PlyBench

PlyBench is a benchmark suite for evaluating the performance of LLMs and LLM agents in simple, fully-observable game environments. It pits players (LLMs, MCTS, optimal solvers, random, or human) against each other across a matrix of games and records every step for later analysis.

Paper: Towards Improving Sequential Decision-Making in LLM Agents via Experience Memory (arXiv:2608.03420) — see Citation.

Package: pypi.org/project/plybench

Live results for a selection of models and games are available at plybench.jakubrada.com.

Installation

PlyBench is on PyPI as plybench:

pip install plybench

Or with uv:

uv add plybench

Requires Python 3.12+.

Each LLM provider SDK is an optional extra — install only the ones you need (or all):

pip install "plybench[openai]"         # OpenAI
pip install "plybench[gemini]"         # Google Gemini
pip install "plybench[grok]"           # Grok (OpenAI-compatible endpoint)
pip install "plybench[anthropic]"      # Anthropic Claude
pip install "plybench[mistral]"        # Mistral
pip install "plybench[metacentrum]"    # Metacentrum (OpenAI-compatible endpoint)
pip install "plybench[huggingface]"    # local HuggingFace models (torch + transformers)
pip install "plybench[all]"            # everything

Providers whose SDK is not installed are simply skipped when building PlyBench().

Quickstart

Building an PlyBench object is the one-stop bootstrap: it creates an instance-scoped registry, registers the built-in games and players, and wires up the LLM router.

import asyncio

from plybench import PlyBench
from plybench.harness.benchmark import Benchmark

op = PlyBench()  # reads provider keys from the environment (see Configuration)

benchmark = Benchmark(
    experiment="quickstart",
    op=op,
    game_configs=["tic_tac_toe:"],
    player_configs=["random:distribution=uniform"],
    opponent_configs=["optimal:stochastic=True"],
    num_games=10,
)

results = asyncio.run(benchmark.run())

Runs are resumable and written under results/benchmarks/<experiment>/ in the current working directory; re-running skips already-completed rounds.

Config strings

Games and players are addressed by name:key=value:key=value strings, e.g. llm:actions:text:openai:gpt-5:thinking_enabled=True or random:distribution=uniform.

Built-in games: tic_tac_toe, modified_tic_tac_toe, magic_square, story_magic_square, nim, modified_nim, inverse_nim, story_nim, connect_four, breakthrough.

Built-in players: human, random, mcts, optimal, llm.

Extend either set at runtime via op.registry.register_game(...) / op.registry.register_player(...).

Configuration

LLM providers are configured through environment variables (a .env file is loaded automatically). PlyBench() self-disables any provider whose key is absent, so bot-only benchmarks run offline. See .env.example:

OPENAI_API_KEY=...
OPENAI_ORGANIZATION=...
OPENAI_PROJECT=...

GEMINI_API_KEY=...
GEMINI_PROJECT=...

GROK_API_KEY=...
GROK_BASE_URL=...   # optional, defaults to https://api.x.ai/v1

CLAUDE_API_KEY=...

MISTRAL_API_KEY=...

METACENTRUM_BASE_URL=...
METACENTRUM_API_KEY=...

HF_TOKEN=...   # only for gated/private HuggingFace models

NTFY_URL=...   # optional, enables progress notifications (see Notifications)
NTFY_TOKEN=...

Rate limits

Two independent layers protect against provider throttling:

• Provider concurrency (ProviderSemaphore) caps in-flight requests per provider. This is the only mechanism most providers need — set it with PlyBench(concurrency=...) or llm.set_concurrency(provider, n).
• Per-model quotas (ModelLimits) additionally pace a single model by in-flight share, requests per second and tokens per minute.

Both layers apply to every provider — the gate is enforced in the shared dispatch path each client routes its API call through, so nothing is provider-specific.

