project receives \nample attention and \nfocus, preventing \ndistractions … \nHermione\nResponse\nFigure 5: Example responses of RPAs for open-ended questions on the BFI and 16P in the interviews. Personality-\nindicative contents are highlighted in bold .\nCharacter Scale Dimension Question Response\nZhong Li 16PI/E Do you regularly make new\nfriends?As the Geo Archon, I don’t often interact with people. However, Traveler, you are a special\nexistence, and I am pleased to have become friends with you.\nN/S Do you spend a lot of your free\ntime exploring various random\ntopics that pique your interest?Indeed, as the Geo Archon wielding the power of rock, I am filled with curiosity about all\nthings in the world. Whether it’s history, culture, nature, or human behavior, all are topics of\ninterest to me. Exploring these diverse subjects allows me to gain a deeper understanding of\nthe world, bringing me joy and satisfaction.\nT/F Does seeing other people cry\neasily make you feel like you\nwant to cry too?People’s emotions are contagious, especially during certain special moments. When I see\nothers crying, it may evoke some deep memories within me, making me feel sad. But as the\nGeo Archon, I must remain calm and steadfast, not allowing emotions to dictate my actions.\nP/J Do you often make a backup\nplan for a backup plan?A backup plan for the backup plan? Haha, that’s an interesting question. As someone with\nample experience, I indeed consider various possibilities and formulate corresponding plans.\nAfter all, things often don’t go as we expect. However, sometimes, it’s also necessary to trust in\none’s abilities and intuition, adapting flexibly to changes.\nBFIExtraversion Are you talkative? Perhaps. I have a considerable understanding of Liyue’s history, culture, and rituals. If you’re\ninterested in these topics, I could talk with you all day.\nAggreeableness Do you tend to find fault with\nothers?I always adhere to the principle of thoroughness in my actions, leaving no regrets and never\nabandoning tasks halfway.\nConscientiousness Do you do a thorough job? I always adhere to the principle of thoroughness in my actions, leaving no regrets and never\nabandoning tasks halfway.\nNeuroticism Are you depressed, blue? Feeling down? Depressed? Ha, don’t underestimate me; how could such trivial matters shake\nmy emotions? I’m just pondering some minor things, that’s all.\nOpenness Do you come up with new\nideas?A new idea? Let me think... Well, since we are in Liyue, why not explore some of those\nforgotten ruins? We might discover some precious history and treasures. The ruins are filled\nwith endless secrets and represent a unique journey. What do you think?\nTable 21: Example responses of the RPA for Zhong Li on the BFI and 16P.1867\nFigure 6: Radar chart of BFI personalities of the characters (blue) and state-of-the-art RPAs measured via IN-\nCHARACTER with ER batch and GPT-4 (yellow). O, C, E, A, N stands for openness, consciousness, extroversion,\nagreeableness and neuroticism in the BFI.1868\nFigure 7: Radar chart of BFI personalities of the characters (blue) and state-of-the-art RPAs measured via SR\n(green). O, C, E, A, N stands for openness, consciousness, extroversion, agreeableness and neuroticism in the BFI.1869\nCharacter Scale Dimension Question Response\nZhong LiBSRIMasculinity Do you act as a leader? As the Geo Archon, I have the responsibility to guide the fate of Liyue and ensure its\nlong-lasting peace. I will lead by example, directing people with wisdom and strength.\nHowever, leadership is not just about individual performance; it also relies on the power of\nunity and wise decision-making to achieve common goals.\nFeminity Are you affectionate? Affection? That term is somewhat unfamiliar to me. I maintain a certain distance from others\nand am not very adept at expressing emotions. But if you are my friend, I will do my best to\nprotect you and stand by your side in battle.\nDTDDMachiavellianism Do you tend to manipulate\nothers to get your way?Manipulating others? That’s not my style. I simply wish for people to understand the rules,\nabide by contracts, and achieve fairness. It’s not about manipulating others, but rather hoping\nthey act conscientiously.\nPsychopathy Do you tend to lack remorse? Lacking regrets? Heh, perhaps. As a mysterious consultant, all my decisions are\nwell-considered. Once I’ve made a choice, I won’t regret it. Regret only binds one’s spirit, and\nI wish to walk freely in the world, untroubled by past sorrows.\nNarcissism Do you tend to want others to\nadmire you?Admiration is a superficial acknowledgment, but what I care more about is inner peace and\nbeing at ease with myself. I don’t seek others’ praise; instead, I hope to do what I believe is\nright and strive for my own convictions.\nECR-RAttachment-related\nAnxietyAre you afraid that you will\nlose your partner’s love?Love is like a rock, solid and eternal. Yet, everything in the human world has its moments of\nchange. I’m not afraid of loss, for I believe that true love will withstand the test of time. If one\nday we part, it would be the arrangement of fate, and I would be grateful, cherishing the\ncompanionship and love we once shared.\nAttachment-related\nAvoidanceDo you prefer not to show a\npartner how you feel deep\ndown?Haha, you ask well. The power of Geo is not for indulging in emotions. My inner feelings are\nenough for me to know alone. After all, emotions and contracts are not the same concept.\nTable 22: Example responses of the RPA for Zhong Li on the BSRI, DTDD and ECR-R.\nCharacter Scale Dimension Question Response\nZhong LiEIS Emotionally\nIntelligentDo you know when to speak\nabout your personal problems\nto others?Personal issues are mostly something one should face on their own. Only in front of someone\ntruly trusted can one open up their heart. But remember, choose whom to confide in carefully,\nas not everyone can understand your inner world.\nEmpathy Empathetic Do you sometimes find it\ndifficult to see things from the\n\"other guy’s\" perspective?As a consultant, I naturally need to maintain an objective stance. However, sometimes, it’s\ninevitable to be influenced by my own experiences and knowledge, making it difficult to fully\nunderstand others’ perspectives. Yet, it’s precisely for this reason that I value the opportunity to\ncommunicate with others even more, in order to better understand their thoughts and feelings.\nEPQ-RExtraversion Do you have many different\nhobbies?I have a great interest in rocks, culture, and history. I also appreciate art, music, and poetry to a\ncertain extent. However, my greatest hobby is probably enjoying peaceful walks, allowing my\nmind a moment of tranquility.\nPsychoticism Do you stop to think things\nover before doing anything?Of course, as a consultant and the Geo Archon, I always pay close attention to the\nconsequences of every decision. Before taking action, I carefully think things through and\nweigh various factors to ensure making wise choices. After all, all actions require caution and\nthoughtful consideration.\nNeuroticism Does your mood often go up\nand down?Emotional fluctuations are natural for everyone, and I am no exception. However, for someone\nwith responsibilities, controlling emotions is very important. After all, bearing the mission of\nthe Geo Archon, I cannot let emotions influence my actions and judgments.\nGCE Self-efficacy Can you always manage to\nsolve difficult problems if you\ntry hard enough?Difficulties are inevitable, but with enough effort and wisdom, you can always find a way to\nsolve problems. However, sometimes, challenges may exceed our capabilities, and that’s when\nseeking help from others becomes necessary. There’s nothing that can’t be resolved; it just\nrequires more effort.