SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.01006 · 2025

Underdamped Diffusion Bridges with Applications to Sampling

arXiv · PDF · Open in the Atlas

Code that ran

10 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 9 have no contract verdict; 1 failure of our instrument; 0 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/2503.01006.json · badge: /api/ran/2503.01006.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
DenisBless/UnderdampedDiffusionBridges canonical (classifier label)
classifier basis not recorded on the link
10 ran · 0 no recorded run
FunctionStatusWhere it lives
check_stop_grad Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/utils.py
code served (permissive licence) · get_code("29ab22cf1e391158")
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 ArrayImpl, 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 jax
import jax.numpy as jnp

def check_stop_grad(expression, stop_grad):
    return jax.lax.stop_gradient(expression) if stop_grad else expression

# Create a simple JAX array as the expression
expression = jnp.array([1.0, 2.0, 3.0, 4.0])

# Test with stop_grad=True
stop_grad_true = True

ARGS = {
    "expression": expression,
    "stop_grad": stop_grad_true,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, 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
get_cosine_schedule Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/scheduler.py
code served (permissive licence) · get_code("3a7853161909da1d")
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 jax.numpy as jnp

ARGS = {
    "total_steps": 100,
    "min": 0.01,
    "s": 0.008,
    "pow": 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — 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
get_learning_rate_scheduler Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/learning_rate_scheduler.py
code served (permissive licence) · get_code("896621e81a8c000f")
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 optax

# Create a minimal config dict that satisfies the function's requirements
cfg = {
    'common': {
        'warmup': 'linear',
        'warmup_iters': 100,
        'iters': 1000,
    }
}

# step_size is a learning rate (float)
step_size = 0.001

# Map all parameters to their values
ARGS = {
    'cfg': cfg,
    'step_size': step_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
optax — 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
get_linear_schedule Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/scheduler.py
code served (permissive licence) · get_code("d85b357fbdeb22d7")
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 numpy as np

torch.manual_seed(0)

ARGS = {
    "total_steps": 100,
    "min": 0.01
}
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 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
init_langevin Ran
ran; our contract draft was wrong, not the code
DenisBless/UnderdampedDiffusionBridges/algorithms/common/init_diffusion_model.py
pointer only (licence: MIT); the stored copy get_code serves, from lizr16/OMAD, is pointer only (licence: NONE) · get_code("d443d828e92a9ae7")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from lizr16/OMAD, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

init_underdamped Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/init_diffusion_model.py
code served (permissive licence) · get_code("ec4e879e1a90f33d")
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.

import jax
import jax.numpy as jnp
import distrax

# Define a simple config class
class AlgorithmConfig:
    def __init__(self):
        self.learn_prior = True
        self.learn_dt = True
        self.per_step_dt = False
        self.learn_friction = True
        self.learn_mass_matrix = True
        self.init_std = 1.0
        self.dt_schedule = lambda step: 1.0

class Config:
    def __init__(self):
        self.algorithm = AlgorithmConfig()

# Create config
cfg = Config()

# Set dimension
dim = 4

# Create params dict with required structure
key = jax.random.PRNGKey(0)
params = {
    'params': {
        'prior_mean': jnp.zeros(dim),
        'prior_std': jnp.zeros(dim),
        'dt': jnp.array([0.1, 0.1, 0.1, 0.1]) if cfg.algorithm.per_step_dt else jnp.array(0.1),
        'friction': jnp.zeros(dim),
        'mass_std': jnp.zeros(dim),
    }
}