No model ships with a quota, because published allowances are account-specific — they depend on your tier and differ per model. Apply your own when you know them:

from plybench.llm import ModelLimits, Provider

op.llm.set_model_limits(Provider.MISTRAL, "mistral-small-4", ModelLimits(max_concurrent=8, rps=1.67, tpm=100_000))
op.llm.set_model_limits(Provider.MISTRAL, "mistral-small-4", None)  # back to unlimited

Every field is optional, so you can pace on tokens alone and leave requests unbounded. Unset limits mean a model is governed only by the provider semaphore, which is why adding this changed nothing for existing providers. The repo-local scripts keep their quotas in LIMITS in scripts/_shared.py, keyed by provider and model name — a useful pattern to copy, since that file is not part of the installed package.

Token pacing reserves an estimate before each call (prompt length plus max_tokens, or the model's default output guess) and reconciles it against reported usage once the response lands, so an inaccurate estimate costs a little throughput rather than correctness. Because quotas are account-wide while each process paces only itself, pass a scale below 1.0 when several runs share an account. safe_call's backoff remains the backstop for any 429 that slips through.

When per-model shares exceed the provider cap they queue behind it, so set the provider concurrency at or above their sum (--concurrency 24 for the two models above) to keep one model from holding slots the other's quota could use.

HuggingFace (local models)

The HuggingFace provider runs models locally instead of calling a remote API — it exposes embed() (generation is not supported yet). hf_models selects which of the supported models (see providers/huggingface/models.py) this environment uses; they are downloaded/verified into the local HF cache at bootstrap:

op = PlyBench(hf_models=["sup-simcse-bert"])
resp = await op.llm.embed(Provider.HUGGINGFACE, "sup-simcse-bert", ["hello"])

Requesting a supported model that was not part of hf_models raises an error telling you to add it to the bootstrap list. Needs the huggingface extra installed.

Notifications

Long benchmark runs can push progress notifications to ntfy.sh (or any compatible endpoint) — one message per finished matchup, with elapsed time, rounds completed and an ETA derived from round throughput, plus a final summary. Set NTFY_URL (and NTFY_TOKEN for protected topics) and opt in per run:

op = PlyBench(notif_enabled=True)
op.notif.notify("hello")  # no-op when disabled or unconfigured

Notifications are off unless notif_enabled=True, and failures are logged as warnings rather than interrupting the run.

Extending PlyBench

Games and players are open registries on op.registry — you can add your own from your own code without modifying the package. Each is registered as a spec that pairs a config-string key with the classes that implement it.

Adding a player

Implement two things:

1. A PlayerParams subclass (configs/player_params.py) — the parsed form of your config string. Implement from_string / to_string / path_suffix. Reuse NoGameParams-style emptiness if your player is parameterless.
2. A Player subclass (player/player.py) — implement initialize_policy (one-time setup per game), the async __call__ (given the game, an InterfaceObservation, and the legal InterfaceActions, return a PlayerOutput — set action=None to forfeit), and format_llm_output.

Optionally, attach a PlayerTracker to persist extra per-step data onto each recorded GameStep.

Then register a PlayerSpec:

from dataclasses import dataclass

from plybench import PlyBench
from plybench.configs.player_config import PlayerConfig
from plybench.configs.player_params import PlayerParams
from plybench.core.game import TurnBasedGame
from plybench.core.interface import InterfaceAction, InterfaceObservation
from plybench.core.prompt_adapter import PromptAdapter
from plybench.player.player import Player, PlayerIdentifier, PlayerOutput
from plybench.player.spec import PlayerSpec


@dataclass(frozen=True, eq=True)
class FirstMoveParams(PlayerParams):
    @classmethod
    def from_string(cls, params_string: str) -> "FirstMoveParams":
        return cls()

    def to_string(self) -> str:
        return ""

    @property
    def path_suffix(self) -> str:
        return ""


class FirstMovePlayer(Player):
    def initialize_policy(self, game: TurnBasedGame, prompt_adapter_template: PromptAdapter) -> None:
        pass

    async def __call__(self, game: TurnBasedGame, observation: InterfaceObservation, legal_moves: list[InterfaceAction]) -> PlayerOutput:
        return PlayerOutput(action=legal_moves[0] if legal_moves else None)

    def format_llm_output(self, player_output: PlayerOutput) -> str:
        return ""


op = PlyBench()
op.registry.register_player(PlayerSpec("first", FirstMoveParams, lambda game, cfg, pid: FirstMovePlayer(cfg, pid)))
# usable anywhere as the config string "first:"