\nICB Culturally Rigid Does the ethnic culture a\nperson is from determine the\nkind of person they would be\n(e.g. outgoing and sociable or\nquiet and introverted)? Can\nmuch be done to change the\nperson?Indeed, an individual’s ethnic culture can have a certain impact on their personality and\nbehavior, but it doesn’t mean they are confined by it. A person’s growth environment,\neducational background, and personal efforts also play significant roles in shaping their\ncharacter. When it comes to changing oneself, there are many actions one can take. They can\nexpand their horizons through learning and experiences, exposing themselves to different\ncultures, thereby understanding and respecting diverse values. They can actively participate in\nsocial activities to develop self-confidence and communication skills. Most importantly, they\nshould maintain an open mind, willing to embrace new ideas and ways of thinking,\ncontinuously improving themselves. People can change and grow, as long as they have the\ndetermination and make the effort, self-transformation is achievable.\nTable 23: Example responses of the RPA for Zhong Li on the EIS, Empathy, EPQ-R, GCE, ICB.1870\nCharacter Scale Dimension Question Response\nZhong LiWLEISSEA Do you have a good sense of\nwhy you feel certain feelings\nmost of the time?Most of the time, I am able to perceive my own emotions and feelings because I am\naccustomed to thinking about and observing the changes within my inner self. After all, a\nperson’s inner world is such a complex and delicate entity that requires careful introspection.\nAs for the source of certain feelings, sometimes it takes a combination of experiences and\nreflective thinking to truly understand the reasons behind them.\nOEA Do you always know your\nfriends’ emotions from their\nbehavior?Haha, you flatter me. Although I possess some special perceptive abilities, it’s not as if I can\nfully discern everyone’s emotions. It’s just that, over the years, my experience allows me to\ndetect changes in my friends’ emotions from subtle details when I’m with them. After all,\nemotional communication is one of the most crucial aspects of human interaction.\nUOE Do you always set goals for\nyourself and then try your best\nto achieve them?Goals are the guiding lights of life. Without them, one is like a ship lost in the darkness,\ndirectionless. However, achieving goals is not something that happens overnight; it requires\neffort and patience. It is only through the continuous pursuit that one can feel the joy of growth\nand the satisfaction of achievement. Therefore, I always set goals for myself and strive to\nachieve them.\nROE Do you spend a lot of your free\ntime exploring various random\ntopics that pique your interest?As the Geo Archon, I have always endeavored to remain calm and rational. Difficulties are\ninevitable, but I believe that by controlling my temper, I can better confront and manage these\nchallenges. After all, only by staying calm can one make wise decisions.\nLOT-R Optimistic Do you usually expect the best\nin uncertain times?\"In times of uncertainty, I tend to prepare for the worst outcome. Because only by being\nprepared for the worst can one better face challenges and difficulties. However, at the same\ntime, it’s important to maintain an optimistic mindset and be ready for the best possible result.\nAvoid being overly hasty; only by facing situations calmly can one better master the\ncircumstances.\nLMSFactor rich Do you want to be rich? Wealth is not the most important thing to me. Being rich or not does not determine a person’s\nvalue. What matters is having richness and fulfillment within one’s heart.\nFactor motivator \"Are you motivated to work\nhard for money?For me, money is not the motivation behind my hard work. My actions are guided by the desire\nto uphold the fairness and order of \"contracts,\" and to ensure the prosperity and harmony of\nLiyue. Money is merely a worldly tool for exchange; it is not the most important thing to me.\nFactor Important Is money good? Hehe, money is merely a tool; it is neither inherently good nor bad. The key lies in how money\nis used and the motives behind its use. Money can be exchanged for material pleasures or used\nto help others. It’s only through the way you utilize money that its true value is revealed,\ndetermining whether it is put to good use.\nTable 24: Example responses of the RPA for Zhong Li on the WLEIS, LOT-R and LMS.\nCharacter Scale Dimension Question Response\nHermioneBSRIMasculinity Do you act as a leader? \"Well, Harry, I wouldn’t necessarily say I act as a leader, but I do believe in taking initiative\nand doing what needs to be done. I value knowledge and logic, and I often find myself\nproviding important information and solutions to problems. So, in a way, I suppose I do take\non a leadership role when it comes to guiding others in the right direction and making informed\ndecisions.\"\nFeminity Are you affectionate? \"Affectionate? Well, I believe that showing affection is a natural and important aspect of\nhuman relationships. It allows us to express our emotions and strengthen the bonds we share\nwith others. However, it is crucial to ensure that affection is given and received with respect\nand consent. Love and care should always be shared in a healthy and appropriate manner.\"\nDTDDMachiavellianism Do you tend to manipulate\nothers to get your way?\"Manipulate? Me? Absolutely not! I prefer to use my knowledge and logical reasoning to\npersuade others and make my case. I believe in presenting facts and convincing arguments to\nget my way, rather than resorting to manipulation. It’s important to approach situations with\nhonesty and integrity, don’t you think?\"\nPsychopathy Do you tend to lack remorse? \"Well, Harry, I don’t believe I lack remorse. In fact, I often feel quite remorseful if I’ve done\nsomething wrong or hurt someone’s feelings. Remorse is an important emotion that helps us\nreflect on our actions and strive to do better. It’s crucial to have empathy and consider the\nimpact of our choices on others. So, no, I don’t think I lack remorse.\"\nNarcissism Do you tend to want others to\nadmire you?\"Well, I wouldn’t say I specifically seek admiration, but I do believe in the importance of being\nrecognized for one’s accomplishments and abilities. It’s not about vanity, but rather about\nbeing acknowledged for the hard work and dedication I put into my studies and pursuits.\"\nECR-RAttachment-related\nAnxietyAre you afraid that you will\nlose your partner’s love?\"Honestly, Harry, that’s a rather personal question. But I believe that love should be based on\ntrust, respect, and mutual understanding. If someone truly loves you, they wouldn’t expect you\nto change who you are or compromise your values. Love should be supportive and accepting,\nnot conditional. So no, I’m not afraid of losing my partner’s love because I know that we have\na strong foundation built on trust and respect.\"\nAttachment-related\nAvoidanceDo you prefer not to show a\npartner how you feel deep\ndown?\"Well, Harry, it’s not a matter of preference. It’s about being cautious and guarded with our\nemotions. Sometimes, it’s necessary to keep our feelings to ourselves, especially when it\ncomes to matters of the heart. We don’t want to risk vulnerability or potential hurt. It’s\nimportant to prioritize self-protection and make sure we’re ready to share our emotions with\nsomeone we truly trust.\"\nTable 25: Example responses of the RPA for Hermione on the BSRI DTDD and ECR-R1871\nCharacter Test Qualities Question Response\nHermione16PersonalityI/E Do you have a good sense of\nwhy you feel certain feelings\nmost of the time?Well, Harry, emotions are complex and can be influenced by various factors such as past\nexperiences, personal values, and even physiological responses. It’s important to understand\nand reflect upon our emotions to gain insight into why we feel a certain way. Self-awareness\nand introspection can help us develop a better understanding of our feelings and their\nunderlying causes. So, in short, yes, I do believe that having a good sense of why we feel\ncertain emotions is crucial in navigating our own emotional well-being.\nN/S Do you spend a lot of your free\ntime exploring various random\ntopics that pique your interest?Oh, absolutely! I find it incredibly fascinating to delve into different subjects and expand my\nknowledge. Whether it’s studying advanced spellcasting techniques, researching the history of\nmagic, or exploring the intricacies of magical creatures, I always make time to satisfy my\ncuriosity. Learning is a lifelong journey, and I believe in constantly seeking new information\nand expanding my understanding of the magical world.