SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2512.04341 · ICML · 2025

Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism

Esther Derman, Tianwei Ni, Pierre-Luc Bacon, Siamak Ravanbakhsh, Vineet Jain, Vincent Taboga

arXiv · PDF · Open in the Atlas

Code that ran

16 of 23 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (3 of them constructed an object rather than computing a result; 3 honoured the contract we drafted, 1 violated it and 3 have no contract verdict; 9 failures of our instrument; 7 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/2512.04341.json · badge: /api/ran/2512.04341.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
twni2016/neubay canonical (classifier label)
classifier basis: census
6 ran · 3 no recorded run
Howuhh/sac-n-jax canonical (classifier label)
classifier basis: census
4 ran · 2 no recorded run
proroklab/memoroids community
Papers with Code archive
6 ran · 2 no recorded run
FunctionStatusWhere it lives
AtariCNN Ran
ran on a synthesized input · constructed an object; no result was computed
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("9a632502a83cb77d")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

Block Ran
ran on a synthesized input · constructed an object; no result was computed
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("0cd85f17db6bef74")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

QHead Ran
ran on a synthesized input · constructed an object; no result was computed
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("40cf3544bbd2c834")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

default_init Ran
ran; our contract draft was wrong, not the code
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("e31de52332dff52c")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

final_linear Ran
ran; our contract draft was wrong, not the code
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("9bcd87ecfba6af31")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

fix_numerical_issue Ran
ran, honoured the contract we drafted
twni2016/neubay/eval_error.py
code served (permissive licence) · get_code("2897edd398c620dc")
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 ndarray, ndim 2, dtype family float, shape (5, 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 numpy as np
import pandas as pd

# Create test data with various numerical issues
# Including positive infinity, negative infinity, and NaN values
data = np.array([
    [1.0, 2.0, 3.0, 4.0],
    [np.inf, 6.0, 7.0, 8.0],
    [9.0, np.nan, 11.0, 12.0],
    [-np.inf, 14.0, 15.0, 16.0],
    [17.0, 18.0, 19.0, 20.0]
], dtype=np.float32)

ARGS = {
    "data": data
}
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 ndarray, ndim 2, dtype family float, shape (5, 4) — shape (5, 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_nll_and_unc Ran
ran; our contract draft was wrong, not the code
twni2016/neubay/offline_world/losses.py
code served (permissive licence) · get_code("a0975bd6994f79e8")
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 equinox as eqx


class SimpleModel(eqx.Module):
    """Minimal model that satisfies the function's usage requirements."""
    
    def forward_same_data(self, x):
        """
        Returns (mu, sigma), unc where:
        - mu and sigma are outputs with shape matching the function's expectations
        - unc is a pytree that can be mapped with jnp.mean
        """
        batch_size = x.shape[0]
        # mu and sigma should have compatible shapes for the NLL computation
        # y_expanded will have shape (1, batch_size, ...)
        # So mu and sigma should broadcast with that
        mu = jnp.ones((1, batch_size, 3))
        sigma = jnp.ones((1, batch_size, 3)) * 0.5
        
        # unc is a pytree that gets mapped with jnp.mean
        unc = {
            'aleatoric': jnp.ones((batch_size, 4)),
            'epistemic': jnp.ones((batch_size, 4)) * 0.1,
        }
        
        return (mu, sigma), unc


# Create minimal tensors
x = jnp.ones((2, 5))  # batch_size=2, input_dim=5
y = jnp.ones((2, 3))  # batch_size=2, output_dim=3

# Create model instance
model = SimpleModel()

ARGS = {
    'model': model,
    'x': x,
    'y': y,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
equinox, 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
leaky_relu Ran
ran, honoured the contract we drafted
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("764e3c422e0df809")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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.

sac_alpha_loss Ran
ran; our contract draft was wrong, not the code
twni2016/neubay/online_rl/ac_losses.py
code served (permissive licence) · get_code("6709bbfd7a31a9c7")
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
from jax import numpy as jnp


class AlphaModule:
    """Mock alpha module that's callable and has target_entropy attribute."""
    def __init__(self, value=0.5, target_entropy=None):
        self.value = value
        self.target_entropy = target_entropy
    
    def __call__(self):
        return jnp.array(self.value)