Adding a game

Games are backed by OpenSpiel: the underlying game must be loadable by pyspiel.load_game(...) (a built-in OpenSpiel game or a custom game you register with OpenSpiel). A new variant implements the same set of classes the built-ins do — use any game under src/plybench/games/ (e.g. tic_tac_toe/tic_tac_toe.py) as a template:

• TurnBasedGame — binds a registry game_type key to an OpenSpiel game_name.
• InterfaceTransformer — renders state/actions for both display and the LLM prompt.
• InterfaceAction / InterfaceObservation — convert to and from OpenSpiel (from_openspiel / to_openspiel).
• PromptAdapter — the game's head prompt and expected action format.
• TurnBasedEngine — wires all of the above together (engine_factory: GameConfig -> TurnBasedEngine).
• Optionally a GameParams subclass for parameterized variants (or reuse NoGameParams).

Then register a GameSpec:

from plybench.configs.game_params import NoGameParams
from plybench.games.spec import GameSpec

op.registry.register_game(
    GameSpec(
        key="my_game",
        params_cls=NoGameParams,  # or a custom GameParams subclass
        engine_factory=MyGameEngine,  # GameConfig -> TurnBasedEngine
        solvable=False,  # True enables minimax optimality/regret analysis (small trees only)
    )
)
# usable anywhere as the config string "my_game:"

Registered games and players work everywhere the built-ins do — in Benchmark, the analysis pipeline, and the scripts.

Reproducing the paper results

The experiment scripts under scripts/ are not part of the installed package — they are the tooling used to produce and reproduce the results of the paper from a repository checkout. Use them when you want to re-run the exact benchmarks the paper reports, extend them with new models, or run the analysis and export pipelines on the resulting transcripts. Everything is resumable, so an interrupted run continues where it left off.

The complete experiment definitions from the paper are prepared under experiments/benchmarks/:

Each file declares the full sweep — the LLM players, the opponents (random, mcts, optimal), and the number of rounds — with per-item enabled toggles so you can narrow a run without editing the sweep.

git clone https://github.com/radajakub/plybench.git
cd plybench
uv sync

# run a prepared experiment (reads experiments/benchmarks/ttt.json)
uv run python scripts/run.py --experiment ttt

# or an ad-hoc smoke run
uv run python scripts/run.py --name smoke \
    --games tic_tac_toe: \
    --players random:distribution=uniform \
    --opponents optimal:stochastic=True \
    --num-games 10

# push progress notifications for a long run (needs NTFY_URL, see Notifications)
uv run python scripts/run.py --experiment ttt --notify

run.py logs matchup and round progress to the console as the run proceeds; rounds already completed by an earlier run are skipped and not logged again.

The LLM matchups require the relevant provider API keys (see Configuration) and will incur API cost; bot-vs-bot matchups run offline. Results are written under results/benchmarks/<experiment>/.

Then analyze or export the transcripts:

uv run python scripts/analyze.py --experiment ttt   # compute per-matchup statistics + confidence intervals
uv run python scripts/export.py --experiment ttt --out ttt.tar.gz   # export results for the PlyBench website
uv run python scripts/play.py --game tic_tac_toe: --i human: --o optimal:stochastic=True  # play interactively

The full result set is large (tens of thousands of game transcripts) and is not stored in this repository.

Citation

If you use PlyBench in your research, please cite the paper (arXiv preprint for now — this entry will be updated once the proceedings version is out):

@misc{rada2026plybench,
  title         = {Towards Improving Sequential Decision-Making in LLM Agents via Experience Memory},
  author        = {Rada, Jakub and Lis{\'y}, Viliam},
  year          = {2026},
  eprint        = {2608.03420},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2608.03420}
}

License

MIT © Jakub Rada