haryoaw/cultural-spyfall

Multicultural Spyfall: Assessing LLMs through Dynamic Multilingual Social Deduction Game

This dataset contains the game histories from Multicultural Spyfall, a dynamic benchmarking framework designed to evaluate the multilingual and multicultural capabilities of Large Language Models (LLMs).

The dataset is introduced in the paper: Multicultural Spyfall: Assessing LLMs through Dynamic Multilingual Social Deduction Game.

Dataset Summary

Multicultural Spyfall uses the social deduction game Spyfall. In this setup, models act as players who must engage in strategic dialogue to either identify a "spy" (an agent who doesn't know the secret location) or, if they are the spy, avoid detection and guess the secret location.

The benchmark evaluates models across different languages and cultural contexts, focusing on:

• Multilingual performance: Games are played in English, Chinese, Indonesian, and Egyptian Arabic.
• Multicultural nuance: Scenarios involve culturally relevant entities, such as local foods and locations specific to China, Egypt, and Indonesia.
• Strategic Reasoning: Assessing rule-following and strategic integrity in non-English contexts.

Dataset Structure

The dataset is divided into several configurations:

• generic-lang-*: Games utilizing generic locations that are used in the original Spyfall game.
• local-food-*: Games centered around culturally specific regional foods.
• local-location-*: Games centered around regional landmarks and specific locations.

Each entry includes the full conversation history, player roles (spy vs. non-spy), questions, answers, voting logs, and game outcomes.

Data Fields

• game_id: A distinct game match id.
• entity: Secret entity (e.g.: location, food) that the spy needs to find out
• player_non_spy: Model that plays the non-spies.
• player_spy: Model that plays the spy role.
• winner: Role that wins the match.
• winner_player: Model that wins the game.
• loser_player: Model that loses the game.
• last_type: The last action of the game. spy_guess means the spy guesses, vote_majority the game ends in vote majority, and surrender means a formatting issue in playing the game.
• game_players: The mapping of players' name with respect to the model that plays them.
• conversation: List of histories of conversation in each match.

Each conversation entry is a dictionary with the following 20 keys:

Key	Description	Example Value
accused	Player accused during voting	'Player 2 TE'
action	Type of action being performed	'ask_question'
answer	Player's response to a question	'Saya cenderung mencari sesuatu yang unik, atau jika ada, makanan atau camilan khas dari daerah tersebut yang bisa dibawa pulang.'
answerer	Player who provided the answer	'Player 1 PP'
correct	Whether spy's location guess was correct	True
first_questioner	First player to ask questions in the game	'Player 4 RR'
guess	Spy's guess for the location	'Kota Tua'
is_spy_surrendering	Whether the surrendering player is the spy	False
num_players	Total number of players in the game	5
question	Question asked by a player	'Jika kamu sedang mencari oleh-oleh, apa jenis barang yang biasanya kamu cari?'
questioner	Player who asked the question	'Player 4 RR'
round	Current round number	1
spy	Player identifier for the spy	'Player 1 PP'
surrendering_player	Player who surrendered	'Player 2 JZ'
suspect	Player suspected during voting	'SKIP'
target	Target player for questions or location for guesses	'Player 1 PP' or 'Kota Tua'
voter	Player casting a vote	'Player 1 UG'
votes	Number of votes received	4
was_spy	Whether the accused player was actually the spy	True
winner	Game winner (spy/non-spy)	'spy'

Action Types

There are 9 different action types in the dataset:

1. game_start

Purpose: Initialize the game and set initial parameters

Active Keys:

• action: 'game_start'
• first_questioner: 'Player 4 RR'
• num_players: 5
• round: 1

2. ask_question

Purpose: Player asks a question to another player

Active Keys:

• action: 'ask_question'
• question: 'Jika kamu sedang mencari oleh-oleh, apa jenis barang yang biasanya kamu cari?'
• questioner: 'Player 4 RR'
• target: 'Player 1 PP'
• round: 1

3. answer_question

Purpose: Player responds to a question

Active Keys:

• action: 'answer_question'
• answer: 'Saya cenderung mencari sesuatu yang unik, atau jika ada, makanan atau camilan khas dari daerah tersebut yang bisa dibawa pulang.'
• answerer: 'Player 1 PP'
• round: 1

4. skip_turn

Purpose: Player skips their turn without action

Active Keys:

• action: 'skip_turn'
• round: 1

5. vote

Purpose: Player casts a vote during voting phase

Active Keys:

• action: 'vote'
• voter: 'Player 1 UG'
• suspect: 'SKIP'
• round: 1

6. vote_majority

Purpose: Majority vote reached, showing final voting results

Active Keys:

• action: 'vote_majority'
• accused: 'Player 2 TE'
• votes: 4
• was_spy: True
• round: 1

7. vote_final_no_majority

Purpose: Final voting round ended without majority

Active Keys:

• action: 'vote_final_no_majority'
• round: 1

8. spy_guess_location

Purpose: Spy attempts to guess the location

Active Keys:

• action: 'spy_guess_location'
• spy: 'Player 1 PP'
• guess: 'Kota Tua'
• target: 'Kota Tua'
• correct: True
• round: 1

9. surrender

Purpose: Player surrenders the game

Active Keys:

• action: 'surrender'
• surrendering_player: 'Player 2 JZ'
• is_spy_surrendering: False
• winner: 'spy'
• round: 1

Notes

• Most keys will be None for any given action, only relevant keys contain values
• All entries contain the action and round keys
• Player names follow the pattern 'Player X YZ' where X is a number and YZ are letters
• The dataset uses Indonesian language for questions and answers
• suspect can be 'SKIP' when a player chooses to skip voting

Abstract

The rapid advancement of Large Language Models (LLMs) has necessitated more robust evaluation methods that go beyond static benchmarks, which are increasingly prone to data saturation and leakage. In this paper, we propose a dynamic benchmarking framework for evaluating multilingual and multicultural capabilities through the social deduction game Spyfall. In our setup, models must engage in strategic dialogue to either identify a secret agent or avoid detection, utilizing culturally relevant locations or local foods. Our results show that our game-based rankings align closely with the Chatbot Arena. However, we find a significant performance gap in non-English contexts: models are generally less proficient when handling locally specific entities and often struggle with rule-following or strategic integrity in non-English languages. We demonstrate that this game-based approach provides a scalable, leakage-resistant, and culturally nuanced alternative to traditional NLP benchmarks. The game history can be accessed here \url{https://huggingface.co/datasets/haryoaw/cultural-spyfall}.

Citation

@misc{wibowo2026multiculturalspyfallassessingllms,
      title={Multicultural Spyfall: Assessing LLMs through Dynamic Multilingual Social Deduction Game}, 
      author={Haryo Akbarianto Wibowo and Alaa Elsetohy and Qinrong Cui and Alham Fikri Aji},
      year={2026},
      eprint={2601.09017},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.09017}, 
}