# Create test data
batch_size = 2
seq_length = 3

# alpha: a callable object with target_entropy attribute
alpha = AlphaModule(value=0.3, target_entropy=-1.0)

# entropy: JAX array of shape (batch_size, seq_length)
entropy = jnp.array([
    [0.5, 0.6, 0.7],
    [0.4, 0.5, 0.6]
])

# tape: dict with 'start' and 'real' keys
# 'start' shape: (batch_size, seq_length)
# 'real' shape: (batch_size, seq_length)
tape = {
    'start': jnp.array([
        [1.0, 0.0, 0.0],
        [1.0, 0.0, 0.0]
    ]),
    'real': jnp.array([
        [0.8, 1.0, 0.0],
        [1.0, 0.9, 0.0]
    ])
}

# real_weight: scalar float between 0 and 1
real_weight = 0.6

ARGS = {
    'alpha': alpha,
    'entropy': entropy,
    'tape': tape,
    'real_weight': real_weight
}
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, 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
soft_clamp Ran
ran, honoured the contract we drafted
twni2016/neubay/offline_world/modules.py
code served (permissive licence) · get_code("f92e635d667b3a0c")
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 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 jax
import jax.numpy as jnp

# Create a simple JAX array with some values
x = jnp.array([[-2.0, 0.5, 1.5, 3.0],
               [-1.0, 0.0, 2.0, 5.0]])

_min = -1.0
_max = 2.0

ARGS = {
    "x": x,
    "_min": _min,
    "_max": _max,
}
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 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
subselect_members Ran
ran, violated the contract we drafted
twni2016/neubay/offline_world/modules.py
code served (permissive licence) · get_code("a6b12c3466cb50c9")
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 jax
import jax.numpy as jnp

# Create a tree structure with various leaves
# The function indexes the first axis of arrays where applicable

# Create some sample arrays with different shapes
array_1d = jnp.arange(10)  # 1D array - should NOT be indexed (ndim == 1 fails condition)
array_2d = jnp.arange(20).reshape(5, 4)  # 5x4 array - will be indexed on axis 0
array_3d = jnp.arange(60).reshape(5, 3, 4)  # 5x3x4 array - will be indexed on axis 0
scalar = jnp.array(42.0)  # scalar - should NOT be indexed (ndim == 0)
small_array = jnp.array([1.0, 2.0])  # small array with shape (2,) - should NOT be indexed

# Create a nested tree structure (dict and list combination)
tree = {
    "tensors": [array_2d, array_3d],
    "nested": {
        "scalar": scalar,
        "small": small_array,
        "one_d": array_1d,
    },
    "direct": array_2d,
}

# Index array to select specific elements from axis 0
# Must be valid indices for the arrays (0-4 range for 5-element axis 0)
idx = jnp.array([0, 2, 4])

ARGS = {
    "tree": tree,
    "idx": idx,
}
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 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
uniform_init Ran
ran; our contract draft was wrong, not the code
Howuhh/sac-n-jax/sac_n_jax_flax.py
code served (permissive licence) · get_code("b5491b8fa060cb74")
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
import jax.numpy as jnp

def uniform_init(bound: float):
    def _init(key, shape, dtype):
        return jax.random.uniform(key, shape=shape, minval=-bound, maxval=bound, dtype=dtype)
    return _init

# Test the function to understand what it does
# It takes a bound parameter and returns an initialization function
# The initialization function takes (key, shape, dtype) and returns random uniform values

ARGS = {
    "bound": 0.5
}
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
update_actor Ran
ran; our contract draft was wrong, not the code
Howuhh/sac-n-jax/sac_n_jax_flax.py
code served (permissive licence) · get_code("a9b0dbc1f6e0b3fb")
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
from flax import linen as nn
from flax.training.train_state import TrainState
import optax

class Actor(nn.Module):
    @nn.compact
    def __call__(self, obs):
        x = nn.Dense(4)(obs)
        x = nn.relu(x)
        x = nn.Dense(8)(x)
        return DiagGaussianDist(x)

class Critic(nn.Module):
    @nn.compact
    def __call__(self, obs, actions):
        x = jnp.concatenate([obs, actions], axis=-1)
        x = nn.Dense(4)(x)
        x = nn.relu(x)
        x = nn.Dense(2)(x)
        return x

class Alpha(nn.Module):
    @nn.compact
    def __call__(self):
        return jnp.array(0.2)

