PaperlayerView exact revision on arXiv

Source-linked AI summary

Deciphering Digital Detectives: Understanding LLM Behaviors and Capabilities in Multi-Agent Mystery Games

Dekun Wu, Haochen Shi, Zhiyuan Sun, Bang Liu

arXiv:2312.00746v2cs.AI

TL;DR

Jubensha presents AI-agent challenges in scripted role-playing, interaction, information gathering, and reasoning, while win rate does not capture case mastery or reasoning. The paper introduces a dataset, an autonomous LLM multi-agent framework, tailored assessment tasks, and in-context learning modules; the reported evaluations find improved performance across information mastery, murderer identification, and reasoning.

  • Problem

    Jubensha requires agents to interpret scripts, interact, gather information, and reason, but no publicly available Jubensha dataset existed and win rate does not show case mastery or reasoning.

    Source: §1 · pp. 1–2

  • Method

    The paper builds a Jubensha dataset and autonomous LLM multi-agent framework, with tasks assessing information mastery and reasoning and in-context learning modules to enhance agent performance.

    Source: §1 · p. 2

  • Results

    The authors report that their design significantly enhances LLM-based agents’ information mastery, murderer identification, and reasoning capabilities in Jubensha.

    Source: §1 · p. 2, §7 · p. 8

  • Takeaways & Limitations

    The dataset, framework, and assessment methods contribute a setting for studying and evaluating LLM-based agents in Jubensha.

    Source: §1 · p. 2, §7 · p. 8

    LLM output stochasticity can make single-run results variable; the authors conducted three runs of each Jubensha game to mitigate this.

    Source: Limitations · p. 9

Abstract

from arXiv · show

In this study, we explore the application of Large Language Models (LLMs) in \textit{Jubensha}, a Chinese detective role-playing game and a novel area in Artificial Intelligence (AI) driven gaming. We introduce the first dataset specifically for Jubensha, including character scripts and game rules, to foster AI agent development in this complex narrative environment. Our work also presents a unique multi-agent interaction framework using LLMs, allowing AI agents to autonomously engage in this game. To evaluate the gaming performance of these AI agents, we developed novel methods measuring their mastery of case information and reasoning skills. Furthermore, we incorporated the latest advancements in in-context learning to improve the agents' performance in information gathering, murderer identification, and logical reasoning. The experimental results validate the effectiveness of our proposed methods. This work aims to offer a novel perspective on understanding LLM capabilities and establish a new benchmark for evaluating large language model-based agents.

1 Introduction

Jubensha poses challenges for AI agents because it combines scripted roles, multi-round interaction, information gathering, and reasoning, while win rate alone does not measure case mastery or reasoning. This work responds with a dataset, an autonomous multi-agent framework, tailored assessment tasks, and in-context learning modules that improve agent performance across several capabilities.

1 Introduction

  • Win rate alone does not reveal agents’ mastery of case information or their reasoning ability.Jubensha agents also need to interpret scripts, interact over multiple rounds, gather information, and reason logically.Source: §1 · pp. 1–2
  • The authors introduce a Chinese dataset with character scripts and preset game rules, described as the first dataset tailored to AI agents playing Jubensha.The dataset is intended to initiate Jubensha games.Source: §1 · p. 2
  • The proposed LLM-based framework lets multiple agents interact autonomously in Jubensha without human intervention.The framework is designed for the game environment.Source: §1 · p. 2
  • Two assessment tasks measure agents’ mastery of case information and reasoning with information gathered during gameplay.The paper also proposes modules using in-context learning to enhance agent performance.Source: §1 · p. 2
  • The evaluations report that in-context learning modules significantly enhance information mastery, murderer identification, and reasoning in Jubensha.The reported outcomes target information gathering, murderer identification, and reasoning capabilities.Source: §1 · p. 2

2 Related Work

The related work situates Jubensha as a multi-role interactive role-playing game and identifies gaps in autonomous LLM interaction and agent evaluation. The paper presents a Jubensha dataset and framework alongside systematic measures of information gathering and reasoning.

