SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2310.02207 · 2023

Language Models Represent Space and Time

arXiv · PDF · Open in the Atlas

Code that ran

11 of 24 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; 2 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 9 failures of our instrument; 13 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/2310.02207.json · badge: /api/ran/2310.02207.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
wesg52/world-models canonical (classifier label)
classifier basis: census
11 ran · 13 no recorded run
FunctionStatusWhere it lives
adjust_precision Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
wesg52/world-models/utils.py
code served (permissive licence) · get_code("c0e20198c0c90a86")
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. 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)

# Create a small activation tensor for testing
# Shape: (2, 4) - batch size 2, 4 channels
activation_tensor = torch.randn(2, 4)

# Define arguments dict for all parameters
ARGS = {
    "activation_tensor": activation_tensor,
    "output_precision": 8,
    "per_channel": True,
    "cos_sim": False
}
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, shape (2, 4) — shape (2, 4) as a later run recorded it
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_model_family Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/utils.py
code served (permissive licence) · get_code("9931b4001e67bb95")
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 str. 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

ARGS = {
    'model_name': 'pythia-70m'
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
get_top_neurons Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
wesg52/world-models/model_composition.py
code served (permissive licence) · get_code("69dc982a438d4f48")
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 2, 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.

import torch
import numpy as np
import torch.nn as nn

torch.manual_seed(0)

# Build a minimal model with W_in and W_out attributes
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        # W_in shape: (n_layers=2, d_model=4, d_mlp=3)
        self.W_in = torch.randn(2, 4, 3)
        # W_out shape: (n_layers=2, d_mlp=3, d_model=4)
        self.W_out = torch.randn(2, 3, 4)
    
    def forward(self, x):
        return x

model = MinimalModel()

# learned_probes shape: (l2=2, d=4) - must match d dimension in W_norm
learned_probes = torch.randn(2, 4)

# Build ARGS dict
ARGS = {
    "learned_probes": learned_probes,
    "model": model,
    "k": 3,
    "W_in": True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, 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
load_activation_probing_dataset Ran
ran, honoured the contract we drafted
wesg52/world-models/save_activations.py
code served (permissive licence) · get_code("80563515302e7660")
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. 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 os
import torch
import tempfile
import shutil

torch.manual_seed(0)

# Create a temporary directory structure to mimic the activation_datasets layout
temp_dir = tempfile.mkdtemp()
activation_datasets_dir = os.path.join(temp_dir, 'activation_datasets')
os.makedirs(activation_datasets_dir, exist_ok=True)

# Create nested directories for model and entity_type
model_name = 'bert-base'
entity_type = 'person'
model_dir = os.path.join(activation_datasets_dir, model_name)
entity_dir = os.path.join(model_dir, entity_type)
os.makedirs(entity_dir, exist_ok=True)

# Create a dummy activation tensor and save it
activation_tensor = torch.randn(2, 4)
prompt_name = 'simple'
layer_ix = 0
activation_aggregation = 'last'
save_name = f'{entity_type}.{activation_aggregation}.{prompt_name}.{layer_ix}.pt'
save_path = os.path.join(entity_dir, save_name)
torch.save(activation_tensor, save_path)

# Change working directory to temp_dir so relative paths work
original_cwd = os.getcwd()
os.chdir(temp_dir)

ARGS = {
    'model': model_name,
    'entity_type': entity_type,
    'prompt_name': prompt_name,
    'layer_ix': layer_ix,
    'activation_aggregation': activation_aggregation
}
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
load_activation_probing_dataset_args Ran
ran, honoured the contract we drafted
wesg52/world-models/save_activations.py
code served (permissive licence) · get_code("a7b32ab38a32b07a")
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. 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 os
import torch
import tempfile

torch.manual_seed(0)

# Create a temporary directory to serve as the activation dataset directory
temp_dir = tempfile.mkdtemp()
os.environ['ACTIVATION_DATASET_DIR'] = temp_dir

# Create the necessary directory structure
model_name = 'test_model'
entity_type = 'test_entity'
model_dir = os.path.join(temp_dir, model_name, entity_type)
os.makedirs(model_dir, exist_ok=True)