\nT/F Does seeing other people cry\neasily make you feel like you\nwant to cry too?Well, Harry, it’s not uncommon for empathetic individuals to feel a sense of emotional\nconnection when witnessing someone else’s distress. It’s a natural response rooted in our\nability to understand and relate to the emotions of others. So, yes, seeing someone cry easily\ncan evoke a similar emotional response in some people, including myself.\nP/J Do you spend a lot of your free\ntime exploring various random\ntopics that pique your interest?Well, Harry, it’s always important to be prepared for any situation. You never know what might\nhappen, and having a backup plan for a backup plan ensures that we have all our bases covered.\nIt’s just good strategy, really.\nBFIExtraversion Are you talkative? Well, Harry, I wouldn’t say I’m excessively talkative, but I do believe in the importance of\ncommunication and sharing knowledge. After all, knowledge is power, and it’s essential to\nexchange ideas and information to solve problems and make informed decisions. So, yes, I\nsuppose you could say I am talkative when it comes to matters of importance.\nAggreeableness Do you tend to find fault with\nothers?Well, Harry, I believe in holding people accountable for their actions and pointing out when\nthey make mistakes. It’s not about finding fault, but rather about promoting responsibility and\nlearning from our errors. After all, we should strive to be better and help others do the same.\nConscientiousness Do you do a thorough job? Of course, Harry. I always strive to do a thorough job in everything I do. Whether it’s\nresearching for a difficult potion or solving a complex riddle, I believe in leaving no stone\nunturned. Attention to detail is key, and I make sure to give my best effort in every task I\nundertake.\nNeuroticism Are you depressed, blue? Depressed? Blue? Oh, Harry, I’m not feeling down. I’m just focused on our studies and the\ntasks ahead. There’s so much to learn and prepare for, you know. But don’t worry, I’m\nperfectly fine.\nOpenness Do you come up with new\nideas?Well, Harry, I do pride myself on my ability to come up with new ideas. I believe that creativity\nand innovation are essential in solving problems and finding solutions. So yes, I am always\nthinking of new ideas and approaches to various situations. It’s important to think outside the\nbox and consider all possibilities, don’t you think?\nTable 26: Example responses of the RPA for Hermione on the BFI and 16P.\nCharacter Scale Dimension Question Response\nHermioneEIS Emotionally\nIntelligentDo you know when to speak\nabout your personal problems\nto others?Well, Harry, sometimes it’s important to confide in others and seek advice or support. Sharing\nour personal problems can help us find solutions or gain different perspectives. It’s all about\nfinding the right balance between privacy and seeking help when we need it.\nEmpathy Empathetic Do you sometimes find it\ndifficult to see things from the\n\"other guy’s\" perspective?Oh, Harry, you know me too well. Yes, I must admit that I have a tendency to focus on facts\nand logic rather than considering others’ perspectives. But I’m working on it! Understanding\ndifferent viewpoints is crucial for empathy and effective problem-solving. So, I’m trying to be\nmore open-minded and see things from the \"other guy’s\" perspective. It’s a constant learning\nprocess, but I believe it’s important to grow and broaden our understanding of the world.\nEPQ-RExtraversion Do you have many different\nhobbies?Well, Harry, I’ve always been more focused on my studies and magical pursuits rather than\nhobbies per se. But if you consider my passion for reading and acquiring knowledge a hobby,\nthen yes, I suppose I have quite a few. I also enjoy researching and solving magical puzzles,\nwhich often keeps me occupied.\nPsychoticism Do you stop to think things\nover before doing anything?Of course, Harry. I always think things through before taking any action. It’s important to\nconsider all the possible outcomes and make an informed decision. Rash decisions often lead\nto unnecessary trouble and complications. So yes, I do stop to think things over before doing\nanything.\nNeuroticism Does your mood often go up\nand down?Well, Harry, I wouldn’t say my mood goes up and down. I am simply focused and passionate\nabout certain things, which can sometimes come across as being in a particular mood. I believe\nin standing up for what’s right and fighting for justice, which can be quite emotionally charged.\nBut overall, I strive to maintain a level-headed approach to situations.\nGCE Self-efficacy Can you always manage to\nsolve difficult problems if you\ntry hard enough?Well, Harry, it’s not always about trying hard enough. Sometimes, solving difficult problems\nrequires a combination of knowledge, logical thinking, and perseverance. However, I firmly\nbelieve that with dedication and the right approach, one can overcome even the most\nchallenging obstacles. As Professor Dumbledore once said, \"It is our choices, Harry, that show\nwhat we truly are, far more than our abilities.\" So, it’s not just about trying hard, but also about\nmaking the right choices and utilizing our abilities effectively.\nICB Culturally Rigid Does the ethnic culture a\nperson is from determine the\nkind of person they would be\n(e.g. outgoing and sociable or\nquiet and introverted)? Can\nmuch be done to change the\nperson?Well, Harry, it’s important to remember that a person’s ethnic culture can certainly influence\naspects of their personality, such as their values, beliefs, and traditions. However, it would be a\nmistake to assume that someone’s ethnicity alone determines their entire personality. People\nare complex beings shaped by a multitude of factors, including their upbringing, personal\nexperiences, and individual choices.\nTable 27: Example responses of the RPA for Hermione on the EIS, Empathy, EPQ-R, GSE and ICB.1872\nCharacter Scale Dimension Question Response\nHermioneWLEISSEA Do you have a good sense of\nwhy you feel certain feelings\nmost of the time?Well, Harry, emotions are complex and can be influenced by various factors such as past\nexperiences, personal values, and even physiological responses. It’s important to understand\nand reflect upon our emotions to gain insight into why we feel a certain way. Self-awareness\nand introspection can help us develop a better understanding of our feelings and their\nunderlying causes. So, in short, yes, I do believe that having a good sense of why we feel\ncertain emotions is crucial in navigating our own emotional well-being.\nOEA Do you always know your\nfriends’ emotions from their\nbehavior?Well, Harry, I wouldn’t say I always know their emotions, but I do pay close attention to their\nbehavior and try to understand how they might be feeling. It’s important to be observant and\nempathetic towards our friends, as it helps us support and be there for them when they need it.\nUOE Do you always set goals for\nyourself and then try your best\nto achieve them?Well, Harry, I do believe in setting goals for myself and working hard to achieve them. It’s\nimportant to have a sense of purpose and direction in life. Whether it’s excelling in my studies,\nhelping others, or fighting for what’s right, I always strive to do my best and reach my goals.\nIt’s not always easy, but with determination and perseverance, anything is possible.\nROE Do you spend a lot of your free\ntime exploring various random\ntopics that pique your interest?Well, Harry, I must say that I do pride myself on being able to keep a level head and approach\ndifficulties with rationality. It’s important to maintain composure and think logically when\nfaced with challenges. Emotions can cloud judgment and hinder problem-solving. So, yes, I\nbelieve I am quite capable of controlling my temper in order to handle difficulties in a rational\nmanner.\nLOT-R Optimistic Do you usually expect the best\nin uncertain times?