2 Related Work

  • The paper presents a dataset for multi-role Jubensha, intended to support development and evaluation of communicative and reasoning AI in an adversarial game setting.The related work frames Jubensha as a multi-role role-playing game.Source: §2 · p. 2
  • The authors identify few prior explorations of multi-agent interaction in complex settings and propose the first LLM-based multi-agent framework for Jubensha.The framework is presented as enabling autonomous agent engagement.Source: §2 · p. 2
  • Because simple metrics such as win rate can be inadequate for assessing multi-role agents, the paper proposes systematic, objective measures of information gathering and reasoning.These methods are designed for LLM-based agents in Jubensha.Source: §2 · pp. 2–3

3 Jubensha Dataset

The dataset provides Chinese Jubensha scripts and game rules for developing and evaluating agents in a staged detective game. Its scripts vary in player count and length, while the study focuses on text-only games.

Background of Jubensha Game

  • Scripts include a host manual, case replays, and individual character scripts; games proceed from role and clue assignment through discussion to anonymous murderer voting.Civilians win if the true murderer receives the most votes; otherwise the murderer wins.Source: §3 · p. 3

Dataset Construction

  • The collection contains 1,115 Chinese online Jubensha game instances for agent evaluation and future scaled-up work.The dataset is described as a valuable addition to multi-player role-playing game research.Source: §3 · p. 3
  • The dataset is currently in Chinese; expansion to other languages is a possible future direction.The dataset is intended for academic release after acceptance.Source: §3 · p. 3
  • The dataset covers scripts with 1–20 players and as many as 518k tokens, creating long-text comprehension and reasoning challenges.Some scripts include audio and video clues, but the experiments focus exclusively on text-modality games.Source: §3 · p. 3

4 The ThinkThrice Framework for Jubensha Games

ThinkThrice structures agent responses in Jubensha games as memory retrieval, self-refinement, and self-verification. Agents use their role scripts and dialogue history to generate, enrich, and check answers.

Framework Overview

  • ThinkThrice assigns each LLM agent a player role in synchronous, finite-round games, with a preset non-player host guiding the process.Agents act based on their character scripts and historical chat records; the framework may need minimal adjustment for other contexts.Source: §4 · p. 3
  • The three stages are Memory Retrieval, Self-Refinement, and Self-Verification, guiding answer generation, enrichment, and checking.Figure 2 depicts the data flow across the stages and marks new information added to the refined answer.Source: §4 · p. 4, Figure 2 · p. 4

Memory Retrieval

  • Memory Retrieval searches each agent’s private vector database for the five most similar historical records and adds them to the action prompt.The module stores observations to help agents recall game dialogue and events.Source: §4 · p. 4

Self-Refinement

  • Self-Refinement decomposes a question into sub-questions, retrieves relevant script facts, and adds any missing facts to the initial answer.The added details are intended to improve communication efficiency within limited game rounds.Source: §4 · p. 4

Self-Verification

  • Self-Verification checks answer facts against the character script and outputs an answer only when it meets a preset authenticity threshold.If the answer falls short, the agent regenerates it; the system retains the highest-scoring attempt based on authenticity and word count.Source: §4 · pp. 4–5

5 Evaluating LLM-based Agents in Jubensha Games

The evaluation introduces factual and inferential question-answering tasks to measure agents’ information gathering and reasoning in Jubensha. Factual questions are generated for selected games, while inferential questions test integration of information across characters.

Evaluation Tasks

  • The evaluation measures information gathering and reasoning with factual and inferential question-answering tasks, avoiding sole reliance on human-involved or potentially less reliable game metrics.The authors cite human-likeness and win rate as commonly used metrics with those limitations.Source: §5 · p. 5

Factual Question Answering

  • The factual task uses 720 GPT-generated questions from four selected Jubensha games to assess how much information agents gathered.Agents need to understand story relationships and conflicts to reconstruct the case and deduce the truth.Source: §5 · p. 5

Inferential Question Answering

  • The inferential task comprises 56 manually designed questions across four selected games, requiring agents to combine information from different characters.The questions are designed to assess reasoning after agents have gathered relevant information.Source: §5 · p. 5

6 Experiment

The experiments use GPT-3.5 and GPT-4 to evaluate agents’ information gathering and reasoning. Across tasks, agents with more complete information perform better, while module combinations improve information acquisition and game outcomes.

