SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2106.09435 · ICML · 2021

Multi-Agent Training beyond Zero-Sum with Correlated Equilibrium Meta-Solvers

Thore Graepel, Marc Lanctot, Luke Marris, Paul Muller, Karl Tuyls

arXiv · PDF · Open in the Atlas

Code that ran

9 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 9 failures of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2106.09435.json · badge: /api/ran/2106.09435.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
deepmind/open_spiel alias (classifier label)
classifier basis: classifier_2025
9 ran · 1 no recorded run
FunctionStatusWhere it lives
_ace_constraints Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("bbfbc44ba5f27cdd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import scipy as sp

# Create a small payoff tensor for a 2-player game
# payoff shape: (num_players, num_actions_player0, num_actions_player1)
# Using a 2x3x4 tensor: 2 players, player 0 has 3 actions, player 1 has 4 actions
payoff = np.random.randn(2, 3, 4).astype(np.float64)

# Create epsilons - one per player
epsilons = np.array([0.1, 0.05], dtype=np.float64)

# Define ARGS dictionary with all parameters
ARGS = {
    "payoff": payoff,
    "epsilons": epsilons,
    "remove_null": True,
    "zero_tolerance": 0.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_eliminate_dominated_decorator Ran
ran; our contract draft was wrong, not the code
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("e718aba422fa4f4b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

def dummy_func(eliminated_payoff, eliminated_action_repeats, *args, **kwargs):
    """Dummy function to be wrapped by the decorator."""
    # Returns eliminated_dist and meta dict
    eliminated_dist = eliminated_payoff.copy()
    meta = {'test': 'value'}
    return (eliminated_dist, meta)

def _eliminate_dominated_payoff(payoff, epsilon, action_labels=None, action_repeats=None, weakly=False):
    """Eliminate epsilon dominated strategies."""
    num_players = payoff.shape[0]
    eliminated = True
    if action_labels is None:
        action_labels = [np.arange(na, dtype=np.int32) for na in payoff.shape[1:]]
    if action_repeats is not None:
        action_repeats = [ar for ar in action_repeats]
    while eliminated:
        eliminated = False
        for p in range(num_players):
            if epsilon > 0.0:
                continue
            num_actions = payoff.shape[1:]
            if num_actions[p] <= 1:
                continue
            for a in range(num_actions[p]):
                index = [slice(None) for _ in range(num_players)]
                index[p] = slice(a, a + 1)
                if weakly:
                    diff = payoff[p] <= payoff[p][tuple(index)]
                else:
                    diff = payoff[p] < payoff[p][tuple(index)]
                axis = tuple(range(p)) + tuple(range(p + 1, num_players))
                less = np.all(diff, axis=axis)
                less[a] = False
                if np.any(less):
                    nonzero = np.nonzero(less)
                    payoff = np.delete(payoff, nonzero, axis=p + 1)
                    action_labels[p] = np.delete(action_labels[p], nonzero)
                    if action_repeats is not None:
                        action_repeats[p] = np.delete(action_repeats[p], nonzero)
                    eliminated = True
                    break
    return (payoff, action_labels, action_repeats)

def _reconstruct_dist(eliminated_dist, action_labels, num_actions):
    """Returns reconstructed dist from eliminated_dist and action_labels."""
    reconstructed_payoff = np.zeros(num_actions)
    reconstructed_payoff[np.ix_(*action_labels)] = eliminated_dist
    return reconstructed_payoff

def _eliminate_dominated_decorator(func):
    """Wrap eliminate dominated."""

    def wrapper(payoff, per_player_repeats, *args, eliminate_dominated=True, **kwargs):
        epsilon = getattr(kwargs, 'epsilon', 0.0)
        if not eliminate_dominated:
            return func(payoff, *args, **kwargs)
        num_actions = payoff.shape[1:]
        eliminated_payoff, action_labels, eliminated_action_repeats = _eliminate_dominated_payoff(payoff, epsilon, action_repeats=per_player_repeats)
        eliminated_dist, meta = func(eliminated_payoff, eliminated_action_repeats, *args, **kwargs)
        meta['eliminated_dominated_dist'] = eliminated_dist
        meta['eliminated_dominated_payoff'] = eliminated_payoff
        dist = _reconstruct_dist(eliminated_dist, action_labels, num_actions)
        return (dist, meta)
    return wrapper