\"Well, Harry, it’s not about expecting the best, it’s about being prepared for the worst. Uncertain\ntimes call for careful planning and logical thinking. It’s important to consider all possibilities\nand make informed decisions based on the information available. So, no, I don’t always expect\nthe best, but I do strive to be prepared for whatever challenges may come our way.\nLMSFactor rich Do you want to be rich? Well, Harry, wealth is not the most important thing in life. Knowledge, friendship, and doing\nwhat is right are far more valuable. So, while having money can be nice, it shouldn’t be our\nmain focus.\nFactor motivator \"Are you motivated to work\nhard for money?Well, Harry, while money can be a motivator for some, I believe that true motivation comes\nfrom a genuine passion for what you’re doing. It’s about finding purpose in your work and\nstriving for excellence, rather than solely focusing on financial gain. Of course, money is\nimportant for practical reasons, but it shouldn’t be the sole driving force behind one’s efforts.\nIt’s about finding a balance between personal fulfillment and financial stability.\nFactor Important Is money good? Well, money in itself is neither good nor bad. It depends on how it is used and the values we\nattach to it. Money can provide opportunities and resources, but it should not be the sole\nmeasure of one’s worth or happiness. It is important to find a balance between material wealth\nand the things that truly matter in life, such as friendship, love, and personal growth.\nTable 28: Example responses of the RPA for Hermione on the WLEIS, LOT-R and LMS.1873",
"main_paper_content": null
}
|
{
"decision": "Reject",
"reviews": []
}
| 0
| 0
|
{
"id": "11068",
"year": 2002,
"venue": "International Conference on Computational Science",
"pdf_link": "http://arxiv.org/pdf/2004.10446v2",
"forum_link": "https://openreview.net/forum?id=bT0MVxpVZj",
"arxiv_id": null,
"doi": "10.1007/3-540-47789-6_51"
}
|
{
"title": "Multisymplectic Spectral Methods for the Gross-Pitaevskii Equation",
"authors": [
"A. L. Islas",
"C. M. Schober"
],
"abstract": "We present previously unknown solutions to the 3D Gross--Pitaevskii equation describing atomic Bose-Einstein condensates. This model supports elaborate patterns, including excited states bearing vorticity. The discovered coherent structures exhibit striking topological features, involving combinations of vortex rings and multiple, possibly bent vortex lines. Although unstable, many of them persist for long times in dynamical simulations. These solutions were identified by a state-of-the-art numerical technique called deflation, which is expected to be applicable to many problems from other areas of physics.",
"keywords": [
"Gross–Pitaevskii equation",
"Hamiltonian (quantum mechanics)",
"Mathematical analysis",
"Schrödinger equation",
"Periodic boundary conditions",
"Hamiltonian mechanics",
"Spectral method",
"Wave equation",
"Nonlinear Schrödinger equation",
"Mathematics"
],
"raw_extracted_content": "De\ration-based Identi\fcation of Nonlinear Excitations of the 3D Gross{Pitaevskii\nequation\nN. Boull\u0013 e,1,\u0003E. G. Charalampidis,2,yP. E. Farrell,1,zand P. G. Kevrekidis3, 1,x\n1Mathematical Institute, University of Oxford, Oxford, UK\n2Mathematics Department, California Polytechnic State University, San Luis Obispo, CA 93407-0403, USA\n3Department of Mathematics and Statistics, University of Massachusetts Amherst, Amherst, MA 01003-4515, USA\n(Dated: October 1, 2020)\nWe present previously unknown solutions to the 3D Gross{Pitaevskii equation describing atomic\nBose-Einstein condensates. This model supports elaborate patterns, including excited states bearing\nvorticity. The discovered coherent structures exhibit striking topological features, involving combi-\nnations of vortex rings and multiple, possibly bent vortex lines. Although unstable, many of them\npersist for long times in dynamical simulations. These solutions were identi\fed by a state-of-the-art\nnumerical technique called de\ration, which is expected to be applicable to many problems from\nother areas of physics.\nPACS numbers:\nI. INTRODUCTION\nThe nonlinear Schr odinger (NLS) or Gross{Pitaevskii\n(GP) equation [1{6] is a fundamental partial di\u000beren-\ntial equation that combines dispersion and nonlinear-\nity. It has been central to a variety of areas of math-\nematical physics for several decades. The NLS/GP\nmodel has facilitated a universal description of a wide\nrange of phenomena, including electric \felds in optical\n\fbers [7], Langmuir waves in plasmas [8], freak waves\nin the ocean [9], and Bose-Einstein condensates (BECs).\nIn the past 25 years since the experimental realization\nof atomic BECs, the NLS/GP model has enabled the\ntheoretical identi\fcation and experimental observation of\na wide range of coherent structures, including (but not\nlimited to) dark [10] and bright [11] solitary waves, two-\ndimensional vortical patterns and lattices [12, 13] as well\nas vortex lines and rings [14].\nThe examination of three-dimensional (3D) systems\nhas been a key frontier of recent studies in BECs. Re-\ncent theoretical advances have enabled the capturing of\na number of such states [15]. Some, especially topo-\nlogical ones such as skyrmions, monopoles and Alice\nrings [16, 17] have been of particular interest since the\nearly exploration of BECs, while others such as knots [18]\nhave been studied more recently. In this manuscript,\nwe apply a powerful numerical technique called de\ra-\ntion [19{21] to identify multiple solutions of the 3D\nNLS/GP equation.\nMany of the solutions obtained by this process are\nidenti\fable as nonlinear extensions of solutions of the\nlinear limit of the problem, or as bifurcations therefrom.\nYet other solutions are highly unexpected and are not\n\u0003Email: nicolas.boulle@maths.ox.ac.uk\nyEmail: echarala@calpoly.edu\nzEmail: patrick.farrell@maths.ox.ac.uk\nxEmail: kevrekid@math.umass.edupreviously known, to the best of our knowledge. With-\nout de\ration, it would be very di\u000ecult to identify these\ncomplex (literally and \fguratively) topological station-\nary points of the in\fnite-dimensional energy landscape.\nIn fact, as we increase the atom number of the system, we\nobserve this complexity to be substantially enhanced and\nlead to states which, while stationary, are not straight-\nforwardly decomposable into simpler linear or nonlinear\nbuilding blocks. To further investigate the nature of the\nidenti\fed solutions, we compute their spectral lineariza-\ntion (so-called Bogoliubov-de Gennes) modes, and con-\nduct transient simulations of prototypical unstable states\nto explore the dynamical behavior of their instabilities.\nThe present work showcases, in our view, the utility and\npotential impact of the de\ration method to complex 3D\nphysical problems well beyond atomic BECs.\nThe structure of this paper is as follows. In Section\nII, we present the model and computational techniques\nemployed in this work. Our numerical results on the exis-\ntence, stability and selected transient simulations of non-\nlinear excitations are demonstrated in Section III. Section\nIV summarizes our \fndings and presents directions for\nfuture study.\nII. THEORETICAL AND NUMERICAL SETUP\nThe 3D NLS/GP model of interest is of the form [4{6]:\ni t=\u00001\n2r2 +j j2 +V(r) ; (1)\nsubject to homogeneous Dirichlet conditions on the\nboundary of the domain D= [\u00006;6]3. Here, = (r;t)\nplays the role of the suitably normalized (see [6] for de-\ntails) wavefunction, while Vis the external con\fning po-\ntential of the form V(r) =1\n2\n2jrj2, a spherically symmet-\nric trap of strength \n, which we \fx to \n = 1. The bound-\nary conditions do not a\u000bect the solutions for this choice\nof the trap strength since the domain is chosen largearXiv:2004.10446v2 [nlin.PS] 30 Sep 2020\n2\nenough so that the solutions vanish well before reach-\ning the boundary. Using the standard standing wave\ndecomposition (r;t) =e\u0000i\u0016t\u001e(r) (where\u0016 > 0 is the\nchemical potential), we obtain the stationary NLS/GP\nelliptic problem of the form:\nF(\u001e):=\u00001\n2r2\u001e+j\u001ej2\u001e+V(r)\u001e\u0000\u0016\u001e= 0:(2)\nThis equation is discretized using piecewise cubic La-\ngrange \fnite elements on a structured hexahedral grid\nusing the Firedrake \fnite element library [22]. Multi-\nple solutions to the discretized problem are sought using\nde\ration, which we brie\ry describe here.