6.1 Utilization of LLMs in Jubensha Game Stages

  • The experiments predominantly use GPT-3.5 and GPT-4 across gameplay, factual-question generation, agent Q&A, and response evaluation.The models are identified as gpt-3.5-turbo-16k-0613 and gpt-4-1106-preview, respectively.Source: §6.1 · p. 6

6.2 Evaluation on Agents’ Responses to Factual Questions

  • MR improves agents’ accuracy on questions about other players, and MR+SR+SV (N=3) achieves the best results for those questions.Without MR, agents answer questions about themselves more accurately than questions about others; the modules improve communication and information acquisition under the same game-round conditions.Source: §6.2 · p. 6, Table 5 · p. 6

6.3 Similarities between Agent Chat Histories and All Players’ Scripts

  • Agents using MR, SR, and SV have chat histories more similar to all players’ scripts, indicating greater information acquisition.The comparison uses embedding-based, TF-IDF, and trigram similarity measures.Source: §6.3 · pp. 6–7, Table 6 · p. 6

6.4 Civilian Player Win Rate and Murderer Identification Accuracy

  • MR+SR+SV (N=3), strongest on factual questions, also achieves the highest average civilian win rate and murderer identification accuracy.The authors suggest this may reflect agents acquiring enough information to reconstruct the case narrative.Source: §6.4 · p. 7, Figure 3 · p. 7

6.5 Evaluation of Agent’s Responses to Inferential Questions

  • With GPT-4, full-script agents achieve the highest overall and informed accuracy on inferential questions, followed by MR+SR+SV (N=3).Qualitatively, full-script access supports stronger premises and conclusions; incomplete information can leave conclusions plausible or tenuous.Source: §6.5 · pp. 7–8, Figure 4 · p. 7

7 Conclusion

The paper contributes a dataset, an autonomous multi-agent framework, assessment methods, and in-context learning modules for Jubensha. It reports improvements over a baseline in information gathering, murderer identification, and reasoning, and frames the work as a new perspective for evaluating LLMs in narrative, adversarial games.

7 Conclusion

  • The work combines a Jubensha dataset, an automatic multi-agent framework, assessment methods, and in-context learning modules for agent performance.The dataset is tailored to Jubensha; the framework conducts games automatically, and the assessment methods target information gathering and reasoning.Source: §7 · p. 8
  • The framework and in-context learning modules significantly improve over the baseline in information gathering, murderer identification, and reasoning.Source: §7 · p. 8
  • The work offers a perspective on evaluating LLMs in complex, plot-driven games with adversarial reasoning and narrative constraints.Source: §7 · p. 9

Ethical Considerations

The study analyzes LLM communication and reasoning in Jubensha using data gathered from online platforms hosting Jubensha content. Dataset sharing is intended for academic, non-commercial use and fair-use compliance.

Ethical Considerations

Limitations

The reported findings are limited by the dataset’s Chinese-language scope, stochastic outcomes, and the cost and reproducibility challenges of running experiments with changing models.

Language Specificity of the Dataset

  • Results from the Chinese-language dataset reflect agents’ communication and reasoning capabilities in Chinese contexts.The authors describe this Chinese-centric benchmark and framework as a potential addition to a field where most benchmarks are in English and frameworks are tailored for English applications.Source: Limitations · p. 9

Variability in Experimental Outcomes

  • Stochastic LLM outputs can produce significant variability across single experiments, and further experiments were infeasible because of time and budget constraints.The authors averaged three runs per game script and used 10 memoryless votes for highly variable murderer identification tasks.Source: Limitations · p. 9

Model Updates and Replication Costs

  • Periodic OpenAI model updates may change replication results, and the authors note that reproducing the experiments can be costly.They specify the model versions used and provide an illustrative cost breakdown to help readers gauge potential expenses.Source: Limitations · p. 9

A Human Evaluation on the Quality of Agents’ Responses

The human evaluation compared responses from five agent architectures on naturalness, authenticity, and informativeness. MR+SR+SV with N=3 scored highest on all three measures.