ARGS = {
    'cfg': cfg,
    'dim': dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
distrax, jax — 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
neg_elbo Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/overdamped/overdamped_rnd.py
code served (permissive licence) · get_code("0a3afc9623a6941a")
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 ArrayImpl, 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 jax
import jax.numpy as jnp

class DiffusionModel:
    def __init__(self, num_steps=5, dim=4):
        self.num_steps = num_steps
        self.dim = dim
    
    def prior_sampler(self, params, key, batch_size):
        # Returns shape (batch_size, dim)
        return jax.random.normal(key, (batch_size, self.dim))
    
    def prior_log_prob(self, x, params):
        # Returns scalar log prob for a single sample
        return -0.5 * jnp.sum(x ** 2)
    
    def target_log_prob(self, x):
        # Returns scalar log prob for a single sample
        return -0.5 * jnp.sum((x - 1.0) ** 2)


class ModelState:
    def __init__(self):
        self.step = 0


def integrator_fn(model_state, params, stop_grad, eval):
    def integrate_step(aux, t):
        init_x, log_ratio, key = aux
        # Simple dummy integrator that doesn't change x much
        key, subkey = jax.random.split(key)
        noise = jax.random.normal(subkey, init_x.shape) * 0.01
        new_x = init_x + noise
        new_log_ratio = log_ratio + 0.1
        
        new_aux = (new_x, new_log_ratio, key)
        per_step_output = new_x if eval else None
        return new_aux, per_step_output
    
    return integrate_step


# Setup arguments
key = jax.random.PRNGKey(0)
model_state = ModelState()
params = {}
integrator = integrator_fn
diffusion_model = DiffusionModel(num_steps=5, dim=4)
batch_size = 2

ARGS = {
    "key": key,
    "model_state": model_state,
    "params": params,
    "integrator": integrator,
    "diffusion_model": diffusion_model,
    "batch_size": batch_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, 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
per_sample_rnd Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/overdamped/overdamped_rnd.py
code served (permissive licence) · get_code("866334c87a4478ef")
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 jax
import jax.numpy as jnp
import jax.random as random

# Define stand-in classes for diffusion_model
class DiffusionModel:
    def __init__(self, num_steps, state_dim):
        self.num_steps = num_steps
        self.state_dim = state_dim
    
    def prior_sampler(self, params, key, batch_size):
        # Return samples from prior with shape (batch_size, state_dim)
        return random.normal(key, (batch_size, self.state_dim))
    
    def prior_log_prob(self, x, params):
        # Return log probability of prior (scalar)
        return jnp.sum(-0.5 * x**2)
    
    def target_log_prob(self, x):
        # Return log probability of target (scalar)
        return jnp.sum(-0.5 * x**2) + 1.0

# Define model_state (can be a simple dict or container)
class ModelState:
    def __init__(self):
        self.value = 0.0

# Define params (can be a simple dict or container)
class Params:
    def __init__(self):
        self.scale = 1.0

# Define integrator function
def create_integrator(num_steps):
    def integrator(model_state, params, stop_grad, eval):
        def step(aux, t):
            init_x, log_ratio, key = aux
            key, subkey = random.split(key)
            # Simulate a step that updates x slightly
            noise = random.normal(subkey, init_x.shape) * 0.01
            new_x = init_x + noise
            new_log_ratio = log_ratio + 0.1
            new_aux = (new_x, new_log_ratio, key)
            # per_step_output is collected via scan
            per_step_output = jnp.expand_dims(new_x, 0)
            return new_aux, per_step_output
        return step
    return integrator

# Create instances
seed = random.PRNGKey(0)
state_dim = 4
num_steps = 3

model_state = ModelState()
params = Params()
diffusion_model = DiffusionModel(num_steps=num_steps, state_dim=state_dim)
integrator = create_integrator(num_steps)

ARGS = {
    "seed": seed,
    "model_state": model_state,
    "params": params,
    "integrator": integrator,
    "diffusion_model": diffusion_model,
    "stop_grad": False,
    "eval": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — 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
rnd Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/overdamped/overdamped_rnd.py
code served (permissive licence) · get_code("8cce7df513db5677")
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 jax
import jax.numpy as jnp
import jax.random as random

