SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2002.01093 · ICLR · 2020

ON THE INTERACTION BETWEEN SUPERVISION AND SELF-PLAY IN EMERGENT COMMUNICATION

Ryan Lowe, Abhinav Gupta, Mila Foerster, Douwe Kiela, Joelle Facebook, A Research

arXiv · PDF · Open in the Atlas

Code that ran

11 of 16 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (3 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 3 have no contract verdict; 7 failures of our instrument; 5 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/2002.01093.json · badge: /api/ran/2002.01093.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
backpropper/s2p canonical (classifier label)
classifier basis: classifier_2025
11 ran · 5 no recorded run
FunctionStatusWhere it lives
Beholder Ran
ran on a synthesized input · constructed an object; no result was computed
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("7a6f9cb1a39ca0b3")
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 an instance of its own class, Beholder: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

class Args:
    def __init__(self):
        self.D_img = 4
        self.D_hid = 3
        self.dropout = 0.5

ARGS = {
    'args': Args()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Beholder — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Beholder 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
Listener Ran
ran on a synthesized input · constructed an object; no result was computed
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("9ecb51a47b201507")
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 an instance of its own class, Listener: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn
import random
import numpy as np

torch.manual_seed(0)

# Define a minimal beholder model that takes images and returns embeddings
class Beholder(nn.Module):
    def __init__(self, img_dim, emb_dim):
        super(Beholder, self).__init__()
        self.linear = nn.Linear(img_dim, emb_dim)
    
    def forward(self, img):
        return self.linear(img)

# Create a minimal args object
class Args:
    def __init__(self):
        self.D_emb = 4
        self.D_hid = 5
        self.vocab_size = 6
        self.dropout = 0.1
        self.lr = 0.001
        self.device = 'cpu'
        self.i2w = {0: 'a', 1: 'b', 2: 'c'}
        self.w2i = {'a': 0, 'b': 1, 'c': 2}

args = Args()

# Create beholder instance
beholder = Beholder(img_dim=7, emb_dim=5)

# Create Listener instance
listener = nn.Module()  # placeholder, will be replaced

# Define actual test data
batch_size = 2
seq_len = 3
img_dim = 7
num_distractors = 2

# Messages: one-hot encoded, shape (batch_size, seq_len, vocab_size)
spk_msg = torch.zeros(batch_size, seq_len, args.vocab_size)
spk_msg[0, 0, 0] = 1
spk_msg[0, 1, 1] = 1
spk_msg[1, 0, 2] = 1

# Message lengths
spk_msg_lens = torch.tensor([2, 1], dtype=torch.long)

# Target image
image = torch.randn(batch_size, img_dim)

# Distractor images
distractor_images = [torch.randn(batch_size, img_dim) for _ in range(num_distractors)]

ARGS = {
    'args': args,
    'beholder': beholder,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Listener — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Listener 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
Speaker Ran
ran on a synthesized input · constructed an object; no result was computed
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("9f509297c82bb3a7")
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 an instance of its own class, Speaker: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn
import numpy as np

torch.manual_seed(0)

# Create a minimal "beholder" model that takes an image and returns hidden state
class MockBeholder(nn.Module):
    def __init__(self, img_dim, hid_dim):
        super().__init__()
        self.linear = nn.Linear(img_dim, hid_dim)
    
    def forward(self, x):
        return self.linear(x)

# Create word-to-index and index-to-word mappings
w2i = {
    '<BOS>': 0,
    '<EOS>': 1,
    '<PAD>': 2,
    'hello': 3,
    'world': 4,
}
i2w = {v: k for k, v in w2i.items()}

# Create args namespace with all required attributes
class Args:
    def __init__(self):
        self.D_emb = 4          # embedding dimension
        self.D_hid = 5          # hidden dimension
        self.vocab_size = 6     # vocab size (includes special tokens)
        self.dropout = 0.1
        self.temp = 0.5         # gumbel softmax temperature
        self.hard = True        # gumbel softmax hard mode
        self.seq_len = 3        # sequence length
        self.lr = 0.001         # learning rate
        self.i2w = i2w
        self.w2i = w2i

args = Args()