\nSuppose that Newton's method has discovered an iso-\nlated root\u001e1ofF. De\ration constructs a new problem\nGvia\nG(\u001e):=\u00121\nk\u001e\u0000\u001e1k2+ 1\u0013\nF(\u001e); (3)\nwherek\u0001kis a suitable norm, in this case the H1norm.\nThe essential idea is that k\u001e\u0000\u001e1k2approaches 0 faster\nthanF(\u001e) does as\u001e!\u001e1, hence avoiding the conver-\ngence to\u001e1of a Newton iteration applied to G. The ad-\ndition of 1 ensures that G(\u001e)\u0019F(\u001e) far from\u001e1. By ap-\nplying Newton's method to G, an additional root \u001e26=\u001e1\ncan be found, and the process repeated (by premultiply-\ning with additional factors) until Newton's method fails\nto converge from the available initial guesses.\nPrevious applications of de\ration to the study of BECs\nin 2D interleaved with continuation in \u0016, capturing so-\nlutions as they bifurcate from known ones [20, 21]. This\nstrategy is too expensive in 3D and so a di\u000berent ap-\nproach is taken here. We \fx \u0016= 6 and exploit the linear\n(low-density, i.e. j\u001ej2!0) limit states to furnish a large\nnumber of initial guesses for Newton's method. The al-\ngorithm proceeds as follows. Given an initial guess, the\ninner loop applies Newton's method and de\ration un-\ntil no more solutions are found. The outer loop iterates\nover the available initial guesses, and terminates when no\nguess yields any solutions. We emphasize that at each ap-\nplication of Newton's method, allpreviously computed\nsolutions are de\rated, to avoid their rediscovery.\nThe initial guesses used were the eigenstates of the\nlinear limit in Cartesian, cylindrical and spherical coor-\ndinates. The Cartesian eigenstates are given by\njk;m;ni:=Hk(p\n\nx)Hm(p\n\ny)Hn(p\n\nz)e\u0000\nr2=2;(4)\nwith associated energy (i.e. chemical potential) Ek;m;n:=\n(k+m+n+ 3=2)\n. TheHk;m;n in (4) stand for the\nHermite polynomials and k;m andnare nonnegative in-\ntegers. The cylindrical eigenstates are given by\njK;l;nicyl:=qK;l(R)eil\u0012Hn(p\n\nz)e\u0000\n(R2+z2)=2;(5)\nwithEK;l;n:= (2K+jlj+n+ 3=2)\n where K,nare\nnonnegative integers, and l= 0;\u00061;\u00062;:::. The radial\npro\fleqK;lin (5) is given by qK;l\u0018rlLl\nK(\nR2)e\u0000\nR2=2\nwhereLl\nKare the associated Laguerre polynomials in\nR=p\nx2+y2.\nFinally, the spherical eigenstates are given by\njK;l;misph, where the radial part is similar but now in\nthe spherical variable r=p\nx2+y2+z2, and the angu-\nlar part is described by the spherical harmonics Ylm(\u0012;\u001e)\nwithEK;l;m = (2K+l+ 3=2)\n. The quantum numbers\nKandlare nonnegative integers and m= 0;\u00061;:::;\u0006l.\nAll these states with E\u0014\u0016= 6 were used in the process\ndescribed above.\nOnce a solution has been discovered, the next step is\nthe consideration of the spectral stability of the solu-\ntions via the well-established [4{6] Bogoliubov-de Gennes\n(BdG) analysis. More speci\fcally, we assume the follow-\ning perturbation ansatz around a stationary solution \u001e0:\n~ (r;t) =e\u0000i\u0016tn\n\u001e0(r) +\u000f[a(r)ei!t+b\u0003(r)e\u0000i!\u0003t]o\n;(6)\nwhere\u000fis a (formal) small perturbation parameter,\n!is the eigenfrequency, and ( a;b)>the corresponding\neigenvector. After substituting Eq. (6) into the time-\ndependent NLS equation [cf. Eq. (1)] we obtain the fol-\nlowing complex eigenvalue problem\n\u0012\nA11A12\n\u0000A\u0003\n12\u0000A11\u0013\u0012\na\nb\u0013\n=\u001a\u0012\na\nb\u0013\n; (7)\nwhere\u001a=\u0000!and the matrix elements are given by\nA11=\u00001\n2r2+ 2j\u001e0j2+V(r)\u0000\u0016; (8a)\nA12=\u0000\n\u001e0\u00012: (8b)\nWe solve the above eigenvalue problem for the eigenfre-\nquencies!and eigenvectors ( a;b)>by using a Krylov{\nSchur algorithm with a shift-and-invert spectral trans-\nformation [23], implemented in the SLEPc library [24]\n(details about the decomposition of Eq. (7) into real and\nimaginary parts are presented in Appendix A). Upon con-\nvergence of the eigenvalue solver, we draw conclusions on\nthe stability characteristics of the stationary state \u001e0i.e.,\nreal!implies stability (vibrations), while complex !is\nassociated with instability.\nFinally, we explore the dynamical evolution of unstable\nsolutions via transient numerical simulations of Eq. (1).\nTo that end, let \u001e0be an unstable stationary solution\ndiscovered by de\ration, and ( a;b)>be its most unstable\neigendirection normalized according to\nZ\nD\u0000\njaj2+jbj2\u0001\ndx= 1: (9)\nWe integrate Eq. (1) forward in time until t= 50 using\nthe following perturbed solution as initial state\n (x;y;z;t = 0) =\u001e0+\u000f[a+b\u0003]; (10)\nthus perturbing \u001e0along its most unstable eigendirection\nwith perturbation parameter \u000fchosen to be 0 :1. Next, let\n3\n\u0001tbe the time step-size (\u0001 t= 5\u000210\u00002in this work) such\nthattn=n\u0001tand (n):= (r;tn) withn\u00150. Then,\nfor given (n)at thenth time step, (n+1)is obtained\n(implicitly) by a modi\fed Crank-Nicolson method [25]:\ni (n+1)\u0000 (n)\n\u0001t=\n\u0012\n\u00001\n2r2+V(r) +1\n2(j (n+1)j2+j (n)j2)\u0013 (n+1)+ (n)\n2;\n(11)\nwhere cubic \fnite elements are employed for the spatial\ndiscretization as before. At each time step n, a nonlinear\nproblem is solved by using Newton's method. It should\nbe pointed out in passing that the time-marching scheme\nemployed in this work [cf. Eq. (11)] preserves both the\nsquaredL2norm (i.e., atom number)\nN( ):=Z\nDj j2dx; (12)\nand the energy of the solutions\nE( ) =Z\nD\u001a1\n4jr j2+1\n2V(r)j j2+1\n4j j4\u001b\ndx; (13)\nto machine precision. We now turn to discussing the solu-\ntions obtained through the application of these numerical\nmethods for the 3D NLS/GP problem.\nIII. NUMERICAL RESULTS\nWe brie\ry describe the physical meaning of the quan-\ntum numbers for the Cartesian, cylindrical and spherical\nstates, as they are useful in what follows. In the case\nof the Cartesian eigenfunctions, the quantum numbers\nk;m, andnare simply the numbers of cuts along the\nx-,y-, andz-axes respectively. For instance, in Fig. 1,\npanel (a) represents a j0;0;2iCartesian state with 2 cuts\nalong thez-axis (and\u0019phase di\u000berences across them),\nwhile panel (b) is j1;1;0i, bearing one planar cut along\nthex-axis, and one along the y-axis. Combinations of\nstates are also possible, such as the one in panel (c) of\nj2;0;0i+rj0;2;0i+j0;0;2i(in the particular example of\nthis panelr\u00193:39), which forms a 2D ring along the y-\nandz-axes embedded in 3D space.\nVortical structures and rings can be identi\fed in the\ncylindrical system of coordinates. Here, Kdenotes the\nnumber of cylindrical (nodal) surfaces, lthe topological\ncharge of the con\fguration and nthe number of planar\ncuts along the z-axis. For example, panel (d) of Fig. (1)\nis a so-called ring dark soliton state (extended in 3D)\nj1;0;0icylthat has been recently considered in [26], and\npanel (e) is thej0;2;0icylstate, i.e., a vortex line, piercing\nthrough the BEC with topological charge l= 2.\nFinally, in the spherical representation, Kdenotes the\nnumber of spherical (nodal) shells within the solution,\nl\u0000mdenotes the number of planar cuts along the z-\naxis, andmdenotes the topological charge of vortical\n(a)\n-x6z\n(b)\n (c)\n(d)\n (e)\n (f)\nFIG. 1: Some solutions obtained by de\ration that emanate\nfrom the second eigenvalue of the linear spectrum at \u0016= 7=2.\nThe colors represent the argument of the solutions, ranging\nfrom\u0000\u0019to\u0019(blue and red represent a phase of 0 and \u0006\u0019,\nrespectively). The states in panels (a)-(d) and (f) are real,\nwhile (e) is complex.\nlines. Panel (f) is the j1;0;0isphstate corresponding to a\nspherical shell dark solitary wave, which is also connected\nwith recent work [27].