KantBench: 93 Game Theory Environments for LLM Training

A comprehensive game theory environment for training and evaluating LLM strategic reasoning via OpenEnv. Supports GRPO/DPO training with the environment as a reward oracle.

Games (93)

2-Player Games (90)

Category	Examples	Count
Classic Matrix	Prisoner's Dilemma, Stag Hunt, Hawk-Dove, Battle of the Sexes	20+
Economic/Market	Cournot, Bertrand, Hotelling, Nash Demand, Double Auction	23
Information & Signaling	Beer-Quiche, Spence Signaling, Bayesian Persuasion, Moral Hazard	21
Cooperative & Repeated	Shapley Allocation, Stable Matching, Discounted PD, Stochastic PD	23
Auctions & Contests	First-Price, Vickrey, All-Pay, Colonel Blotto, Tullock Contest	10+
Sequential	Ultimatum, Trust, Centipede, Stackelberg, Dictator	6

N-Player Games (3)

Game	Players	Description
nplayer_public_goods	5	Each player contributes from an endowment; pot is multiplied and split equally
nplayer_volunteer_dilemma	5	At least one must volunteer for everyone to benefit; volunteers pay a cost
nplayer_el_farol	5	Attend a bar that's fun when uncrowded but unpleasant when full

Opponent Strategies (17)

random, always_cooperate, always_defect, tit_for_tat, tit_for_two_tats, grudger, pavlov, suspicious_tit_for_tat, generous_tit_for_tat, adaptive, mixed, ultimatum_fair, ultimatum_low, trust_fair, trust_generous, public_goods_fair, public_goods_free_rider

Reset Parameters

# Specific game and strategy
result = env.reset(game="stag_hunt", strategy="grudger")

# N-player game (strategy applies to all opponents)
result = env.reset(game="nplayer_volunteer_dilemma", strategy="random")

# Random game and strategy (default)
result = env.reset()

API Endpoints

• Web Interface at /web — Interactive UI for exploring the environment
• API Docs at /docs — Full OpenAPI/Swagger interface
• Health Check at /health — Container health monitoring
• WebSocket at /ws — Persistent session endpoint (reset/step with state)

Environment Details

Action

KantBenchAction: Single field

• move (str) — Your move (e.g. "cooperate", "defect", "hawk", "contribute_10")

Observation

KantBenchObservation: Full round result and episode state

• game_name, game_description — Current game info
• available_moves — Valid moves for this game
• your_move, opponent_move — Moves played this round (2-player)
• your_payoff, opponent_payoff — Payoffs this round (2-player)
• cumulative_score — Your total score
• round_number, max_rounds — Episode progress
• opponent_strategy — Opponent strategy name
• history — Full round-by-round history
• num_players — Number of players (N-player games only, null for 2-player)
• player_index — Your player index (N-player games only)
• all_scores — Scores for all players (N-player games only)