SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2601.21026 · 2026

Diffusion-based Annealed Boltzmann Generators : benefits, pitfalls and hopes

Louis Grenioux

arXiv · PDF · Open in the Atlas

Code that ran

18 of 23 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; 18 failures of our instrument; 5 no recorded run), measured 2026-11-22; 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/2601.21026.json · badge: /api/ran/2601.21026.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
h2o64/dabg canonical (classifier label)
classifier basis: census
18 ran · 5 no recorded run
FunctionStatusWhere it lives
compute_grad_and_laplacian_batched_fn Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/utils.py
code served (permissive licence) · get_code("34a3d5ccc2adbf8c")
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 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 torch
import torch.nn as nn

torch.manual_seed(0)

# The function takes grad_f_batched as input
# grad_f_batched is called with batched t and x tensors
# It should return gradients in batched form
# Looking at the usage: grad_f_batched(t.unsqueeze(0), x.unsqueeze(0)).squeeze(0)
# This suggests grad_f_batched takes (batch_size, ...) shaped tensors

def grad_f_batched(t, x):
    """
    Mock gradient function that takes batched t and x.
    t: shape (batch_size, 1) - time variable
    x: shape (batch_size, 4) - spatial variable
    returns: shape (batch_size, 4) - gradient w.r.t. x
    """
    batch_size = x.shape[0]
    # Return a simple linear function of x as the gradient
    return x * 2.0

ARGS = {
    "grad_f_batched": grad_f_batched
}
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 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
compute_laplacian_fn Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/ebm.py
code served (permissive licence) · get_code("8e282faaa16c04a5")
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 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 torch
import torch.nn as nn

torch.manual_seed(0)

def simple_function(t, x):
    """
    A simple scalar function that takes time t (scalar) and position x (vector).
    Returns a scalar value.
    """
    # x shape: (batch_size, dim)
    # t shape: (batch_size,)
    # We'll create a simple quadratic function
    return torch.sum(x ** 2) + torch.sum(t.unsqueeze(-1) * x)

ARGS = {
    "f_non_batched": simple_function
}
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 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
compute_pairwise_sq_distance_same Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/metrics/mmd.py
code served (permissive licence) · get_code("e63b2f2f8ef209d1")
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 3, 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

torch.manual_seed(0)

# A is used with torch.mm(A, A.t()), so it needs to be a 2D tensor
# We'll create a small matrix: batch_size x feature_dim
# Using 2 x 4 to keep it tiny and distinct
A = torch.randn(2, 4)

ARGS = {
    "A": A
}
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, ndim 3, 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
compute_random_proj_cdf Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/metrics/ks.py
code served (permissive licence) · get_code("33708cbaa252489c")
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, 5) 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

torch.manual_seed(0)

# Create samples tensor of shape (batch_size, dim)
batch_size = 2
dim = 4
samples = torch.randn(batch_size, dim)

# Create random projections tensor of shape (n_random_projections, dim)
n_random_projections = 3
random_projs = torch.randn(n_random_projections, dim)

# Number of bins for histogram
n_bins = 5

# Optional min_x and max_x - set to None to let function compute them
min_x = None
max_x = None

# Optional weights tensor of shape (batch_size,)
weights = None

# Whether to return min_max values
return_min_max = False

ARGS = {
    "samples": samples,
    "random_projs": random_projs,
    "n_bins": n_bins,
    "min_x": min_x,
    "max_x": max_x,
    "weights": weights,
    "return_min_max": return_min_max,
}
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 (3, 5) — shape (3, 5) 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
compute_sliced_ks Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/metrics/ks.py
code served (permissive licence) · get_code("4dd14b25d03f623f")
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 float, ndim 0, dtype family float, shape () 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

torch.manual_seed(0)

# Create small sample tensors for two distributions
batch_size_1 = 2
batch_size_2 = 3
dim = 4

samples1 = torch.randn(batch_size_1, dim)
samples2 = torch.randn(batch_size_2, dim)

# Optional weights for samples2
weights = torch.ones(batch_size_2)

# Other parameters
n_random_projections = 8
n_bins = 16

ARGS = {
    "samples1": samples1,
    "samples2": samples2,
    "weights": weights,
    "n_random_projections": n_random_projections,
    "n_bins": n_bins,
}
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 float, ndim 0, dtype family float, shape () — shape () 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
compute_value_and_grad_and_laplacian_fn Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/ebm.py
code served (permissive licence) · get_code("30f1c3a3a2ba56d0")
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 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 torch

torch.manual_seed(0)

def value_and_grad_f_non_batched(t, x):
    """
    Mock function that returns (grad, value) for a non-batched input.
    t: scalar tensor
    x: 1D tensor of shape (4,)
    Returns: (grad of shape (4,), value scalar)
    """
    # Simple quadratic function: f(x) = sum(x^2)
    value = torch.sum(x ** 2)
    grad = 2 * x
    return grad, value