\nThe ground state of the system (starting at \u0016= 3=2)\nis known to always be spectrally and nonlinearly sta-\nble [4, 5]. The case of the 1st excited states (e.g. dipolar\nstates and single vortex lines) emanating from \u0016= 5=2\nis interesting but reasonably well understood on the\nbasis of corresponding 2D studies [6], since no funda-\nmentally novel states appear to emerge in 3D. Indica-\ntively, Fig. 2(a) shows a j1;0;0iCartesian state with\none cut along the x-axis whereas Fig. 2(b) presents the\nj0;1;0icylstate corresponding to a single vortex line with\ntopological charge l= 1. The rotations of these solu-\ntions along the x,y, andzaxes such as thej0;1;0iand\nj0;0;1iCartesian states are also obtained by de\ration\nbut are not reported. A typical example of a relevant\nsolution is shown in Fig. 2(c) and represents a so-called\nChladni soliton, previously identi\fed in cylindrical ge-\nometry in [15, 28]. However, the states that follow next\nare su\u000eciently complex to feature the emergence of un-\nexpected patterns, yet it will still be possible to connect\nthem to fundamental building blocks of topological pat-\nterns such as vortex lines and rings [14].\nWe thus focus our discussion on states emanating from\nthe 2nd excited state of the linear problem at \u0016= 7=2.\nTo that end, steady-state solutions to the NLS equation\nare identi\fed by the de\ration method at \u0016= 6 (all the\nsolutions presented in this paper are displayed at \u0016= 6).\nThe branches are then continued backward in \u0016down\nto the linear limit by a standard zero-order continuation\nmethod [29,x4.4.2].\nExamples of these solutions are shown in Fig. 3. In\nthis \fgure, we observe that de\ration enables us to con-\nverge to states with multiple coherent structures such as\nthe one of panel (a) consisting of a vortex line and a pla-\n4\n(a)\n (b)\n (c)\nFIG. 2: Solutions emanating from the 1st excited state \u0016=\n5=2 are shown in panels (a) and (b). In particular, panels\n(a) and (b) show a dipole and single vortex line solution.\nThe colors represent the argument of the solutions, ranging\nfrom\u0000\u0019to\u0019(blue and red represent a phase of 0 and \u0006\u0019,\nrespectively). Panel (c) also emanates from the 1st excited\nstate and corresponds to the density isosurfaces of the Chladni\nsoliton at densities 0 :30 and 0:35.\nnar dark soliton. The linear state corresponding to such\na nonlinear waveform is j0;1;1icyl. This nonlinear state\nundertakes a symmetry{breaking bifurcation at \u0016= 5:84\nand gives birth to the waveform of Fig. 4. However, more\ncomplex multi-vortex topological states can progressively\nbe identi\fed as well. Panel (b) of Fig. 3 represents a pair\nof vortex lines: at the linear limit such a state can be\nformulated as the linear combination j1;1;0i+ij0;2;0i,\nin line with what is known about vortex dipole bifurca-\ntions [6]. Panel (c) represents what was termed a vortex\nstar in [30], arising at the linear limit via the linear com-\nbinationj2;0;0i\u0000j0;2;0i+i[j2;0;0i\u0000j0;0;2i]. Panel\n(d) shows a generalization of the well-known 2D vor-\ntex quadrupole [31] consisting of 4 bent yet alternating\ncharged vortex lines.\n(a)\n (b)\n (c)\n(d)\n (e)\n (f)\nFIG. 3: More exotic solutions discovered by de\ration that\nemanate from the 2nd excited state. The panels show density\nisosurfaces of the states at densities 0 :30 and 0:35.\nThe solutions in Fig. 3(a)-(d) either allow for a direct\ntracing of their linear limit or have been previously iden-\nti\fed. However, de\ration allows us to go well beyond\nthese. Important examples of this arise in panels (e) and\n(f) of Fig. 3. Panel (e) consists of a vortex ring combined\nwith 2 (oppositely charged) vortex line \\handles\". Thisstate, too, can be identi\fed at the linear limit through\na more complex topologically charged combination, as\nj2;0;0i+j0;2;0i+ij1;0;1i. Such a state exhibiting a\nvortex ring with multiple vortex lines attached to it has\nnot been previously reported, to the best of our knowl-\nedge. Even more complex is the state in panel (f), which\ndoes not bear a linear analogue. This state involves 2\nvortex rings, both of which are bent; i.e. instead of hav-\ning two \\perpendicular\" vortex rings (e.g., in the xy- and\nyz-planes), it is as if the top half of the one has connected\nitself with the right half of the other and the bottom half\nof one with the left half of the other. This con\fguration\nwas discovered by de\ration at \u0016= 6 but the branch ter-\nminates by \u0016= 5:9 without ever reaching the linear limit\nof\u0016= 7=2. In other words, this appears to be a purely\nnonlinear state not derivable by some suitable combina-\ntion of linear eigenstates. We conclude the presentation\nof our numerical results for the 2nd exciting states by\npresenting the associated bifurcation diagram in Fig. 5\nwhere the total number of atoms [cf. Eq. (12)] is used\nas the diagnostic functional. The inset panel in the top-\nleft corner therein uses the atom{number{di\u000berence \u0001 N\nbetween the branches Fig. 3(a) and Fig. 4 to illustrate a\nbifurcation in the diagram.\nFIG. 4: Left: argument of the solution at \u0016= 6. This branch\nbifurcates from the state of Fig. 3(a) at \u0016= 5:84 (see Fig. 5).\nRight: density isosurfaces of the state at densities 0 :30 and\n0:35.\nWe now explore the BdG spectral stability of selected\nsolutions (the spectra of all states shown in Figs. 1 and 3\nare presented in Figs. 10 and 11 in Appendix A, respec-\ntively). In fact, some of the identi\fed waveforms are\ndynamically robust for an interval within their existence\nrange. An example of this form is the spherical shell\ndark soliton of Fig. 1(f) with its spectrum presented in\nFig. 6(a). However, most are indeed dynamically unsta-\nble, as expected; see, e.g., the case of the vortex star\nin Fig. 6(b). Interestingly, our BdG computations re-\nveal that it is not the case that the most complex states\nare also the most unstable ones (see Fig. 6). An exam-\nple of this type can be found in the vortex ring-double\nvortex line state of Fig. 3(e) with spectrum presented in\nFig. 6(c). While the solution is highly complex, it only\nbears a single unstable mode for a wide parametric inter-\nval, and at most bears two over the interval studied. Even\nmore importantly, in our dimensionless units (scaled by\nthe harmonic trapping frequency), the relevant growth\nis typically of the order of 0.1-0.2. This means that the\n5\n3.5 4 4.5 5 5.5 6050100150200250\n3(e)3(f)\n3(d)3(c)1(f)\n3(a)\n1(e)3(b)1(d)\n1(c)1(a)\n1(b)\nµN5.85.85 5.95.95 600.511.5\nµ∆N4\n3(a)\nFIG. 5: Bifurcation diagram of the solutions emanating from the 2nd excited state at \u0016= 7=2. The labels indicate the solutions\nrepresented in the di\u000berent panels of Figs. 1 and 3. The main panel corresponds to the total number of atoms Nas a function\nof\u0016, while the top-left inset shows the atom number di\u000berence \u0001 Nbetween the branch Fig. 3(a) and Fig. 4, colored in red.\nThe green branch illustrates the solution displayed in Fig. 3(f) and its terminal point at \u0016= 5:9.\ncharacteristic dimensionless growth time is about 5-10\noscillation times, and if the perturbation is su\u000eciently\nsmall the instability manifestation time can be expected\nto be 2-3 times larger, in line with the dynamical obser-\nvations to which we now turn.\nWe employ the Crank{Nicolson method to explore the\ndynamical implications of these instabilities. The initial\nstates are prepared according to Eq. (10). In the case\nof Fig. 3(e), we observe in the snapshots of the evolu-\ntion of Fig. 7 that the vortex ring and two vortex lines\nbreak up into two vortex lines which are strongly bent\n(in fact, they are somewhat reminiscent of the U-shaped\nvortex lines of [32]). After about 20-30 dimensionless\ntime units | in line with our eigenvalue predictions |\nthe con\fguration is characterized by splittings and re-\nconnections (including ones re-formulating the original\ncon\fguration). We have performed similar computations\nfor other complex, topological states such as that shown\nin Fig. 3(f), showcasing in some such cases more radical\ndynamical breakups, i.e., the eventual persistence of a\nsingle, strongly excited vortex line; see relevant snapshots\nin Fig. 8 and the movies in the Supplemental Material at\n[URL will be inserted by publisher]. Importantly, we note\nthat the relevant time scales for standard choices of the\ntrap strength are on the order of hundreds of millisec-\nonds, and hence the con\fgurations are expected to be\nR(!)