SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2301.13616 · ICML · 2023

Anti-Exploration by Random Network Distillation

Sergey Kolesnikov, Vladislav Kurenkov, Alexander Nikulin, Denis Tarasov

arXiv · PDF · Open in the Atlas

Code that ran

20 of 26 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (18 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 18 have no contract verdict; 1 failure of our instrument; 6 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/2301.13616.json · badge: /api/ran/2301.13616.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
tinkoff-ai/sac-rnd canonical (classifier label)
classifier basis: classifier_2025
17 ran · 1 no recorded run
shidilrzf/anti-exploration-rl canonical (classifier label)
classifier basis: classifier_2025
0 ran · 1 no recorded run
zzmtsvv/sac_rnd community
Papers with Code archive
3 ran · 1 no recorded run
zzmtsvv/sac_drnd community
Papers with Code archive
0 ran · 3 no recorded run
FunctionStatusWhere it lives
BilinearDense Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("d69bcc085240a24a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, BilinearDense: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Create a simple test setup for BilinearDense
# The function takes x and z as inputs in __call__
# x should have shape (batch, feature_dim)
# z should have shape (batch, z_dim)
# out_dim is a required parameter

# Define small test tensors
batch_size = 2
x_dim = 4
z_dim = 3
out_dim = 5

# Create dummy inputs as JAX arrays
x = jnp.ones((batch_size, x_dim))
z = jnp.ones((batch_size, z_dim))

# Create ARGS dict with all parameters
ARGS = {
    "out_dim": out_dim,
    "kernel_init": nn.initializers.lecun_normal(),
    "bias_init": nn.initializers.zeros,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type BilinearDense — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BilinearDense 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
BilinearFirstMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("8f6f4cff35de62cf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, BilinearFirstMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
from jax import random
import math

default_bias_init = nn.initializers.zeros
default_kernel_init = nn.initializers.lecun_normal()

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

# Create minimal JAX arrays for testing
key = random.PRNGKey(0)
batch_size = 2
feature_dim = 4
context_dim = 3

# Dummy feature and context inputs
feature = jnp.ones((batch_size, feature_dim), dtype=jnp.float32)
context = jnp.ones((batch_size, context_dim), dtype=jnp.float32)

ARGS = {
    "hidden_dim": 8,
    "out_dim": 5,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type BilinearFirstMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BilinearFirstMLP 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
BilinearFullMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("0f917b109efda65c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, BilinearFullMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
from typing import Sequence, Dict, Callable, Tuple
import math

default_bias_init = nn.initializers.zeros

default_kernel_init = nn.initializers.lecun_normal()

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class BilinearDense(nn.Module):
    """
    Bilinear implementation as in:
    Multiplicative Interactions and Where to Find Them, ICLR 2020, https://openreview.net/forum?id=rylnK6VtDH
    """
    out_dim: int
    kernel_init: Callable = default_kernel_init
    bias_init: Callable = default_bias_init

    @nn.compact
    def __call__(self, x, z):
        W_layer = nn.Dense(self.out_dim * x.shape[-1], kernel_init=self.kernel_init, bias_init=self.bias_init)
        b_layer = nn.Dense(self.out_dim, kernel_init=self.kernel_init, bias_init=self.bias_init)
        W = W_layer(z).reshape(-1, self.out_dim, x.shape[-1])
        b = b_layer(z)[..., None]
        out = W @ x[..., None] + b
        return out.squeeze(-1)

class BilinearFullMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        bilinear1 = BilinearDense(self.hidden_dim, kernel_init=pytorch_init(f_d), bias_init=pytorch_init(f_d))
        bilinear2 = BilinearDense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        bilinear3 = BilinearDense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        out = nn.relu(bilinear1(feature, context))
        out = nn.relu(bilinear2(out, context))
        out = nn.relu(bilinear3(out, context))
        out = linear4(out)
        return out

# Create sample JAX arrays for feature and context
feature = jnp.ones((2, 4))  # batch_size=2, feature_dim=4
context = jnp.ones((2, 3))  # batch_size=2, context_dim=3

ARGS = {
    "hidden_dim": 5,
    "out_dim": 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type BilinearFullMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BilinearFullMLP 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
BilinearLastMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("a231dfdd79465ac6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, BilinearLastMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Set seed for reproducibility
jax.config.update('jax_default_prng_impl', 'threefry2x32')