# Create a small dummy activation tensor and save it
prompt_name = 'test_prompt'
layer_ix = 0
activation_tensor = torch.randn(2, 4)
save_name = f'{entity_type}.default_aggregation.{prompt_name}.{layer_ix}.pt'
save_path = os.path.join(model_dir, save_name)
torch.save(activation_tensor, save_path)

# Create a simple args object
class Args:
    def __init__(self, model, entity_type, activation_aggregation):
        self.model = model
        self.entity_type = entity_type
        self.activation_aggregation = activation_aggregation

args_obj = Args(
    model=model_name,
    entity_type=entity_type,
    activation_aggregation='default_aggregation'
)

ARGS = {
    'args': args_obj,
    'prompt_name': prompt_name,
    'layer_ix': layer_ix
}
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
load_all_probe_evals Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/evaluate_probes.py
code served (permissive licence) · get_code("e40d500a0bbcefbe")
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 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 os
import pickle
import tempfile
import shutil

# Create a temporary directory structure for the probe evaluations
temp_dir = tempfile.mkdtemp()
results_dir = os.path.join(temp_dir, 'results')

# Set up the directory structure
experiment_name = 'test_experiment'
model_name = 'test_model'
entity_type = 'test_entity'
feature_name = 'test_feature'

probe_dir = os.path.join(results_dir, experiment_name, model_name, entity_type, feature_name, 'evaluations')
os.makedirs(probe_dir, exist_ok=True)

# Create some dummy probe evaluation files
probe_results_1 = {'accuracy': 0.95, 'loss': 0.05}
probe_results_2 = {'accuracy': 0.92, 'loss': 0.08}

pickle.dump(probe_results_1, open(os.path.join(probe_dir, 'probe_1.p'), 'wb'))
pickle.dump(probe_results_2, open(os.path.join(probe_dir, 'probe_2.p'), 'wb'))

# Set the environment variable
os.environ['RESULTS_DIR'] = results_dir

ARGS = {
    'experiment_name': experiment_name,
    'model_name': model_name,
    'entity_type': entity_type,
    'feature_name': feature_name,
}
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 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
load_evaluation Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/evaluate_probes.py
code served (permissive licence) · get_code("cd39e9fbf2e85bb4")
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 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 os
import pickle
import tempfile
from pathlib import Path

# Create a temporary directory structure to simulate the results directory
temp_dir = tempfile.mkdtemp()
os.environ['RESULTS_DIR'] = temp_dir

# Create the nested directory structure that the function expects
experiment_name = "test_experiment"
model = "test_model"
entity_type = "test_entity"
feature_name = "test_feature"

save_path = os.path.join(
    temp_dir,
    experiment_name,
    model,
    entity_type,
    feature_name,
    'evaluations'
)
os.makedirs(save_path, exist_ok=True)

# Create a dummy evaluation file with pickle
evaluation_type = "accuracy"
experiment_type = "classification"
normalization_type = "none"
label_processing = "none"
activation_aggregation = "last"

evaluation_metadata = [evaluation_type, experiment_type, normalization_type, label_processing, activation_aggregation]
eval_name = '.'.join(evaluation_metadata) + '.p'
save_file = os.path.join(save_path, eval_name)

# Create dummy evaluation data and save it
dummy_evaluation_data = {
    "accuracy": 0.95,
    "precision": 0.92,
    "recall": 0.88,
    "f1": 0.90
}

with open(save_file, 'wb') as f:
    pickle.dump(dummy_evaluation_data, f)

# Define the ARGS dictionary with all parameters
ARGS = {
    "experiment_name": "test_experiment",
    "model": "test_model",
    "entity_type": "test_entity",
    "feature_name": "test_feature",
    "evaluation_type": "accuracy",
    "experiment_type": "classification",
    "normalization_type": "none",
    "label_processing": "none",
    "activation_aggregation": "last"
}
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 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
load_model Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
wesg52/world-models/load.py
code served (permissive licence) · get_code("cd24389a9fa9621c")
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 MockHookedTransformer. 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
from transformers import LlamaForCausalLM, AutoTokenizer

torch.manual_seed(0)