# Create the beholder
beholder = MockBeholder(img_dim=8, hid_dim=args.D_hid)

# Define ARGS dictionary
ARGS = {
    'args': args,
    'beholder': beholder,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Speaker — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Speaker 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
filter_caps Ran
ran; our contract draft was wrong, not the code
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("02b107d78fd0fca7")
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 torch

torch.manual_seed(0)

# Construct w2i (word-to-index dictionary)
w2i = {
    '<UNK>': 0,
    'hello': 1,
    'world': 2,
    'test': 3,
}

# Construct captions - list of lists, where each inner list contains token sequences
# The function checks cap[0] and counts occurrences of w2i['<UNK>']
captions = [
    [[0, 1, 2]],  # First caption: [UNK, hello, world] - has 1 UNK out of 3 = 33%
    [[1, 2, 3]],  # Second caption: [hello, world, test] - has 0 UNK out of 3 = 0%
]

# Construct images - list of tensors that will be stacked
# Each image tensor should have compatible shape for stacking
images = [
    torch.randn(3, 4),  # Image 1: shape (3, 4)
    torch.randn(3, 4),  # Image 2: shape (3, 4)
]

# Percentage threshold - filter captions with UNK ratio below this
perc = 0.5

ARGS = {
    'captions': captions,
    'images': images,
    'w2i': w2i,
    'perc': perc,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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_s2p_batch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("5a0aaf20a82cf665")
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. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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
import random

torch.manual_seed(0)
random.seed(0)

# Small stand-in args class with required attributes
class Args:
    def __init__(self):
        self.s2p_batch_size = 2
        self.num_distrs = 3
        self.seq_len = 8
        self.vocab_size = 16

args = Args()

# Create seed_train_images: shape (5, 3, 4) - 5 images, 3 channels, 4x4
seed_train_images = torch.randn(5, 3, 4)

# Create seed_train_captions: list of 5 caption arrays, each with variable length up to seq_len
seed_train_captions = [
    np.array([1, 2, 3], dtype=np.long),
    np.array([4, 5], dtype=np.long),
    np.array([6, 7, 8, 9, 10], dtype=np.long),
    np.array([11, 12, 13], dtype=np.long),
    np.array([14], dtype=np.long),
]

ARGS = {
    "seed_train_images": seed_train_images,
    "seed_train_captions": seed_train_captions,
    "args": args,
    "batch_size": None,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — 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
gumbel_softmax Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("779be7e9ddb772f0")
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. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn.functional as F

torch.manual_seed(0)

# Create a small batch of logits
# Shape: (batch_size=2, num_classes=4)
logits = torch.randn(2, 4)

# Temperature parameter for gumbel softmax
temp = 0.5

# Hard parameter to control whether to use straight-through estimator
hard = True

ARGS = {
    'logits': logits,
    'temp': temp,
    'hard': hard
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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
gumbel_softmax_sample Ran
ran, honoured the contract we drafted
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("626e470a849ef8d9")
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 Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch
import torch.nn.functional as F

torch.manual_seed(0)

# Create small logits tensor (batch=2, num_classes=4)
logits = torch.randn(2, 4)

# Create a temperature value (scalar float)
temp = 0.5

ARGS = {
    "logits": logits,
    "temp": temp,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
sample_gumbel Ran
ran; our contract draft was wrong, not the code
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("eb35bf807af0e213")
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 Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch

torch.manual_seed(0)

device = torch.device('cpu')

def sample_gumbel(shape, eps=1e-20):
    U = torch.empty(shape, device=device).uniform_(0, 1)
    return -torch.log(-torch.log(U + eps) + eps)

ARGS = {
    'shape': (2, 4),
    'eps': 1e-20
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
trim_caps Ran
ran; our contract draft was wrong, not the code
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("7b8ff1df3ed9fbd2")
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 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

torch.manual_seed(0)

# caps is a list of lists of strings
# Each inner list contains captions (strings)
# minlen and maxlen are integers for filtering by string length

caps = [
    ["a", "hello world", "hi", "testing caption here"],
    ["ok", "short", "this is a longer caption", "x"]
]

minlen = 2
maxlen = 10

ARGS = {
    "caps": caps,
    "minlen": minlen,
    "maxlen": maxlen
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
truncate_captions Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("bfdb7c5db11d3441")
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 torch
import numpy as np

torch.manual_seed(0)

