BenchMAX policy store

October 2026 Early Preview — Official Release Planned for January 2027

Reference policy collection for BenchMAX, a library for training and evaluating multi-agent policies.

Connect

Install BenchMAX with its training extra to load neural policies.

Browse

from benchmax.storage import reference_store

store = reference_store()
for ref in store.find_policies():
    training = ref.training
    print(training.game.name, training.algorithm.name, training.seed)

Discovery includes all configurations when no game or algorithm is supplied. When filtering by name, use training_config="any" or game_config="any" to include nondefault configurations.

Some examples via the CLI:

benchmax store info
benchmax store show --game kuhn_poker --algorithm mappo
benchmax store show --json

Evaluate or export

This example selects the current rock-paper-scissors MAPPO policy. Adjust the algorithm and seed using the listing above.

from benchmax import load_game

game = load_game("rock_paper_scissors")
reference, = store.find_policies(
    game, algorithm="mappo", seed=2009, training_config="any"
) # requires exactly one matching artifact
metrics = game.evaluate(reference)
local = reference.save(path="saved-policy")

Find more advanced usage and information in the official storage guide.