# Create a minimal stand-in for HookedTransformer
class MockHookedTransformer(nn.Module):
    def __init__(self):
        super().__init__()
        self.cfg = type('Config', (), {'device': 'cpu'})()
        
    def eval(self):
        return self
    
    def to(self, device):
        return self
    
    @classmethod
    def from_pretrained(cls, model_name, device='cpu', checkpoint_index=None, hf_model=None, fold_ln=False, center_writing_weights=False, center_unembed=False, tokenizer=None):
        return cls()

# Monkey-patch HookedTransformer for this test
import sys
current_module = sys.modules[__name__]
current_module.HookedTransformer = MockHookedTransformer

# Create stand-in objects
class MockTokenizer:
    def __call__(self, text):
        return {"input_ids": torch.tensor([[1, 2, 3]])}

class MockLlamaModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(10, 10)
    
    def eval(self):
        return self
    
    def to(self, device):
        return self

# Build arguments for load_model
ARGS = {
    'model_name': 'pythia-70m',
    'device': 'cpu',
    'checkpoint_index': None,
    'use_hf': False,
    'dtype': torch.float32
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MockHookedTransformer — a MockHookedTransformer output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MockHookedTransformer 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
load_probe_results Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/probe_experiment.py
code served (permissive licence) · get_code("3bf568dbe6a7fbb8")
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 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 os
import pickle
import tempfile

# Create a temporary directory to serve as RESULTS_DIR
temp_dir = tempfile.mkdtemp()
os.environ['RESULTS_DIR'] = temp_dir

# Create the necessary directory structure and a dummy probe results file
experiment_name = 'test_experiment'
model = 'bert_base'
entity_type = 'subject'
feature_name = 'attention'
prompt = 'default'
experiment_type = 'lsq_train'
normalization_type = 'none'
label_processing = 'none'
activation_aggregation = 'last'
pca = False

save_path = os.path.join(temp_dir, experiment_name, model, entity_type, feature_name, 'probes')
os.makedirs(save_path, exist_ok=True)

probe_metadata = [experiment_type, normalization_type, label_processing, activation_aggregation, prompt]
probe_name = '.'.join(probe_metadata) + '.p'
probe_file_path = os.path.join(save_path, probe_name)

# Create a dummy probe results object (could be dict, list, or any picklable object)
dummy_probe_results = {
    'accuracy': 0.95,
    'loss': 0.05,
    'weights': [0.1, 0.2, 0.3]
}

# Write the dummy results to a pickle file
with open(probe_file_path, 'wb') as f:
    pickle.dump(dummy_probe_results, f)

ARGS = {
    'experiment_name': experiment_name,
    'model': model,
    'entity_type': entity_type,
    'feature_name': feature_name,
    'prompt': prompt,
    'experiment_type': experiment_type,
    'normalization_type': normalization_type,
    'label_processing': label_processing,
    'activation_aggregation': activation_aggregation,
    'pca': pca
}
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 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
process_activation_batch Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/save_activations.py
code served (permissive licence) · get_code("c4e9a9eec8e92cc3")
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 (3, 4) 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 torch
import einops

torch.manual_seed(0)

class Args:
    def __init__(self, activation_aggregation):
        self.activation_aggregation = activation_aggregation

# Test with activation_aggregation = None
batch_size = 2
num_tokens = 3
d_act = 4

batch_activations_none = torch.randn(batch_size, num_tokens, d_act)
batch_mask_none = torch.tensor([
    [1, 1, 0],
    [1, 0, 0]
], dtype=torch.float32)
flattened_mask = einops.rearrange(batch_mask_none, 'b c -> (b c)')
batch_mask_flat = flattened_mask > 0.5

args_none = Args(activation_aggregation=None)