ARGS = {
    'hidden_dim': 8,
    'out_dim': 4,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type BilinearLastMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BilinearLastMLP 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
ConcatFirstMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("bd41551ec9638504")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ConcatFirstMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class ConcatFirstMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        network = nn.Sequential([nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d + c_d), bias_init=pytorch_init(f_d + c_d)), nn.relu, nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d)), nn.relu, nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d)), nn.relu, nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))])
        out = network(jnp.hstack([feature, context]))
        return out

ARGS = {
    "hidden_dim": 8,
    "out_dim": 3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ConcatFirstMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ConcatFirstMLP 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
ConcatFullMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("41028f940725e4a0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ConcatFullMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class ConcatFullMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        linear1 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d + c_d), bias_init=pytorch_init(f_d + c_d))
        linear2 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d + c_d), bias_init=pytorch_init(h_d + c_d))
        linear3 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d + c_d), bias_init=pytorch_init(h_d + c_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        out = nn.relu(linear1(jnp.hstack([feature, context])))
        out = nn.relu(linear2(jnp.hstack([out, context])))
        out = nn.relu(linear3(jnp.hstack([out, context])))
        out = linear4(out)
        return out

ARGS = {
    "hidden_dim": 8,
    "out_dim": 3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ConcatFullMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ConcatFullMLP 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
ConcatLastMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("21b612feebc3499d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ConcatLastMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class ConcatLastMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        linear1 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d), bias_init=pytorch_init(f_d))
        linear2 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear3 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d + c_d), bias_init=pytorch_init(h_d + c_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        out = nn.relu(linear1(feature))
        out = nn.relu(linear2(out))
        out = nn.relu(linear3(jnp.hstack([out, context])))
        out = linear4(out)
        return out

# Create model instance
model = ConcatLastMLP(hidden_dim=8, out_dim=3)

# Create sample inputs
key = jax.random.PRNGKey(0)
batch_size = 2
feature_dim = 4
context_dim = 5

feature = jnp.ones((batch_size, feature_dim))
context = jnp.ones((batch_size, context_dim))

# Initialize model parameters
params = model.init(key, feature, context)

# Test forward pass
output = model.apply(params, feature, context)

ARGS = {
    "hidden_dim": 8,
    "out_dim": 3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ConcatLastMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ConcatLastMLP 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
FiLM Ran
ran on a synthesized input · constructed an object; no result was computed
zzmtsvv/sac_rnd/rnd_modules.py
pointer only (licence: NONE) · get_code("ceefac1ae561a6c3")
How it ran

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

FilmFirstMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("be3f4b3f06b0e6bd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, FilmFirstMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class FilmFirstMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        film = nn.Dense(2 * self.hidden_dim, kernel_init=pytorch_init(c_d), bias_init=pytorch_init(c_d))
        linear1 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d), bias_init=pytorch_init(f_d))
        linear2 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear3 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        gamma, beta = jnp.split(film(context), 2, axis=-1)
        out = nn.relu(gamma * linear1(feature) + beta)
        out = nn.relu(linear2(out))
        out = nn.relu(linear3(out))
        out = linear4(out)
        return out

# Create test inputs
key = jax.random.PRNGKey(0)
batch_size = 2
feature_dim = 4
context_dim = 3
hidden_dim = 5
out_dim = 6

feature = jax.random.normal(key, shape=(batch_size, feature_dim))
context = jax.random.normal(key, shape=(batch_size, context_dim))

ARGS = {
    "hidden_dim": hidden_dim,
    "out_dim": out_dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FilmFirstMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FilmFirstMLP 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
FilmFullMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("91e0406ad5665e10")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, FilmFullMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class FilmFullMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        film = nn.Dense(2 * self.hidden_dim * 3, kernel_init=pytorch_init(c_d), bias_init=pytorch_init(c_d))
        linear1 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d), bias_init=pytorch_init(f_d))
        linear2 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear3 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        gamma, beta = jnp.split(film(context).reshape(context.shape[0], 3, -1), 2, axis=-1)
        out = nn.relu(gamma[:, 0] * linear1(feature) + beta[:, 0])
        out = nn.relu(gamma[:, 1] * linear2(out) + beta[:, 1])
        out = nn.relu(gamma[:, 2] * linear3(out) + beta[:, 2])
        out = linear4(out)
        return out

key = jax.random.PRNGKey(0)
batch_size = 2
feature_dim = 4
context_dim = 3
hidden_dim = 5
out_dim = 6

feature = jax.random.normal(key, (batch_size, feature_dim))
context = jax.random.normal(key, (batch_size, context_dim))

model = FilmFullMLP(hidden_dim=hidden_dim, out_dim=out_dim)
params = model.init(key, feature, context)
output = model.apply(params, feature, context)