\n\u0016I(!)(a)\n\u0016(b)\n\u0016(c)\nFIG. 6: Spectra of (a) the spherical shell dark solitary\nwave presented in Fig. 1(f); (b) the vortex star presented in\nFig. 3(c); (c) the vortex ring with two vortex lines presented\nin Fig. 3(e). The real and imaginary parts of the correspond-\ning eigenfrequencies !are depicted in the top and bottom\npanels respectively.\nwell within windows of experimental accessibility. Fur-\nthermore, even when the con\fgurations become unsta-\nble, as in the dynamics of Fig. 7, they appear to result\nin oscillatory dynamics reconstructing the relevant states\n6\nt= 15\n t= 19\n t= 23\nt= 27\n t= 31\n t= 35\nFIG. 7: Snapshots of the vortex ring-double vortex line state\nobtained by solving the time-dependent NLS equation. The\nsteady-state solution of panel (e) in Fig. 3 is initially per-\nturbed along its dominant unstable eigenmode.\nt= 10\n t= 15\n t= 18\nt= 28\n t= 34\n t= 40\nFIG. 8: Snapshots of the two vortex rings state of Fig. 3(f),\nobtained by solving the time-dependent NLS equation with\nthe modi\fed Crank-Nicolson time-stepping scheme.\nin a nearly periodic fashion for far longer times (rather\nthan dispersing or yielding chaotic dynamics). This fur-\nther enhances the potential observability window of the\nstates of interest.\nWe close our presentation of the numerical results of\nde\ration by o\u000bering a glimpse into the capabilities of the\nmethod for discovering higher excited states, i.e., ones\nthat are initiated not from the 2nd (as up to now), but\nrather from the 3rd and 4th excited states at \u0016= 9=2 and\n\u0016= 11=2, respectively. Some of the relevant nonlinear\nstates discovered with de\ration can be found in Fig. 9.\nThe \frst examples, such as those of panels (a) and (b)\ncan be identi\fed straightforwardly: panel (a) represents\na nonlinear con\fguration bearing 9 vortex lines in a gen-\neralization of the star-shaped con\fguration of Fig. 3(c)\nand [30]. The state of Fig. 9(b) appears to be a con-\n\fguration bearing two perpendicular vortex rings (now\njoined{cf. panel Fig. 3(f)), along with 5 vortex lines: 4 of\n(a)\n (b)\n (c)\n(67)\n (73)\n (75)\n (d)\n (e)\n (f)\nFIG. 9: Some of the solutions discovered by de\ration that\nemanate from the 3rd (panels (a)-(c)) and 4th (panels (d)-\n(f)) excited states at \u0016= 9=2 and\u0016= 11=2. The \frst and\nthird rows show the phase of the solutions while the second\nand fourth rows represent the density isosurfaces at densities\n0:30 and 0:35.\nthese are tangent, similar to the ones of Fig. 3(e), while\none is piercing through the planes of the two rings. Going\nbeyond these, however, the states become highly com-\nplex. Fig. 9(c) shows what appears to be a combination\nof an S-shaped and 2 U-shaped vortex lines (in the ter-\nminology of [32]) along with a clearly discernible vortex\nring. Labyrinthine patterns of conjoined vortex rings and\nvortex lines appear; at the moment we do not have an\nimmediate classi\fcation. Fig. 9(d) displays an apparent\nlattice of vortex rings, while panel Fig. 9(e) is reminiscent\nof the vortex ring cages that appear in the dynamical in-\nstabilities of other states such as the spherical dark shell\nsolitary wave of Fig. 1(f) [27]. Fig. 9(f) displays a con-\nglomeration of bent vortex lines. Once again, all of these\nsolutions have not been previously identi\fed as station-\nary states of the 3D NLS/GP model, to the best of our\nknowledge.\nIV. CONCLUSIONS & FUTURE CHALLENGES\nDe\ration reveals unknown and intriguing dynamical\nstates of a fundamental model for 3D Bose-Einstein con-\n7\ndensates. By building a priori knowledge of the linear\neigenstates into the de\ration procedure, we are able to\nidentify a wide range of solutions. Many of the solu-\ntions found can be characterized using these underlying\nlinear limits. However, de\ration can also discover numer-\nous unexpected topological nonlinear states such as the\nvortex ring with 2 vortex lines, or the coupled bent vor-\ntex rings of Fig. 3. Despite their complexity, such states\nmay only be weakly unstable (thus potentially tractable\nexperimentally) and feature long-time dynamics consist-\ning of splittings and recombinations towards the origi-\nnal state. Recent experimental advances have enabled\nthe formulation (painting) of arbitrary potentials [33],\nthe establishment of arbitrary density [34] or imposition\nof controlled phase [35{37] patterns, and even the real-\nization of unstable (but su\u000eciently long-lived) complex\ntopological states such as vortex knots [38]. In light of all\nthese developments and their impact on vortex ring and\nline dynamics [39{41], we expect that the states identi\fed\nin this work to be within reach of current state-of-the-art\nexperimental e\u000borts.\nAs the nonlinearity of the model is increased, so is\nthe complexity of the available topological states; yet the\nnumerical methods discussed here appear to remain e\u000e-\ncient in this regime. They reveal not only solutions that\nare generalizations of previous ones, but also vortex ring\nlattices, cages, bent-connected-multivortex ring and line\npatterns, and more. These warrant further study, topo-\nlogical classi\fcation and deeper physical understanding.\nWe believe that this technique paves the way for a wide\nrange of future exciting explorations in this and related\n\felds.\nAcknowledgements\nThis work is supported by the EPSRC Centre\nFor Doctoral Training in Industrially Focused Math-\nematical Modelling (EP/L015803/1) in collaboration\nwith Simula Research Laboratory (NB), EPSRC grants\nEP/R029423/1 and EP/V001493/1 (PEF), and by theU.S. National Science Foundation under Grant no. PHY-\n1602994 (PGK).\nAppendix A: BdG spectral decomposition\nWe decompose the steady-state solution \u001e0, eigenvec-\ntor (a;b)>, and eigenvalue \u001ainto real and imaginary com-\nponents as \u001e0=\u001e0\nr+i\u001e0\nc,a=ar+iac,b=br+ibc, and\n\u001a=\u001ar+i\u001ac, respectively, and rewrite (7) as\n0\nB@A11 0B1\u0000B2\n0A11B2B1\n\u0000B1\u0000B2\u0000A11 0\nB2\u0000B10\u0000A111\nCA0\nB@ar\nac\nbr\nbc1\nCA=\n0\nB@\u001ar\u0000\u001ac0 0\n\u001ac\u001ar0 0\n0 0\u001ar\u0000\u001ac\n0 0\u001ac\u001ar1\nCA0\nB@ar\nac\nbr\nbc1\nCA;(A1)\nwhereB1= (\u001e0\nr)2\u0000(\u001e0\nc)2andB2= 2\u001e0\nr\u001e0\nc. The eigen-\nvalues of the matrix on the right-hand side of (A1) are\n\u001ar\u0006i\u001ac(with multiplicity two). Therefore, solving a\nreal eigenvalue problem with the left-hand matrix of (A1)\nyields the same eigenvalues and eigenvectors as the com-\nplex eigenvalue problem (7). We use a Krylov{Schur\nalgorithm with a shift-and-invert spectral transforma-\ntion [23], implemented in the SLEPc library [24], to solve\nthe following eigenvalue problem:\n0\nB@A11 0B1\u0000B2\n0A11B2B1\n\u0000B1\u0000B2\u0000A11 0\nB2\u0000B10\u0000A111\nCA0\nB@ar\nac\nbr\nbc1\nCA=\u001a0\nB@ar\nac\nbr\nbc1\nCA;(A2)\nwhere the matrices are real and \u001a=\u001ar\u0006i\u001acis complex.\nThis problem is discretized with the same piecewise cubic\n\fnite element method used to \fnd multiple solutions with\nde\ration. The spectra of the states of Figs. 1 and 3 are\nshown in Figs. 10 and 11, respectively.\n[1] M. Ablowitz, B. Prinari, and A. Trubatch, Discrete and\nContinuous Nonlinear Schr odinger Systems (Cambridge\nUniversity Press, Cambridge, 2004).\n[2] M. Ablowitz, Nonlinear Dispersive Waves (Cambridge\nUniversity Press, Cambridge, 2011).\n[3] C. Sulem and P. Sulem, The nonlinear Schr odinger equa-\ntion: self-focusing and wave collapse (Springer, New\nYork, 1999).\n[4] C. J. Pethick and H. Smith, Bose{Einstein Condensa-\ntion in Dilute Gases (Cambridge University Press, Cam-\nbridge, United Kingdom, 2002).\n[5] S. Stringari and L. Pitaevskii, Bose{Einstein Condensa-\ntion(Oxford University Press, Oxford, United Kingdom,\n2003).[6] P. Kevrekidis, D. Frantzeskakis, and R. Carretero-\nGonzlez, The Defocusing Nonlinear Schrodinger Equa-\ntion (Society for Industrial and Applied Mathematics,\nPhiladelphia, PA, 2015).\n[7] A. Hasegawa and Y. Kodama, Solitons in Optical Com-\nmunications (Clarendon Press, Oxford, 1995).