# Build word-to-index and index-to-word mappings
w2i = {
    '<UNK>': 0,
    'hello': 1,
    'world': 2,
    'foo': 3,
    'bar': 4,
}

i2w = {
    0: '<UNK>',
    1: 'hello',
    2: 'world',
    3: 'foo',
    4: 'bar',
}

# Build nested list structure: [batch][captions_per_item][word_indices]
# Some indices are valid (in i2w), some are invalid (not in i2w) to test UNK replacement
train_captions = [
    [[1, 2, 3], [4, 1, 999]],  # 999 is invalid, should become 0
    [[2, 4, 1], [3, 2, 888]],  # 888 is invalid, should become 0
]

valid_captions = [
    [[1, 2], [3, 4, 777]],  # 777 is invalid, should become 0
]

test_captions = [
    [[4, 1, 2, 3], [1, 666]],  # 666 is invalid, should become 0
]

ARGS = {
    'train_captions': train_captions,
    'valid_captions': valid_captions,
    'test_captions': test_captions,
    'w2i': w2i,
    'i2w': i2w,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — 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
truncate_dicts Ran
ran; our contract draft was wrong, not the code
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("d004ca3c465884a6")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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.

from collections import OrderedDict

# Construct w2i (word to index) mapping
w2i = OrderedDict([
    ('<PAD>', 0),
    ('<UNK>', 1),
    ('<BOS>', 2),
    ('<EOS>', 3),
    ('hello', 4),
    ('world', 5),
    ('foo', 6),
    ('bar', 7),
    ('baz', 8),
    ('qux', 9),
])

# Construct i2w (index to word) mapping
i2w = OrderedDict([
    (0, '<PAD>'),
    (1, '<UNK>'),
    (2, '<BOS>'),
    (3, '<EOS>'),
    (4, 'hello'),
    (5, 'world'),
    (6, 'foo'),
    (7, 'bar'),
    (8, 'baz'),
    (9, 'qux'),
])

# Truncation size
trunc_size = 6

ARGS = {
    'w2i': w2i,
    'i2w': i2w,
    'trunc_size': trunc_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
SpeakerListener Not yet run
unverified (harvested, no recorded run)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("057dbfd6c2a4131d")
How it ran

no invocation recorded.

load_listener Not yet run
unverified (harvested, no recorded run)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("6509889b8433dd2e")
How it ran

no invocation recorded.

main Not yet run
unverified (harvested, no recorded run)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("5dffc621e220f31f")
How it ran

no invocation recorded.

save_to_file Not yet run
unverified (harvested, no recorded run)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("975549298f10836c")
How it ran

no invocation recorded.

torch_save_to_file Not yet run
unverified (harvested, no recorded run)
backpropper/s2p/ibr_game/finetune.py
code served (permissive licence) · get_code("409eb80fda6b01b0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 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 16 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

A promising approach for teaching artificial agents to use natural language involves using human-in-the-loop training. However, recent work suggests that current machine learning methods are too data inefficient to be trained in this way from scratch. In this paper, we investigate the relationship between two categories of learning signals with the ultimate goal of improving sample efficiency: imitating human language data via supervised learning, and maximizing reward in a simulated multi-agent environment via self-play (as done in emergent communication), and introduce the term supervised self-play (S2P) for algorithms using both of these signals. We find that first training agents via supervised learning on human data followed by self-play outperforms the converse, suggesting that it is not beneficial to emerge languages from scratch. We then empirically investigate various S2P schedules that begin with supervised learning in two environments: a Lewis signaling game with symbolic inputs, and an image-based referential game with natural language descriptions. Lastly, we introduce population based approaches to S2P, which further improves the performance over single-agent methods. 1

For agents

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

get_harvested_code_for_paper("2002.01093")
get_code_for_paper("2002.01093")
have("2002.01093")

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

curl https://syntology.ai/api/ran/2002.01093.json

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

[![Syntology run record](https://syntology.ai/api/ran/2002.01093.svg)](https://syntology.ai/paper/2002.01093)

Connect an agent — have() is free.