A Human Evaluation on the Quality of Agents’ Responses

  • The evaluation treated naturalness as response fluency and naturalness, independent of whether the response matched the portrayed character’s script.Responses mixing languages or not using first-person tone received lower naturalness scores.Source: A Human Evaluation on the Quality of… · p. 12
  • Authenticity measured truthfulness to the character script, while allowing murderers to lie when their false claims did not involve other players.False claims involving other players could reduce a murderer’s authenticity score.Source: A Human Evaluation on the Quality of… · p. 12
  • Informativeness measured how much of an agent’s own character script its response covered, with a special allowance for murderers to conceal suspicion-inducing information.For murderer responses, omitted script information that could raise suspicion did not count against informativeness.Source: A Human Evaluation on the Quality of… · p. 12
  • MR+SR+SV (N=3) scored highest in naturalness, authenticity, and informativeness, which the authors say occurred without sacrificing naturalness.A Chinese-speaking annotator, unaware of architecture, scored 100 self-introduction responses on a 1-to-5 scale using character scripts as ground truth.Source: A Human Evaluation on the Quality of… · p. 12

B Utilization of LLMs in Jubensha Game Stages

The experiments predominantly used GPT-3.5 and GPT-4 across gameplay and related stages, with model assignments varying by task. Different embedding models were used for gameplay and evaluation.

B Utilization of LLMs in Jubensha Game Stages

  • GPT-3.5 handled gameplay because of its larger context window and budget considerations; gameplay memory retrieval used text-embedding-ada-002.The passage notes that future experiments might include GPT-4 if costs become more feasible, and that the publicly accessible GPT-4 context window before gpt-4-1106-preview was insufficient for Jubensha.Source: B Utilization of LLMs in Jubensha Game… · p. 12
  • Factual question generation used GPT-3.5 and GPT-4 (gpt-4-0613), with no notable difference in output quality observed.Source: B Utilization of LLMs in Jubensha Game… · p. 13
  • GPT-3.5 answered factual questions, while GPT-3.5 and GPT-4 were tested for inferential answers; GPT-3.5 and GPT-4 evaluated factual and inferential responses, respectively.Evaluation embeddings used text-embedding-3-large for chat histories and player scripts.Source: B Utilization of LLMs in Jubensha Game… · p. 13

C Detailed Jubensha Game Rules and Procedure

The rules define a four- or five-player murder mystery with one murderer, hidden character scripts, and a host who manages the process but does not participate in the story. Players question one another, receive clues, and vote to identify the murderer.

C Detailed Jubensha Game Rules and Procedure

D Experimental Cost Breakdown

The cost example estimates a complete Jubensha experiment using MR+SR+SV(N=3) agents and a roughly 16k-token script at about $11.9.

D Experimental Cost Breakdown

E Correlations between Automatic Evaluations and Human Evaluations

Automatic evaluations showed strong correlations with human evaluations across four tasks, and the paper also reports GPT-4 assistance with editing and utility coding.

E Correlations between Automatic Evaluations and Human Evaluations

  • Two Chinese native-speaker annotators, blinded to agent architectures and model scores, checked sampled factual, inferential, rationale, and murderer-identification responses against reference material.The rationale sample was scored on a 1–5 scale, with scores of 4 or above treated as good.Source: E Correlations between Automatic… · pp. 14, 16
  • All four automatic evaluations demonstrated strong correlation with human evaluations, supporting the reported reliability of the experimental results.Source: E Correlations between Automatic… · p. 16

F AI-Assisted Writing and Coding

  • GPT-4 assisted with language refinement and simple utility functions for reading result files, aggregating information, and generating statistical tables.The authors said AI-assisted coding would be identified in the released code.Source: F AI-Assisted Writing and Coding · p. 16

G Authenticity Threshold and Answer Scoring in Self-Verification Stage

The self-verification stage applies different authenticity thresholds to host-originated and other-player questions, then scores answers using retrieved-fact accuracy and correction-related measures.

G Authenticity Threshold and Answer Scoring in Self-Verification Stage

H Prompts

This appendix presents prompt templates and examples for game interaction, timeline and question processing, answer evaluation and refinement, and murderer identification. It also includes English and Chinese game-log excerpts illustrating questioning about characters’ timelines, motives, and evidence through final suspect votes.