# Create a small payoff array: 2 players, 3 actions for player 0, 4 actions for player 1
payoff = np.random.randn(2, 3, 4).astype(np.float32)

# Create per_player_repeats
per_player_repeats = [np.array([1, 1, 1], dtype=np.int32), np.array([1, 1, 1, 1], dtype=np.int32)]

# Wrap the dummy function
wrapped_func = _eliminate_dominated_decorator(dummy_func)

ARGS = {
    'func': dummy_func,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_eliminate_dominated_payoff Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("29cf1f05cd26f3dd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a simple 2-player game payoff matrix
# Shape: (num_players, *action_spaces)
# For 2 players with 3 and 4 actions respectively
payoff = np.array([
    # Player 0's payoffs
    [[1, 2, 3, 4],
     [2, 3, 4, 5],
     [0, 1, 2, 3]],
    # Player 1's payoffs
    [[5, 4, 3, 2],
     [6, 5, 4, 3],
     [4, 3, 2, 1]]
], dtype=np.float32)

# Action labels for each player
action_labels = [
    np.array([0, 1, 2], dtype=np.int32),  # Player 0 has 3 actions
    np.array([0, 1, 2, 3], dtype=np.int32)  # Player 1 has 4 actions
]

# Action repeats (optional, tracks how many times each action was repeated)
action_repeats = [
    np.array([1, 1, 1], dtype=np.int32),
    np.array([1, 1, 1, 1], dtype=np.int32)
]

ARGS = {
    "payoff": payoff,
    "epsilon": 0.0,
    "action_labels": action_labels,
    "action_repeats": action_repeats,
    "weakly": False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_get_repeat_factor Ran
ran; our contract draft was wrong, not the code
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("dbb9554288a2c120")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import string
import numpy as np

# The function takes action_repeats as a parameter
# action_repeats should be a list/sequence of numpy arrays
# Each array represents repeat factors for a player's actions

# Looking at the function:
# - It gets the length of action_repeats (num_players)
# - It uses einsum with these arrays
# - For einsum to work, the arrays need compatible shapes for broadcasting

# Create a simple case with 2 players
# Each player has an array of action repeat factors
action_repeats = [
    np.array([2, 3, 1, 4]),  # Player 0's repeat factors
    np.array([1, 2, 3, 2])   # Player 1's repeat factors
]

ARGS = {
    'action_repeats': action_repeats
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_partition_by_player Ran
ran; our contract draft was wrong, not the code
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("66258bd7ca20a035")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned list. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct arguments for _partition_by_player
# The function partitions a value array by a player vector

# val: should be an array-like that supports indexing with boolean masks
val = np.array([10, 20, 30, 40, 50, 60])

# p_vec: a vector indicating which player each element belongs to
# Should have same length as val
p_vec = np.array([0, 1, 0, 1, 0, 2])

# num_players: the number of distinct players
num_players = 3

ARGS = {
    "val": val,
    "p_vec": p_vec,
    "num_players": num_players
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_qp_ce Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("1110ee2cafc9c0d6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import scipy as sp
import cvxpy as cp

# Set random seed for reproducibility
np.random.seed(0)

# Create a small 2-player game payoff tensor
# Shape: [NUM_PLAYER=2, NUM_ACT_0=3, NUM_ACT_1=4]
payoff = np.random.randn(2, 3, 4)

# Create constraint matrices for each player
# a_mats is a list of constraint matrices, one per player
# Each matrix has shape [NUM_CON, PROD(NUM_ACTIONS)]
# PROD(NUM_ACTIONS) = 3 * 4 = 12
num_actions_prod = 3 * 4

# For player 0: 2 constraints
a_mat_0 = np.random.randn(2, num_actions_prod)

# For player 1: 2 constraints
a_mat_1 = np.random.randn(2, num_actions_prod)

a_mats = [a_mat_0, a_mat_1]