class DiagGaussianDist:
    def __init__(self, params):
        self.params = params
    
    def sample_and_log_prob(self, seed):
        mean = self.params[..., :4]
        logstd = self.params[..., 4:]
        std = jnp.exp(logstd)
        noise = jax.random.normal(seed, mean.shape)
        actions = mean + noise * std
        logp = -0.5 * ((actions - mean) / std) ** 2 - logstd - 0.5 * jnp.log(2 * jnp.pi)
        return actions, logp

key = jax.random.PRNGKey(0)
key_init = jax.random.PRNGKey(1)

obs_dim = 4
action_dim = 4
batch_size = 2

actor_model = Actor()
critic_model = Critic()
alpha_model = Alpha()

actor_params = actor_model.init(key_init, jnp.ones((batch_size, obs_dim)))
critic_params = critic_model.init(key_init, jnp.ones((batch_size, obs_dim)), jnp.ones((batch_size, action_dim)))
alpha_params = alpha_model.init(key_init)

actor_tx = optax.adam(learning_rate=1e-3)
critic_tx = optax.adam(learning_rate=1e-3)
alpha_tx = optax.adam(learning_rate=1e-3)

actor = TrainState.create(
    apply_fn=actor_model.apply,
    params=actor_params,
    tx=actor_tx
)

critic = TrainState.create(
    apply_fn=critic_model.apply,
    params=critic_params,
    tx=critic_tx
)

alpha = TrainState.create(
    apply_fn=alpha_model.apply,
    params=alpha_params,
    tx=alpha_tx
)

batch = {
    'obs': jnp.ones((batch_size, obs_dim)),
}

ARGS = {
    'key': key,
    'actor': actor,
    'critic': critic,
    'alpha': alpha,
    'batch': batch,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, optax — 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
update_alpha Ran
ran; our contract draft was wrong, not the code
Howuhh/sac-n-jax/sac_n_jax_eqx.py
code served (permissive licence) · get_code("2be8994bb96607b4")
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 optax
import dataclasses
import equinox as eqx
from typing import Dict, Any, Tuple


class SimpleAlphaModel(eqx.Module):
    """A simple model that outputs a scalar alpha value."""
    weight: jax.Array
    
    def __init__(self):
        self.weight = jnp.array([0.5])
    
    def __call__(self):
        return self.weight


class TrainState(eqx.Module):
    model: eqx.Module
    optim: optax.GradientTransformation
    optim_state: optax.OptState

    @classmethod
    def create(cls, *, model, optim, **kwargs):
        optim_state = optim.init(eqx.filter(model, eqx.is_array))
        return cls(model, optim, optim_state, **kwargs)

    def apply_updates(self, grads):
        updates, new_optim_state = self.optim.update(grads, self.optim_state)
        new_model = eqx.apply_updates(self.model, updates)
        return dataclasses.replace(self, model=new_model, optim_state=new_optim_state)


# Create the model
alpha_model = SimpleAlphaModel()

# Create optimizer
optim = optax.adam(learning_rate=1e-3)

# Create TrainState
alpha_state = TrainState.create(model=alpha_model, optim=optim)

# Create entropy array (batch of 2 entropy values, dim 4)
entropy = jnp.array([[0.5, 0.6, 0.7, 0.8],
                      [0.4, 0.5, 0.6, 0.7]])

# Target entropy scalar
target_entropy = 0.5

ARGS = {
    'alpha': alpha_state,
    'entropy': entropy,
    'target_entropy': target_entropy
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
equinox, jax, optax — 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
update_alpha Ran
ran; our contract draft was wrong, not the code
Howuhh/sac-n-jax/sac_n_jax_flax.py
code served (permissive licence) · get_code("43914e1e4188a279")
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
from flax.training.train_state import TrainState
from flax import linen as nn
import optax


# Define a simple alpha network
class AlphaNetwork(nn.Module):
    @nn.compact
    def __call__(self, x=None):
        # Returns a scalar alpha value
        return nn.Dense(1)(jnp.array([1.0]))


# Create the alpha model
alpha_model = AlphaNetwork()

# Initialize parameters
rng = jax.random.PRNGKey(0)
alpha_params = alpha_model.init(rng, jnp.array([1.0]))

# Create optimizer
optimizer = optax.adam(learning_rate=1e-3)

# Create TrainState
alpha_train_state = TrainState.create(
    apply_fn=alpha_model.apply,
    params=alpha_params,
    tx=optimizer
)