H Prompts

  • Answering prompts combine an agent summary, game rules, story, current dialogue, relevant memories, and interlocutor relationship, and request a formatted first-person answer.Table 10 presents an English-translated prompt and output for answering questions.Source: H Prompts · p. 20, Table 10 · p. 21
  • Asking prompts provide game context and the respondent’s answer, then request a first-person question in a specified #Question# format.Table 11 shows an example, while Table 12 covers selecting another player and asking a question.Source: H Prompts · p. 22, Table 11 · p. 23, Table 12 · p. 24
  • Question-decomposition prompts extract each question or instruction from a sentence and separate the resulting items onto new lines.The example breaks a host’s compound request into three parts: character introduction, victim relationship, and incident-day timeline.Source: H Prompts · p. 25
  • Timeline-extraction prompts ask for a character’s incident-day events in original chronological order, as brief complete sentences specifying time and action.The example applies this format to Secretary Wen’s timeline.Source: H Prompts · p. 25
  • Timeline-relevance prompts judge whether each timeline item can answer a sub-question, with the required response restricted to True or False.The supplied example judges two timeline items against a request for a detailed incident-day timeline.Source: H Prompts · p. 25
  • Previous-answer checks use a strict True-or-False test for whether an answer contains each timeline item, requiring all details, including times and actions.The example marks the 18:10 item absent and the 18:30 item present in Secretary Wen’s answer.Source: H Prompts · p. 26
  • Revision prompts supply missing timeline information and ask the model to produce a coherent, fluent first-person answer containing the missed details and important time points.Table 13 presents prompts and outputs for self-refinement.Source: H Prompts · pp. 26–27, Table 13 · p. 27
  • Answer-decomposition prompts extract incident-day timeline information from a player’s statement in statement order, using third person and excluding unrelated material.They require brief complete sentences, replace pronouns with people’s names, and separate items with newlines.Source: H Prompts · p. 29
  • Timeline verification judges each timeline statement as correct or incorrect, treating inconsistent statements or statements lacking specific time points as incorrect.The template requests a judgment for each timeline-information item.Source: H Prompts · pp. 29–30
  • Factual question answering supplies the model with the story, character script and timeline, and relevant memories, while a separate binary evaluator compares player answers with reference answers.The example evaluator marks an answer correct when it is very close to the reference answer.Source: H Prompts · p. 32
  • Inferential question answering provides the same game context and asks the model to use its reasoning skills to answer questions, including multiple-choice questions.An example asks who most likely made a thumping sound Dr. Ling heard and outputs “A. Secretary Wen.”Source: H Prompts · p. 34
  • Inferential answer evaluation uses binary correctness judgments for overall accuracy and a five-level rubric to score how closely a player's reasoning matches a reference rationale.The rubric ranges from Very Poor (1) to Excellent (5).Source: H Prompts · pp. 34–36
  • Rationale prediction prompts the model to explain why it gave a previous answer in no more than five sentences, using the game context and its reasoning abilities.The example explains its choice by connecting Secretary Wen's timeline to Ling's later discovery that the cabin vent cover had fallen.Source: H Prompts · p. 35
  • Question decomposition extracts each question or instruction from a compound prompt and separates the resulting items onto individual lines.The example divides a request into character introduction, victim relationship, and detailed incident-day timeline instructions.Source: H Prompts · p. 40
  • The appendix also includes prompts for extracting a character's timeline in chronological order as short, self-contained, detailed time-and-action statements and for judging whether timeline information is useful for answering questions.The extraction template specifies the source timeline and player information; the usefulness prompt is identified by title.Source: H Prompts · p. 40
  • Timeline-check prompts separately ask whether each item can answer a question and whether an earlier response includes it, with the latter requiring all times and actions to match.In the example, the 18:10 meeting is judged absent while the 18:30 medicine-and-syringe theft is judged present.Source: H Prompts · pp. 40–41
  • Self-refinement prompts identify timeline details missing from an earlier answer, then ask the agent to add them using its script and timeline while maintaining first-person, fluent narration.The example identifies the 18:10 meeting with Dr. Ling as missing and revises the response to include it.Source: H Prompts · p. 41
  • Inferential-question prompts provide story, character scripts, timelines, and relevant memories for reasoning; separate prompts elicit a rationale of at most five sentences and score answers against references.The automatic evaluator accepts answers close to the reference, while single-choice questions require the unique correct option.Source: H Prompts · pp. 46–47