# Create epsilon vectors for each player
e_vecs = [0.1, 0.1]

# Action repeats - one repeat factor per player
action_repeats = [np.array([1, 2, 1]), np.array([1, 1, 2, 1])]

# Solver kwargs - use default
solver_kwargs = dict(
    solver='OSQP',
    max_iter=1000000000,
    eps_abs=1e-08,
    eps_rel=1e-08,
    eps_prim_inf=1e-08,
    eps_dual_inf=1e-08,
    polish_refine_iter=100,
    check_termination=1000,
    sigma=1e-07,
    delta=1e-07,
    verbose=False
)

ARGS = {
    'payoff': payoff,
    'a_mats': a_mats,
    'e_vecs': e_vecs,
    'assume_full_support': False,
    'action_repeats': action_repeats,
    'solver_kwargs': solver_kwargs,
    'min_epsilon': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
cvxpy, numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_reconstruct_dist Ran
ran; our contract draft was wrong, not the code
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("6e6fe595eda68d92")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned ndarray, ndim 2, dtype family float, shape (4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Construct valid arguments for _reconstruct_dist

# Create a simple case with 2 actions per dimension
# Let's say we have 2 dimensions of actions
# First dimension has 4 possible actions, second has 3 possible actions
# We eliminate one action from each dimension

num_actions = [4, 3]  # A0=4, A1=3

# action_labels specifies which indices to keep for each dimension
# For dimension 0: keep indices [0, 1, 2] (eliminate 1 action, so A0E=3)
# For dimension 1: keep indices [0, 1] (eliminate 1 action, so A1E=2)
action_labels = [np.array([0, 1, 2]), np.array([0, 1])]

# eliminated_dist has shape [3, 2] (A0E=3, A1E=2)
eliminated_dist = np.array([[0.1, 0.2],
                            [0.3, 0.4],
                            [0.5, 0.6]])

ARGS = {
    "eliminated_dist": eliminated_dist,
    "action_labels": action_labels,
    "num_actions": num_actions,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (4, 3) — shape (4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_sparse_indices_generator Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("d8e86d7df8877857")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned product. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import itertools

def _sparse_indices_generator(player, action, num_actions):
    indices = [(action,) if p == player else range(na) for p, na in enumerate(num_actions)]
    return itertools.product(*indices)

ARGS = {
    "player": 1,
    "action": 2,
    "num_actions": [3, 4, 5]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type product — a product output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a product output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_try_two_solvers Ran
ran; our contract draft was wrong, not the code
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("9188dc04332e853e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from absl import logging

torch.manual_seed(0)

# Define a simple function that accepts solver_kwargs and other kwargs
def mock_optimization_func(*args, **kwargs):
    """Mock function that simulates an optimization routine."""
    solver_kwargs = kwargs.get('solver_kwargs', {})
    solver_type = solver_kwargs.get('solver', 'CVXOPT')
    
    # Simulate success with OSQP, failure with CVXOPT
    if solver_type == 'CVXOPT':
        raise ValueError("CVXOPT solver failed")
    
    # Return a mock result
    return {'status': 'success', 'solver': solver_type}

# Define arguments for _try_two_solvers
ARGS = {
    'func': mock_optimization_func,
    'args': (np.array([1.0, 2.0, 3.0]),),
    'kwargs': {'extra_param': 42, 'tolerance': 1e-6}
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
absl, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_mgce Not yet run
unverified (harvested, no recorded run)
deepmind/open_spiel/open_spiel/python/algorithms/jpsro.py
code served (permissive licence) · get_code("d652e989b3b72e17")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 10 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Two-player, constant-sum games are well studied in the literature, but there has been limited progress outside of this setting. We propose Joint Policy-Space Response Oracles (JPSRO), an algorithm for training agents in n-player, general-sum extensive form games, which provably converges to an equilibrium. We further suggest correlated equilibria (CE) as promising meta-solvers, and propose a novel solution concept Maximum Gini Correlated Equilibrium (MGCE), a principled and computationally efficient family of solutions for solving the correlated equilibrium selection problem. We conduct several experiments using CE meta-solvers for JPSRO and demonstrate convergence on n-player, general-sum games.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2106.09435")
get_code_for_paper("2106.09435")
have("2106.09435")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2106.09435.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2106.09435.svg)](https://syntology.ai/paper/2106.09435)

Connect an agent — have() is free.