ARGS = {
    "hidden_dim": hidden_dim,
    "out_dim": out_dim
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FilmFullMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FilmFullMLP 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
FilmLastMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("3ea09b9fb948e867")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, FilmLastMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
import math

def pytorch_init(fan_in: float):
    """
    Default init for PyTorch Linear layer weights and biases:
    https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
    """
    bound = math.sqrt(1 / fan_in)

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

class FilmLastMLP(nn.Module):
    hidden_dim: int
    out_dim: int

    @nn.compact
    def __call__(self, feature, context):
        f_d, c_d, h_d = (feature.shape[-1], context.shape[-1], self.hidden_dim)
        film = nn.Dense(2 * self.hidden_dim, kernel_init=pytorch_init(c_d), bias_init=pytorch_init(c_d))
        linear1 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(f_d), bias_init=pytorch_init(f_d))
        linear2 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear3 = nn.Dense(self.hidden_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        linear4 = nn.Dense(self.out_dim, kernel_init=pytorch_init(h_d), bias_init=pytorch_init(h_d))
        gamma, beta = jnp.split(film(context), 2, axis=-1)
        out = nn.relu(linear1(feature))
        out = nn.relu(linear2(out))
        out = nn.relu(gamma * linear3(out) + beta)
        out = linear4(out)
        return out

ARGS = {
    "hidden_dim": 8,
    "out_dim": 3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FilmLastMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FilmLastMLP 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
GatedMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("12f4f05b82493956")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, GatedMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Initialize random keys for JAX
key = jax.random.PRNGKey(0)

# Create small test inputs
batch_size = 2
feature_dim = 4
context_dim = 3
hidden_dim = 5
out_dim = 6

# Create dummy feature and context arrays
feature = jnp.ones((batch_size, feature_dim))
context = jnp.ones((batch_size, context_dim))

# Define ARGS dictionary with all GatedMLP parameters
ARGS = {
    "hidden_dim": hidden_dim,
    "out_dim": out_dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GatedMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a GatedMLP 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
PredictorNetwork Ran
ran on a synthesized input · constructed an object; no result was computed
zzmtsvv/sac_rnd/rnd_modules.py
pointer only (licence: NONE) · get_code("590f9905be494072")
How it ran

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

PriorNetwork Ran
ran on a synthesized input · constructed an object; no result was computed
zzmtsvv/sac_rnd/rnd_modules.py
pointer only (licence: NONE) · get_code("d33f9e805bac2d2e")
How it ran

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

TorchBilinearDense Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("f08bbf7286cf9812")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TorchBilinearDense: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Set seed for reproducibility
key = jax.random.PRNGKey(0)

# Define tiny input and output dimensions for testing
batch_size = 2
x_dim = 4
z_dim = 3
out_dim = 5

# Create sample input tensors
x = jnp.ones((batch_size, x_dim), dtype=jnp.float32)
z = jnp.ones((batch_size, z_dim), dtype=jnp.float32)

# Define kernel and bias initializers
kernel_init = nn.initializers.lecun_normal()
bias_init = nn.initializers.zeros

# Construct ARGS dictionary with all parameters for TorchBilinearDense
ARGS = {
    'out_dim': out_dim,
    'kernel_init': kernel_init,
    'bias_init': bias_init,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TorchBilinearDense — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TorchBilinearDense 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
TorchBilinearFirstMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("7131c9d471efe1a4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TorchBilinearFirstMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Set random seed for reproducibility
key = jax.random.PRNGKey(0)

# Define small tensors for feature and context inputs
# batch_size=2, feature_dim=4, context_dim=3
feature = jnp.ones((2, 4), dtype=jnp.float32)
context = jnp.ones((2, 3), dtype=jnp.float32)

# Define the arguments for TorchBilinearFirstMLP
ARGS = {
    'hidden_dim': 5,
    'out_dim': 6,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TorchBilinearFirstMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TorchBilinearFirstMLP 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
TorchBilinearFullMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("93814161ebae72e3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TorchBilinearFullMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn

# Set seed for reproducibility
jax.config.update("jax_default_prng_impl", "threefry2x32")

ARGS = {
    "hidden_dim": 8,
    "out_dim": 5,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TorchBilinearFullMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TorchBilinearFullMLP 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
TorchBilinearLastMLP Ran
ran on a synthesized input · constructed an object; no result was computed
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("a9a4771d15f5a324")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TorchBilinearLastMLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import flax.linen as nn
from jax import random

# Set up random key for JAX
key = random.PRNGKey(0)

# Create tiny input tensors
batch_size = 2
feature_dim = 4
context_dim = 3
hidden_dim = 5
out_dim = 6

# feature shape: (batch_size, feature_dim)
feature = jnp.ones((batch_size, feature_dim), dtype=jnp.float32)

# context shape: (batch_size, context_dim)
context = jnp.ones((batch_size, context_dim), dtype=jnp.float32)

ARGS = {
    "hidden_dim": hidden_dim,
    "out_dim": out_dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TorchBilinearLastMLP — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TorchBilinearLastMLP 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
normalize Ran
ran, honoured the contract we drafted
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("adb37ed22c2703a9")
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 small JAX arrays for normalization
arr = jnp.array([[1.0, 2.0, 3.0, 4.0],
                  [5.0, 6.0, 7.0, 8.0]], dtype=jnp.float32)

mean = jnp.array([3.0, 4.0, 5.0, 6.0], dtype=jnp.float32)

std = jnp.array([1.0, 1.5, 2.0, 2.5], dtype=jnp.float32)

eps = 1e-8

ARGS = {
    "arr": arr,
    "mean": mean,
    "std": std,
    "eps": eps,
}
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
pytorch_init Ran
ran; our contract draft was wrong, not the code
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("b62cf8b42971d060")
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 math

# The function pytorch_init takes a single parameter: fan_in (a float)
# It returns an initialization function that uses jax.random.uniform
# The fan_in parameter is used to calculate a bound for uniform initialization

fan_in_value = 10.0

ARGS = {
    "fan_in": fan_in_value,
}
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
DRND Not yet run
unverified (harvested, no recorded run)
zzmtsvv/sac_drnd/drnd_modules.py
pointer only (licence: NONE) · get_code("c901b83a37fd3070")
How it ran

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

EnsembledFiLM Not yet run
unverified (harvested, no recorded run)
zzmtsvv/sac_drnd/drnd_modules.py
pointer only (licence: NONE) · get_code("7120bed03164d5ad")
How it ran

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

EnsembledTargetNetwork Not yet run
unverified (harvested, no recorded run)
zzmtsvv/sac_drnd/drnd_modules.py
pointer only (licence: NONE) · get_code("bc04133ae453cf88")
How it ran

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

RND Not yet run
unverified (harvested, no recorded run)
tinkoff-ai/sac-rnd/offline_sac/networks.py
code served (permissive licence) · get_code("ffdc71693d91c600")
How it ran

no invocation recorded.

RND Not yet run
unverified (harvested, no recorded run)
zzmtsvv/sac_rnd/rnd_modules.py
pointer only (licence: NONE) · get_code("28ed96e0bf8e9eab")
How it ran

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

train Not yet run
unverified (harvested, no recorded run)
shidilrzf/anti-exploration-rl/examples/train_ae.py
code served (permissive licence) · get_code("226a9594f55e107f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 26 samples lifted from this paper's linked repositories, 26 carry a function identity and 0 do not. Of the 26 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 26 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Despite the success of Random Network Distillation (RND) in various domains, it was shown as not discriminative enough to be used as an uncertainty estimator for penalizing out-of-distribution actions in offline reinforcement learning. In this paper, we revisit these results and show that, with a naive choice of conditioning for the RND prior, it becomes infeasible for the actor to effectively minimize the anti-exploration bonus and discriminativity is not an issue. We show that this limitation can be avoided with conditioning based on Feature-wise Linear Modulation (FiLM), resulting in a simple and efficient ensemble-free algorithm based on Soft Actor-Critic. We evaluate it on the D4RL benchmark, showing that it is capable of achieving performance comparable to ensemble-based methods and outperforming ensemble-free approaches by a wide margin.

For agents

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

get_harvested_code_for_paper("2301.13616")
get_code_for_paper("2301.13616")
have("2301.13616")

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

curl https://syntology.ai/api/ran/2301.13616.json

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

[![Syntology run record](https://syntology.ai/api/ran/2301.13616.svg)](https://syntology.ai/paper/2301.13616)

Connect an agent — have() is free.