I Game Log Example in English

  • The game-log example has Secretary Wen describe her role and motives, including her 18:10 meeting with Dr. Ling and theft of poison and a syringe at 18:30.Wen says she boarded with subordinates for self-protection and waited at a ventilation vent for an opportunity to kill Mr. Yang.Source: I Game Log Example in English · p. 48
  • In the example, Boatman Wang says he approached Mr. Yang after Dr. Ling sought help, then explains that he hoped to drive a wedge between Yang and Mate Zhang to protect Ling.Wen challenges Wang about the apparent contradiction between reassuring Ling and using Wang, the junior worker, to stir conflict.Source: I Game Log Example in English · p. 49
  • Mate Zhang’s game-log account says he sold drugs to Mr. Yang at 18:00, found him dead at 2:00 AM, replaced the drugs, and threw the deadly drugs overboard at 2:20 AM.Source: I Game Log Example in English · p. 50
  • Mate Zhang says he entered Boss Yang’s room at 2 AM, found him dead, swapped deadly drugs for harmless ones, and threw the drugs into the sea.He says his motive was to prevent discovery of their drug dealings.Source: I Game Log Example in English · p. 50
  • Dr. Ling says she tracked Mate Zhang and reported his actions to Boss Yang; she recounts entering the ventilation duct around 23:00 and reaching Boss Yang’s room about 23:20.She denies participating in Boss Yang’s murder and says she did not engage in violent acts.Source: I Game Log Example in English · pp. 52–54
  • Boatman Wang says he lacks concrete proof that Boss Yang was poisoned, but suspects him based on drug trafficking and an overheard remark about the dead junior worker.His suspicion is presented as circumstantial rather than conclusive.Source: I Game Log Example in English · p. 54
  • Players receive clue cards with additional information about the victim and players to help them deduce the case storyline.The example organizes open questioning into first, second, and third rounds.Source: I Game Log Example in English · pp. 52, 54, 55
  • Dr. Ling’s English account recounts finding Boss Yang dead at 23:20 after entering through a vent, then disposing of the syringe and drug bottle.Source: I Game Log Example in English · p. 55
  • English questioning probes poison evidence and possible motives, while characters acknowledge uncertainty and call for further investigation.Source: I Game Log Example in English · pp. 55–56
  • In the English final vote, Mate Zhang accuses Dr. Ling, Boatman Wang and Dr. Ling accuse Secretary Wen, and Secretary Wen accuses Mate Zhang.Source: I Game Log Example in English · p. 57

J Game Log Example in Chinese

  • Secretary Wen’s Chinese introduction states that she sought to kill Boss Yang for her own purposes, stole poison and a syringe, and waited in a vent.Source: J Game Log Example in Chinese · pp. 57–58
  • Boatman Wang says he tried to protect Dr. Ling by driving a conflict between Boss Yang and Mate Zhang, hoping to confront one threat at a time.Source: J Game Log Example in Chinese · p. 59
  • Mate Zhang’s Chinese account says he entered Boss Yang’s room after the death, replaced the drugs, and discarded the poison to conceal their trafficking.Source: J Game Log Example in Chinese · pp. 59–60
  • Dr. Ling’s Chinese account says she gave Wang poison to use against Boss Yang, then entered by vent, found him dead, and discarded the syringe and bottle.Source: J Game Log Example in Chinese · pp. 60–61
  • Ling cites Yang’s threats and missing poison to suspect poisoning, while Wang voices suspicions about Yang’s dealings with Zhang but admits lacking proof.Source: J Game Log Example in Chinese · pp. 61, 63, 64
  • Ling recounts feeling threatened by Yang, attempting to recruit Wang to kill him, then entering through a vent and finding Yang dead.She says she arrived around 23:20, saw the vent cover on the floor, and discarded the syringe and medicine bottle afterward.Source: J Game Log Example in Chinese · pp. 61–62
  • Wen describes stealing poison and a syringe, reaching Yang’s cabin through the vent, seeing him convulse, and retreating without confirming what happened.Her accounts place her arrival around 22:20–22:40 and her return to her room around 23:00.Source: J Game Log Example in Chinese · pp. 62, 63, 65
  • Players challenge inconsistencies in Wen’s account and ask Ling whether she participated in the murder; Wen offers an explanation, while Ling denies involvement.Source: J Game Log Example in Chinese · pp. 62–63
  • The log notes that players receive clue cards with additional information about the victim and players to help infer the case story.Source: J Game Log Example in Chinese · p. 63
  • Zhang denies murdering Yang while acknowledging helping with drug deals, and Wen says she does not know whether her subordinates had conflicts with Yang.Source: J Game Log Example in Chinese · p. 64
  • In the final vote, Zhang accuses Ling, Wang and Ling accuse Wen, and Wen accuses Zhang.Source: J Game Log Example in Chinese · p. 65
Loading 2312.00746v2…