ARGS = {
    "value_and_grad_f_non_batched": value_and_grad_f_non_batched,
}
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 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
compute_value_and_value_d_t Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/ebm.py
code served (permissive licence) · get_code("ffbc442b55fe8cce")
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 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 torch

torch.manual_seed(0)

# f_non_batched should be a function that:
# 1. Takes two arguments (t, x)
# 2. Returns a scalar value
# 3. Is differentiable with respect to the first argument (t)

def f_non_batched(t, x):
    """
    A simple function that takes time t and state x,
    and returns a scalar that depends on both.
    """
    # t is scalar, x is a vector
    # Return a scalar value
    return torch.sum(t * x ** 2)

ARGS = {
    "f_non_batched": f_non_batched,
}
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 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
cov_of_mog Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/distr/utils.py
code served (permissive licence) · get_code("944ffc1e825b0214")
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 1, dtype family float, shape (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

torch.manual_seed(0)

# Build arguments for cov_of_mog function
batch_size = 2
dim = 4

# weights: shape (batch_size,) - mixture weights that sum to 1
weights = torch.tensor([0.3, 0.7])

# means: shape (batch_size, dim) - means of each component
means = torch.randn(batch_size, dim)

# variances: shape (batch_size, dim, dim) - covariance matrices for each component
# For return_diag=True case, can also be shape (batch_size, dim)
# We'll use the full covariance form (batch_size, dim, dim)
variances = torch.randn(batch_size, dim, dim)

# Make variances symmetric positive semi-definite by doing A @ A.T
variances = torch.matmul(variances, variances.transpose(1, 2))

ARGS = {
    "weights": weights,
    "means": means,
    "variances": variances,
    "return_diag": True,
    "return_mean": 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, ndim 1, dtype family float, shape (4,) — shape (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
cov_of_mog_batch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
h2o64/dabg/dabg/distr/utils.py
code served (permissive licence) · get_code("eb1c1a67385d4822")
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 torch

torch.manual_seed(0)

# Batch size: 2
# Dimension: 4
# Number of mixture components: 3

batch_size = 2
num_components = 3
dim = 4

# weights: shape (batch_size, num_components)
# Should sum to 1 for each batch element
weights = torch.ones(batch_size, num_components) / num_components

# means: shape (batch_size, num_components, dim)
means = torch.randn(batch_size, num_components, dim)

# variances: shape (batch_size, num_components, dim)
# Diagonal covariances (one variance per dimension per component)
variances = torch.abs(torch.randn(batch_size, num_components, dim)) + 0.1

ARGS = {
    "weights": weights,
    "means": means,
    "variances": variances,
    "return_diag": True,
    "return_mean": 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, 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
diag_standardize_gauss_params Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/distr/gauss.py
code served (permissive licence) · get_code("90bfee50e7465f2e")
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)

# Dimension parameters
dim = 4
n_modes = 2

# Create scaling vector S of shape (dim,)
S = torch.ones(dim) * 2.0

# Create shift vector b of shape (dim,)
b = torch.ones(dim) * 0.5

# Create means of shape (n_modes, dim)
means = torch.randn(n_modes, dim)

# Create covariances of shape (n_modes, dim, dim) - full covariance matrices
covariances = torch.eye(dim).unsqueeze(0).expand(n_modes, -1, -1).clone()

ARGS = {
    "S": S,
    "b": b,
    "means": means,
    "covariances": covariances,
    "upper": 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 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
full_standardize_gauss_params Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/distr/gauss.py
code served (permissive licence) · get_code("0c17cc6547d5135b")
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)

dim = 4
n_modes = 2

# Create a lower triangular matrix L
L = torch.tril(torch.randn(dim, dim))
# Ensure L is invertible by making diagonal non-zero
L = L + torch.eye(dim) * 2

# Create shift vector b
b = torch.randn(dim)

# Create means: (n_modes, dim)
means = torch.randn(n_modes, dim)

# Create covariances: (n_modes, dim, dim)
# Make them symmetric positive definite
covariances = torch.randn(n_modes, dim, dim)
covariances = covariances @ covariances.transpose(-1, -2)  # Make symmetric PD

ARGS = {
    "L": L,
    "b": b,
    "means": means,
    "covariances": covariances,
    "upper": 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 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
gauss_kernel Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/metrics/mmd.py
code served (permissive licence) · get_code("dc8452534f652dbb")
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, 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 torch

torch.manual_seed(0)