# Define a minimal diffusion model that satisfies the interface used
class MinimalDiffusionModel:
    def __init__(self, num_steps=5, state_dim=4):
        self.num_steps = num_steps
        self.state_dim = state_dim
    
    def prior_sampler(self, params, key, batch_size):
        # Returns samples of shape (batch_size, state_dim)
        return random.normal(key, (batch_size, self.state_dim))
    
    def prior_log_prob(self, x, params):
        # Returns a scalar log probability
        return -0.5 * jnp.sum(x ** 2)
    
    def target_log_prob(self, x):
        # Returns a scalar log probability
        return -0.5 * jnp.sum((x - 1.0) ** 2)

# Define a minimal integrator function
def minimal_integrator(model_state, params, stop_grad, eval):
    def integrate_step(aux, step):
        init_x, log_ratio, key = aux
        key, subkey = random.split(key)
        # Simple integration: add small noise
        noise = random.normal(subkey, init_x.shape) * 0.01
        next_x = init_x + noise
        next_log_ratio = log_ratio + 0.1
        new_aux = (next_x, next_log_ratio, key)
        per_step_output = next_x if eval else None
        return new_aux, per_step_output
    return integrate_step

# Create minimal model_state
class MinimalModelState:
    pass

# Set up the arguments
key = random.PRNGKey(0)
model_state = MinimalModelState()
params = {}  # Empty dict for parameters
integrator = minimal_integrator
diffusion_model = MinimalDiffusionModel(num_steps=5, state_dim=4)
batch_size = 2
stop_grad = False
eval = False

ARGS = {
    'key': key,
    'model_state': model_state,
    'params': params,
    'integrator': integrator,
    'diffusion_model': diffusion_model,
    'batch_size': batch_size,
    'stop_grad': stop_grad,
    'eval': eval
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — 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
sample_kernel Ran
ran on a synthesized input
DenisBless/UnderdampedDiffusionBridges/algorithms/common/utils.py
code served (permissive licence) · get_code("80005c2f78f46bf7")
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 ArrayImpl, 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 jax
import jax.numpy as jnp

# Set seed for reproducibility
jax.config.update("jax_enable_custom_prng", True)

# Create a PRNGKey for JAX
rng_key = jax.random.PRNGKey(0)

# Create mean and scale as JAX arrays
mean = jnp.array([1.0, 2.0, 3.0, 4.0])
scale = jnp.array([0.5, 0.5, 0.5, 0.5])

ARGS = {
    "rng_key": rng_key,
    "mean": mean,
    "scale": scale,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, 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

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 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 9 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
init_langevin 1 other paper
1 later
2602.18291 (2026-02)
Diffusing to Coordinate: Efficient Online Multi-Agent Diffusion Policies

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

We provide a general framework for learning diffusion bridges that transport prior to target distributions. It includes existing diffusion models for generative modeling, but also underdamped versions with degenerate diffusion matrices, where the noise only acts in certain dimensions. Extending previous findings, our framework allows to rigorously show that score matching in the underdamped case is indeed equivalent to maximizing a lower bound on the likelihood. Motivated by superior convergence properties and compatibility with sophisticated numerical integration schemes of underdamped stochastic processes, we propose \emph{underdamped diffusion bridges}, where a general density evolution is learned rather than prescribed by a fixed noising process. We apply our method to the challenging task of sampling from unnormalized densities without access to samples from the target distribution. Across a diverse range of sampling problems, our approach demonstrates state-of-the-art performance, notably outperforming alternative methods, while requiring significantly fewer discretization steps and no hyperparameter tuning.

For agents

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

get_harvested_code_for_paper("2503.01006")
get_code_for_paper("2503.01006")
have("2503.01006")

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

curl https://syntology.ai/api/ran/2503.01006.json

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

[![Syntology run record](https://syntology.ai/api/ran/2503.01006.svg)](https://syntology.ai/paper/2503.01006)

Connect an agent — have() is free.