ARGS = {
    "args": args_none,
    "batch_activations": batch_activations_none,
    "step": 0,
    "batch_mask": batch_mask_flat
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, 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 (3, 4) — shape (3, 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
remove_lower_specificity_duplicates Ran
ran; our contract draft was wrong, not the code
wesg52/world-models/feature_datasets/space_world.py
code served (permissive licence) · get_code("f13f4939cf537d87")
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 DataFrame. 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 pandas as pd
import numpy as np

# Create a sample DataFrame with entity data
entity_df = pd.DataFrame({
    'type': ['Person', 'Person', 'Organization', 'Person', 'Organization', 'Location', 'Location'],
    'name': ['Alice', 'Bob', 'TechCorp', 'Alice', 'TechCorp', 'NYC', 'NYC'],
    'value': [1, 2, 3, 4, 5, 6, 7]
})

column_name = 'name'

ARGS = {
    'entity_df': entity_df,
    'column_name': column_name
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 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
compute_prox_error Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/generalization_experiment.py
code served (permissive licence) · get_code("1be061eaded8779b")
How it ran

no invocation recorded.

load_generalization_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/generalization_experiment.py
code served (permissive licence) · get_code("7f131954ed6e43de")
How it ran

no invocation recorded.

load_tokenizer Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/make_prompt_datasets.py
code served (permissive licence) · get_code("ec78e0b3d6261ed5")
How it ran

no invocation recorded.

make_correlation_df Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/model_composition.py
code served (permissive licence) · get_code("4a1ed06862b5c9c0")
How it ran

no invocation recorded.

make_generalization_rdf Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/analysis/generalization.py
code served (permissive licence) · get_code("403d8927d15ed841")
How it ran

no invocation recorded.

make_natural_place_entity_df Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/feature_datasets/space_world.py
code served (permissive licence) · get_code("c070305f38faca0e")
How it ran

no invocation recorded.

make_structure_entity_df Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/feature_datasets/space_world.py
code served (permissive licence) · get_code("926cbc77347d1214")
How it ran

no invocation recorded.

place_neuron_correlations Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/model_composition.py
code served (permissive licence) · get_code("5043e7864198f632")
How it ran

no invocation recorded.

place_probe_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/probe_experiment.py
code served (permissive licence) · get_code("94dc1fb81c6a70d6")
How it ran

no invocation recorded.

place_probe_generalization_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/prompt_gen_experiment.py
code served (permissive licence) · get_code("cd2f271c58a3f210")
How it ran

no invocation recorded.

run_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/generalization_experiment.py
code served (permissive licence) · get_code("eeab6f68b890a0fb")
How it ran

no invocation recorded.

time_probe_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/probe_experiment.py
code served (permissive licence) · get_code("af9c1c4eb6c780a6")
How it ran

no invocation recorded.

time_probe_generalization_experiment Not yet run
unverified (harvested, no recorded run)
wesg52/world-models/prompt_gen_experiment.py
code served (permissive licence) · get_code("88794b7e8476006e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 24 samples lifted from this paper's linked repositories, 24 carry a function identity and 0 do not. Of the 24 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 24 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

The capabilities of large language models (LLMs) have sparked debate over whether such systems just learn an enormous collection of superficial statistics or a set of more coherent and grounded representations that reflect the real world. We find evidence for the latter by analyzing the learned representations of three spatial datasets (world, US, NYC places) and three temporal datasets (historical figures, artworks, news headlines) in the Llama-2 family of models. We discover that LLMs learn linear representations of space and time across multiple scales. These representations are robust to prompting variations and unified across different entity types (e.g. cities and landmarks). In addition, we identify individual "space neurons" and "time neurons" that reliably encode spatial and temporal coordinates. While further investigation is needed, our results suggest modern LLMs learn rich spatiotemporal representations of the real world and possess basic ingredients of a world model.

For agents

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

get_harvested_code_for_paper("2310.02207")
get_code_for_paper("2310.02207")
have("2310.02207")

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

curl https://syntology.ai/api/ran/2310.02207.json

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

[![Syntology run record](https://syntology.ai/api/ran/2310.02207.svg)](https://syntology.ai/paper/2310.02207)

Connect an agent — have() is free.