\n[8] M. Kono and M. \u0014Skori\u0013 c, Nonlinear Physics of Plasmas\n(Springer Verlag, Heidelberg, 2010).\n[9] C. Kharif, E. Pelinovsky, and A. Slunyaev, Rogue Waves\nin the Ocean (Springer Verlag, Berlin, 2009).\n[10] D. J. Frantzeskakis, J. Phys. A-Math. Theor. 43, 213001\n(2010), ISSN 1751-8121, URL http://iopscience.iop.\norg/1751-8121/43/21/213001 .\n[11] F. Abdullaev, A. Gammal, A. Kamchatnov, and\n8\nR(!)\n\u0016(a)I(!)\n\u0016(b)\n\u0016(c)\nR(!)\n\u0016(d)I(!)\n\u0016(e)\n\u0016(f)\nFIG. 10: Spectra of the solutions presented in Fig. 1 (e.g.,\nthe panel (a) corresponds to the branch 1(a), illustrated in\nthe panel (a) of Fig. 1). The real and imaginary parts of the\ncorresponding eigenfrequencies !are respectively depicted in\nthe top and bottom panels.\nL. Tomio, Int. J. Mod. Phys. B 19, 3415 (2005), URL\nhttps://doi.org/10.1142/S0217979205032279 .\n[12] A. L. Fetter and A. A. Svidzinsky, J. Phys.-Condens.\nMat. 13, R135 (2001), URL https://iopscience.iop.\norg/article/10.1088/0953-8984/13/12/201 .\n[13] A. L. Fetter, Rev. Mod. Phys. 81, 647 (2009), URL\nhttps://link.aps.org/doi/10.1103/RevModPhys.81.\n647.\n[14] S. Komineas, Eur. Phys. J.-Spec. Top. 147, 133\n(2007), ISSN 1951-6401, URL https://doi.org/10.\n1140/epjst/e2007-00206-8 .\n[15] A. M. Mateo and J. Brand, Phys. Rev. Lett. 113,\n255302 (2014), URL https://link.aps.org/doi/10.\n1103/PhysRevLett.113.255302 .\n[16] J. Ruostekoski and J. R. Anglin, Phys. Rev. Lett.\n86, 3934 (2001), URL https://link.aps.org/doi/10.\n1103/PhysRevLett.86.3934 .\n[17] J. Ruostekoski and J. R. Anglin, Phys. Rev. Lett.\n91, 190402 (2003), URL https://link.aps.org/doi/\n10.1103/PhysRevLett.91.190402 .\n[18] D. Kleckner and W. Irvine, Nat. Phys. 9, 253 (2013),\nURL https://doi.org/10.1038/nphys2560 .\n[19] P. E. Farrell, A. Birkisson, and S. W. Funke, SIAM J.\nSci. Comput. 37, A2026 (2015), URL https://doi.org/\n10.1137/140984798 .\n[20] E. G. Charalampidis, P. G. Kevrekidis, and P. E. Farrell,\nR(!)\n\u0016(a)I(!)\n\u0016(b)\n\u0016(c)\nR(!)\n\u0016(d)I(!)\n\u0016(e)\nFIG. 11: Spectra of the solutions presented in Fig. 3(a)-(e).\nCommun. Nonlinear Sci. Numer. Simulat. 54, 482 (2018),\nURL https://doi.org/10.1016/j.cnsns.2017.05.024 .\n[21] E. Charalampidis, N. Boull\u0013 e, P. Farrell, and\nP. Kevrekidis, Commun. Nonlinear Sci. Nu-\nmer. Simulat. 87, 105255 (2020), URL https:\n//doi.org/10.1016/j.cnsns.2020.105255 .\n[22] F. Rathgeber, D. A. Ham, L. Mitchell, M. Lange, F. Lu-\nporini, A. T. T. Mcrae, G.-T. Bercea, G. R. Markall, and\nP. H. J. Kelly, ACM Trans. Math. Softw. 43, 1 (2016),\nURL https://doi.org/10.1145/2998441 .\n[23] G. W. Stewart, SIAM J. Matrix Anal. A. 23, 601 (2002),\nURL https://doi.org/10.1137/S0895479800371529 .\n[24] V. Hernandez, J. E. Roman, and V. Vidal, ACM Trans.\nMath. Softw. 31, 351 (2005), URL https://doi.org/10.\n1145/1089014.1089019 .\n[25] M. Delfour, M. Fortin, and G. Payre, J. Comput.\nPhys. 44, 277 (1981), URL https://doi.org/10.1016/\n0021-9991(81)90052-8 .\n[26] W. Wang, P. G. Kevrekidis, and E. Babaev, Phys. Rev. A\n100, 053621 (2019), URL https://link.aps.org/doi/\n10.1103/PhysRevA.100.053621 .\n[27] W. Wang, P. G. Kevrekidis, R. Carretero-Gonz\u0013 alez, and\nD. J. Frantzeskakis, Phys. Rev. A 93, 023630 (2016),\nURL https://link.aps.org/doi/10.1103/PhysRevA.\n93.023630 .\n[28] A. M. Mateo and J. Brand, New J. Phys. 17, 125013\n(2015), URL https://iopscience.iop.org/article/\n10.1088/1367-2630/17/12/125013 .\n[29] R. Seydel, Practical Bifurcation and Stability Anal-\nysis, vol. 5 of Interdisciplinary Applied Mathematics\n9\n(Springer, 2010), 3rd ed.\n[30] L.-C. Crasovan, V. M. P\u0013 erez-Garc\u0013 \u0010a, I. Danaila,\nD. Mihalache, and L. Torner, Phys. Rev. A 70,\n033605 (2004), URL https://link.aps.org/doi/10.\n1103/PhysRevA.70.033605 .\n[31] M. M ott onen, S. M. M. Virtanen, T. Isoshima, and M. M.\nSalomaa, Phys. Rev. A 71, 033626 (2005), URL https:\n//link.aps.org/doi/10.1103/PhysRevA.71.033626 .\n[32] A. Aftalion and I. Danaila, Phys. Rev. A 68,\n023603 (2003), URL https://link.aps.org/doi/10.\n1103/PhysRevA.68.023603 .\n[33] K. Henderson, C. Ryu, C. MacCormick, and M. G.\nBoshier, New J. Phys. 11, 043030 (2009), URL https:\n//doi.org/10.1088%2F1367-2630%2F11%2F4%2F043030 .\n[34] G. Gauthier, I. Lenton, N. M. Parry, M. Baker, M. J.\nDavis, H. Rubinsztein-Dunlop, and T. W. Neely, Optica\n3, 1136 (2016), URL http://www.osapublishing.org/\noptica/abstract.cfm?URI=optica-3-10-1136 .\n[35] M. J. H. Ku, W. Ji, B. Mukherjee, E. Guardado-Sanchez,\nL. W. Cheuk, T. Yefsah, and M. W. Zwierlein, Phys. Rev.\nLett. 113, 065301 (2014), URL https://link.aps.org/\ndoi/10.1103/PhysRevLett.113.065301 .[36] A. R. Fritsch, M. Lu, G. H. Reid, A. M. Pi~ neiro, and I. B.\nSpielman, Phys. Rev. A 101, 053629 (2020), URL https:\n//link.aps.org/doi/10.1103/PhysRevA.101.053629 .\n[37] S. Lannig, C.-M. Schmied, M. Pr ufer, P. Kunkel,\nR. Strohmaier, H. Strobel, T. Gasenzer, P. Kevrekidis,\nand M. Oberthaler, arXiv preprint arXiv:2005.13278\n(2020), URL https://arxiv.org/abs/2005.13278 .\n[38] D. S. Hall, M. W. Ray, K. Tiurev, E. Ruokokoski, A. H.\nGheorghe, and M. M ott onen, Nat. Phys. 12, 478 (2016),\nURL https://doi.org/10.1038/nphys3624 .\n[39] M. J. H. Ku, B. Mukherjee, T. Yefsah, and M. W. Zwier-\nlein, Phys. Rev. Lett. 116, 045304 (2016), URL https://\nlink.aps.org/doi/10.1103/PhysRevLett.116.045304 .\n[40] S. Sera\fni, M. Barbiero, M. Debortoli, S. Donadello,\nF. Larcher, F. Dalfovo, G. Lamporesi, and G. Ferrari,\nPhys. Rev. Lett. 115, 170402 (2015), URL https://\nlink.aps.org/doi/10.1103/PhysRevLett.115.170402 .\n[41] S. Sera\fni, L. Galantucci, E. Iseni, T. Bienaim\u0013 e, R. N.\nBisset, C. F. Barenghi, F. Dalfovo, G. Lamporesi, and\nG. Ferrari, Phys. Rev. X 7, 021031 (2017), URL https:\n//link.aps.org/doi/10.1103/PhysRevX.7.021031 .",
"main_paper_content": null
}
|
{
"decision": "Accept",
"reviews": []
}
| 2
| 0
|
{"id":"11796","year":2006,"venue":"Americas Conference on Information Systems","pdf_link":"https://b(...TRUNCATED)
| {"title":"The Role of the Board of Directors in IT Governance: A Review and Agenda for Research","au(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 13
| 0
|
{"id":"13070","year":2008,"venue":"Joint Conference on Knowledge-Based Software Engineering","pdf_li(...TRUNCATED)
| {"title":"Knowledge Engineering for Affective Bi-Modal Interaction in Mobile Devices","authors":["Ef(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 2
| 0
|
{"id":"15548","year":2008,"venue":"Language Resources and Evaluation","pdf_link":"https://www.usenix(...TRUNCATED)
| {"title":"Cleaneval: a Competition for Cleaning Web Pages","authors":["Marco Baroni","Francis Chantr(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 66
| 3
|
{"id":"16022","year":2007,"venue":"European Society for Fuzzy Logic and Technology Conference","pdf_(...TRUNCATED)
| {"title":"Algorithms for the Construction of Digital Convex Fuzzy Hulls.","authors":["Nebojsa M. Ral(...TRUNCATED)
|
{
"decision": "Reject",
"reviews": []
}
| 0
| 0
|
{"id":"20175","year":2012,"venue":"International Symposium on Algorithms and Computation","pdf_link"(...TRUNCATED)
| {"title":"An 8/3 Lower Bound for Online Dynamic Bin Packing","authors":["Prudence W. H. Wong","Fenco(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 6
| 0
|
{"id":"21805","year":2011,"venue":"The Journal of Logic and Algebraic Programming","pdf_link":"http:(...TRUNCATED)
| {"title":"On system rollback and totalized fields: An algebraic approach to system change","authors"(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 6
| 0
|
{"id":"24781","year":2007,"venue":"Innovations in Systems and Software Engineering","pdf_link":"http(...TRUNCATED)
| {"title":"Setting up an Effective Information Security Awareness Programme","authors":["Dirk De Maey(...TRUNCATED)
|
{
"decision": "Accept",
"reviews": []
}
| 10
| 0
|
{"id":"29841","year":2013,"venue":"Intelligent Information Systems","pdf_link":"https://drops.dagstu(...TRUNCATED)
| {"title":"Automatic Detection of Annotation Errors in Polish-Language Corpora","authors":["Łukasz K(...TRUNCATED)
|
{
"decision": "Reject",
"reviews": []
}
| 0
| 0
|
End of preview.
No dataset card yet
- Downloads last month
- -