# Create a small pairwise squared distance matrix
# Shape: (2, 3) - representing pairwise distances between 2 points in some space
pairwise_sq_matrix = torch.tensor([
    [0.5, 1.2, 2.0],
    [1.2, 0.3, 1.5]
], dtype=torch.float32)

# Create a small bandwidth squared parameter
bandwidth_sq = torch.tensor(0.8, dtype=torch.float32)

ARGS = {
    "pairwise_sq_matrix": pairwise_sq_matrix,
    "bandwidth_sq": bandwidth_sq
}
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, 3) — shape (2, 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
heuristics_step_size Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/mcmc/utils.py
code served (permissive licence) · get_code("e98548faaafafd2a")
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 torch
import math

torch.manual_seed(0)

# Construct arguments for heuristics_step_size
# The function expects:
# - stepsize: a tensor (used with torch.where and shape operations)
# - mean_log_acceptance: a tensor (used in comparisons)
# - target_acceptance: float (default 0.75)
# - factor: float (default 1.01)
# - tol: float (default 0.05)

# Create a batch of step sizes with shape (2, 4) - 2 chains, 4 parameters
stepsize = torch.ones(2, 4) * 0.1

# Create mean log acceptance rates with shape (2,) - one per chain
mean_log_acceptance = torch.tensor([-0.2, -0.5])

# Standard parameters
target_acceptance = 0.75
factor = 1.01
tol = 0.05

ARGS = {
    "stepsize": stepsize,
    "mean_log_acceptance": mean_log_acceptance,
    "target_acceptance": target_acceptance,
    "factor": factor,
    "tol": tol,
}
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
kaiming_normal_zeros_ Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/utils.py
code served (permissive licence) · get_code("050f49ca2d4ec694")
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 (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 torch
import torch.nn as nn
import math

torch.manual_seed(0)

init_weight_scale = 1e-06

def kaiming_normal_zeros_(m):
    return torch.nn.init.kaiming_normal_(m, a=math.sqrt(6.0 / init_weight_scale ** 2 - 1))

# The function expects a tensor or tensor-like object with in_features and out_features
# or at least a tensor that can be passed to kaiming_normal_
# Looking at the function body, it calls torch.nn.init.kaiming_normal_(m, a=...)
# kaiming_normal_ expects a tensor (typically a weight tensor from a layer)

# Create a simple weight tensor
weight_tensor = torch.randn(4, 3)

ARGS = {
    "m": weight_tensor
}
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 (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
kaiming_uniform_zeros_ Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/networks/utils.py
code served (permissive licence) · get_code("dffbfb0a413afade")
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 (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 torch
import torch.nn as nn
import math

torch.manual_seed(0)

init_weight_scale = 1e-06

def kaiming_uniform_zeros_(m):
    return torch.nn.init.kaiming_uniform_(m, a=math.sqrt(6.0 / init_weight_scale ** 2 - 1))

# The function expects a tensor or a module with a weight tensor
# kaiming_uniform_ modifies a tensor in-place, so we create a simple tensor
test_tensor = torch.empty(4, 3)

ARGS = {
    "m": test_tensor
}
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 (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
log_prob_multivariate_normal_diag Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/mcmc/utils.py
code served (permissive licence) · get_code("611d01886cc9b963")
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 1, dtype family float, shape (2,) 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

torch.manual_seed(0)

# Create small tensors for testing
batch_size = 2
dim = 4

# samples: tensor of shape (batch_size, dim) representing sampled values
samples = torch.randn(batch_size, dim)

# mean: tensor of shape (batch_size, dim) representing the mean of the distribution
mean = torch.randn(batch_size, dim)

# variance: can be a scalar or a tensor
# Testing with a scalar variance
variance = torch.tensor(1.0)

# sum_indexes: tuple of dimensions to sum over
# Typically this would be the dimensions over which to aggregate the log probability
sum_indexes = (1,)

ARGS = {
    "samples": samples,
    "mean": mean,
    "variance": variance,
    "sum_indexes": sum_indexes,
}
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 1, dtype family float, shape (2,) — shape (2,) 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
mean_of_mog Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/distr/utils.py
code served (permissive licence) · get_code("5994f86d61124681")
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 1, dtype family float, shape (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

torch.manual_seed(0)

# weights: should be a 1D tensor of shape (num_components,)
# means: should be a 2D tensor of shape (num_components, dim)
# The function unsqueezes weights to (num_components, 1), multiplies with means,
# and sums along dim 0, resulting in output of shape (dim,)

weights = torch.tensor([0.3, 0.7], dtype=torch.float32)  # 2 components
means = torch.randn(2, 4)  # 2 components, 4 dimensions