# Create entropy and target_entropy values
entropy = jnp.array(0.5)
target_entropy = jnp.array(0.8)

ARGS = {
    'alpha': alpha_train_state,
    'entropy': entropy,
    'target_entropy': target_entropy
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, optax — 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
update_selected_members Ran
ran; our contract draft was wrong, not the code
twni2016/neubay/offline_world/modules.py
code served (permissive licence) · get_code("22bf8ddfe38b20bc")
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 jax
import jax.numpy as jnp

# Set seed for reproducibility
jax.config.update('jax_default_prng_impl', 'threefry2x32')
key = jax.random.PRNGKey(0)

# Create a mask with 3 ensemble members, selecting the first and third
mask = jnp.array([True, False, True])

# Create dst: a PyTree with arrays that have ensemble axis as leading dimension
dst = {
    'weights': jnp.ones((3, 4)),  # (N, 4) where N=3 is ensemble axis
    'bias': jnp.ones((3, 2)),      # (N, 2) where N=3 is ensemble axis
    'scalar': jnp.array(5.0),      # scalar, no leading ensemble axis
    'nested': {
        'data': jnp.ones((3, 5)),  # (N, 5) where N=3 is ensemble axis
        'const': jnp.array([1, 2, 3])  # shape (3,) but not ensemble axis
    }
}

# Create src: same structure with different values
src = {
    'weights': jnp.ones((3, 4)) * 10.0,
    'bias': jnp.ones((3, 2)) * 20.0,
    'scalar': jnp.array(7.0),
    'nested': {
        'data': jnp.ones((3, 5)) * 30.0,
        'const': jnp.array([10, 20, 30])
    }
}

ARGS = {
    'dst': dst,
    'src': src,
    'mask': mask
}
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 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
RandomSequential Not yet run
unverified (harvested, no recorded run)
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("ac49693a84a1de1c")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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. No invocation recorded.

RecurrentQNetwork Not yet run
unverified (harvested, no recorded run)
proroklab/memoroids/modules.py
pointer only (licence: NONE) · get_code("9b9bf1240105b1d6")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, proroklab/memoroids 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. No invocation recorded.

ensemble_predict Not yet run
unverified (harvested, no recorded run)
Howuhh/sac-n-jax/sac_n_jax_eqx.py
code served (permissive licence) · get_code("4873b7ac678b7322")
How it ran

no invocation recorded.

get_mse_and_unc Not yet run
unverified (harvested, no recorded run)
twni2016/neubay/offline_world/losses.py
code served (permissive licence) · get_code("0424e46f070a475b")
How it ran

no invocation recorded.

get_train_loss Not yet run
unverified (harvested, no recorded run)
twni2016/neubay/offline_world/losses.py
code served (permissive licence) · get_code("6c22b2d26f00f6a8")
How it ran

no invocation recorded.

sac_pi_loss Not yet run
unverified (harvested, no recorded run)
twni2016/neubay/online_rl/ac_losses.py
code served (permissive licence) · get_code("810d7ec6ea4d4cf9")
How it ran

no invocation recorded.

update_actor Not yet run
unverified (harvested, no recorded run)
Howuhh/sac-n-jax/sac_n_jax_eqx.py
code served (permissive licence) · get_code("1058e2af882503c3")
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 (1 with at least one paper dated earlier, 0 the same month, 0 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
uniform_init 1 other paper
1 dated earlier
2110.01548 (2021-10)
Uncertainty-Based Offline Reinforcement Learning with Diversified Q-Ensemble

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

Popular offline reinforcement learning (RL) methods rely on explicit conservatism, penalizing outof-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective for test-time adaptation. By modeling a posterior over world models and training a historydependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty without explicit conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that long-horizon rollouts are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding model errors, yielding our algorithm, NEUBAY, grounded in the NEUtral BAYesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-theart on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods. Code is available at github.com/twni2016/neubay.

For agents

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

get_harvested_code_for_paper("2512.04341")
get_code_for_paper("2512.04341")
have("2512.04341")

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

curl https://syntology.ai/api/ran/2512.04341.json

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

[![Syntology run record](https://syntology.ai/api/ran/2512.04341.svg)](https://syntology.ai/paper/2512.04341)

Connect an agent — have() is free.