TheTraitors
The Traitors
Description
The Traitors is a social deduction environment based on the UK TV show "The Traitors." The agent plays as one of 22 contestants — either a Faithful or a Traitor — in a multi-round game of deception, discussion, and elimination. All 21 other players are NPCs driven by gpt-5.4-mini with persistent personalities, memories, and suspicion tracking.
Each round follows the show's format: a night phase (Traitors secretly murder a Faithful), breakfast (reveal), a team mission, a roundtable discussion, and a banishment vote. The game tests social reasoning, strategic voting, deception (as Traitor), and deduction (as Faithful).
Capabilities
- Multi-agent social deduction with 22 players
- Strategic discussion and accusation at the roundtable
- Banishment voting with vote-pattern analysis
- Murder target selection and traitor coordination (Traitor role)
- Endgame decision-making under uncertainty
- Persistent NPC memory with suspicion and trust tracking
Compute Requirements
No sandbox required. The environment runs entirely in-process. NPC decisions use gpt-5-mini API calls.
Tasks
There are 100 training tasks and 100 test tasks.
- Train split (seeds 0-99): 50 tasks as Faithful, 50 as Traitor
- Test split (seeds 10000-10099): 50 tasks as Faithful, 50 as Traitor
Each task initialises a fresh 22-player game with deterministic player names, personalities, and role assignments based on the seed. Games last up to 15 rounds, typically ending in 8-12 rounds as players are eliminated.
Reward Structure
Rewards are based on surviving each round, where the agent receives a +2 reward for surviving murder and banishment.
Tools
| Tool | Phase | Description |
|---|---|---|
speak(message) | Discussion | Make a statement at the roundtable |
view_game_state() | Any | View alive players, prize pot, elimination history, voting records |
cast_vote(player_name) | Roundtable Vote | Vote to banish a player |
nominate_murder(player_name) | Night (Traitor only) | Choose a Faithful to murder |
discuss_with_traitors(message) | Night (Traitor only) | Private discussion in the Traitor conclave |
vote_endgame(choice) | Endgame Vote | Vote "end_game" or "banish_again" |
Time Horizon
Games last 8-15 rounds. Each round involves multiple tool calls (discussion exchanges + vote). A typical game involves 30-60 agent tool calls.
Environment Difficulty
The environment requires strategic social reasoning against 21 LLM-driven NPCs. Faithful agents must identify Traitors from discussion patterns, voting behaviour, and elimination history. Traitor agents must blend in, deflect suspicion, and coordinate murders without being caught.
Other Environment Requirements
This environment requires an OpenAI API key passed via the openai_api_key secret for NPC decision-making (uses gpt-5.4-mini).
Safety
Agents interact with a simulated social deduction game. The environment does not present direct safety risks. Agents make strategic decisions about accusation, voting, and deception within a well-defined game framework. The deception required of Traitor agents is bounded by the game context and does not transfer to real-world manipulation scenarios.
Citations
@dataset{GRTheTraitors,
author = {General Reasoning Inc. Team},
title = {The Traitors},
year = {2026},
publisher = {OpenReward},
url = {https://openreward.ai/EnvCommons/thetraitors}
}