ARGS = {
    "weights": weights,
    "means": means,
}
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 1, dtype family float, shape (4,) — shape (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_multivariate_normal_diag Ran
ran; our contract draft was wrong, not the code
h2o64/dabg/dabg/mcmc/utils.py
code served (permissive licence) · get_code("76fbbcaa72d4be50")
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 torch
import math

torch.manual_seed(0)

# Construct arguments for sample_multivariate_normal_diag
batch_size = 2
mean = torch.randn(2, 4)  # shape (batch_size, feature_dim)
variance_tensor = torch.ones(2, 4) * 0.5  # tensor variance

ARGS = {
    "batch_size": batch_size,
    "mean": mean,
    "variance": variance_tensor,
}
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
compute_pairwise_sq_distance_diff Not yet run
unverified (harvested, no recorded run)
h2o64/dabg/dabg/metrics/mmd.py
code served (permissive licence) · get_code("81772ea13e2c90e8")
How it ran

no invocation recorded.

mala_step Not yet run
unverified (harvested, no recorded run)
h2o64/dabg/dabg/mcmc/samplers.py
code served (permissive licence) · get_code("d63a5f3defe5b8b1")
How it ran

no invocation recorded.

rejection_sampling Not yet run
unverified (harvested, no recorded run)
h2o64/dabg/dabg/distr/manywell.py
code served (permissive licence) · get_code("612d95cca0e9d605")
How it ran

no invocation recorded.

rwmh_step Not yet run
unverified (harvested, no recorded run)
h2o64/dabg/dabg/mcmc/samplers.py
code served (permissive licence) · get_code("f0d257a9370efb5e")
How it ran

no invocation recorded.

ula_step Not yet run
unverified (harvested, no recorded run)
h2o64/dabg/dabg/mcmc/samplers.py
code served (permissive licence) · get_code("4fb1af22577efcaa")
How it ran

no invocation recorded.

Also present in other papers' code

Of 23 samples lifted from this paper's linked repositories, 23 carry a function identity and 0 do not. Of the 23 distinct functions with an identity, 1 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 1 later) and 22 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.

FunctionAlso present inEarliest-datedOthers named
kaiming_uniform_zeros_ 1 other paper
1 later
2607.16987 (2026-07)
Twisted Schrödinger Bridge Matching

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

Sampling configurations at thermodynamic equilibrium is a central challenge in statistical physics. Boltzmann Generators (BGs) address this problem by pairing a generative model with a Monte Carlo (MC) correction scheme, yielding asymptotically consistent samples from an unnormalized target density. However, most existing BGs rely on classic MC mechanisms such as importance sampling, which (i) impose strong constraints on the backbone model (typically requiring exact and efficient likelihood evaluation) and (ii) suffer from severe scalability issues in high-dimensional, multi-modal settings. This work investigates BGs built around annealed Monte Carlo (aMC) schemes, which mitigate the limitations of classic MC by bridging a simple reference distribution to the target through a sequence of intermediate densities. In this context, diffusion models (DMs) are particularly appealing backbones: they are powerful generative models and naturally induce density paths that have been leveraged in prior aMC-based methods. We provide an empirical meta-analysis of this DM-based aMC-BG design choice on controlled yet challenging synthetic benchmarks based on multi-modal Gaussian mixtures, varying inter-mode separation, number of modes, and dimensionality. To disentangle learning effects from inference effects, we first study an idealized setting in which the DM is perfectly learned, and then turn to realistic settings where the DM is trained from data. Even in the idealized regime, we find that standard aMC integrations of DMs that rely only on first-order stochastic denoising kernels systematically fail in the proposed scenarios. In contrast, incorporating second-order denoising kernels can substantially improve performance when the required covariance information is available. Motivated by this gap, we propose an alternative aMC integration based on deterministic first-order transport maps derived from DMs; empirically, this approach consistently outperforms its stochastic first-order counterpart, albeit at increased computational cost. Overall, while results in the perfect-learning regime suggest that exploiting DM-induced dynamics within aMC is a promising route to building effective BGs, our experiments with learned DMs show that DM-aMC combinations still struggle to produce accurate BGs in practice. We attribute this limitation primarily to inaccuracies in DM log-density estimation. Code available at https://github.com/h2o64/dabg.

For agents

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

get_harvested_code_for_paper("2601.21026")
get_code_for_paper("2601.21026")
have("2601.21026")

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

curl https://syntology.ai/api/ran/2601.21026.json

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

[![Syntology run record](https://syntology.ai/api/ran/2601.21026.svg)](https://syntology.ai/paper/2601.21026)

Connect an agent — have() is free.