SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.05066 · 2026

Reward-Conditioned Reinforcement Learning

Pieter Abbeel, Michal Nauman, Marek Cygan

arXiv · PDF · Open in the Atlas

Code that ran

15 of 29 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (10 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 4 failures of our instrument; 14 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/2603.05066.json · badge: /api/ran/2603.05066.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
sukhijab/maxinforl_jax canonical (classifier label)
classifier basis: census
1 ran · 0 no recorded run
naumix/BiggerRegularizedCategorical community
Papers with Code archive
11 ran · 11 no recorded run
DAVIAN-Robotics/SimbaV2 community
Papers with Code archive
3 ran · 3 no recorded run
FunctionStatusWhere it lives
BroNet Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("397e1bc43033edf4")
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, BroNet: 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 ARGS dictionary with all parameters for BroNet
ARGS = {
    "hidden_dims": 8,
    "depth": 2,
    "add_final_layer": True,
    "output_nodes": 16,
    "activations": nn.relu,
}
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 BroNet — 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 BroNet 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
BronetBlock Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("03073dfb6f111e5e")
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, BronetBlock: 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 activation function
def relu_activation(x):
    return nn.relu(x)

# Define ARGS dict with all required parameters for BronetBlock
ARGS = {
    "hidden_dims": 4,
    "activations": relu_activation,
}
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 BronetBlock — 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 BronetBlock 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
Critic Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("5366a0d342198b2d")
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, Critic: 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 ARGS dictionary with all parameters for Critic
ARGS = {
    'num_tasks': 5,
    'embedding_size': 8,
    'ensemble_size': 2,
    'hidden_dims': 64,
    'depth': 2,
    'activations': nn.relu,
    'output_nodes': 101,
    'multitask': True,
}
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 Critic — 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 Critic 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
Model Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("6e0a024aaf80cb02")
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, Model: 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 optax
import flax
import flax.linen as nn
from typing import Any, Optional, Sequence
from flax.core import FrozenDict

# Define a minimal Flax model for testing
class SimpleModel(nn.Module):
    features: int = 4
    
    @nn.compact
    def __call__(self, x):
        x = nn.Dense(self.features)(x)
        return x

# Create a minimal model instance
model_def = SimpleModel(features=4)

# Create input arrays for initialization
key = jax.random.PRNGKey(0)
batch_size = 2
input_dim = 3
inputs = [key, jnp.ones((batch_size, input_dim))]

# Initialize the model to get params
variables = model_def.init(*inputs)
params = variables.pop('params')

# Create an optimizer
tx = optax.adam(learning_rate=0.001)
opt_state = tx.init(params)

# Define the ARGS dictionary
ARGS = {
    'step': 1,
    'apply_fn': model_def,
    'params': params,
    'tx': tx,
    'opt_state': opt_state
}
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 Model — 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 Model 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
Network Ran
ran on a synthesized input · constructed an object; no result was computed
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("ebeac4bf6f7894e1")
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, Network: 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 optax
from flax.core import FrozenDict

class SimpleNetwork(nn.Module):
    features: int = 8
    
    @nn.compact
    def __call__(self, x):
        x = nn.Dense(self.features)(x)
        x = nn.relu(x)
        x = nn.Dense(4)(x)
        return x

network_def = SimpleNetwork(features=8)

network_inputs = {
    'x': jnp.ones((2, 4))
}

variables = network_def.init(jax.random.PRNGKey(0), **network_inputs)
params = variables['params']

tx = optax.adam(learning_rate=0.001)
opt_state = tx.init(params)

ARGS = {
    'network_def': network_def,
    'params': params,
    'tx': tx,
    'opt_state': opt_state,
    'update_step': 0,
}
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 Network — 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 Network 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
QValue Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("8f9ccab1e13cfe0d")
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, QValue: 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 the ARGS dictionary with all parameters for QValue
ARGS = {
    "hidden_dims": 64,
    "depth": 2,
    "activations": nn.relu,
    "output_nodes": 101,
}
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 QValue — 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 QValue 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
QValueEnsemble Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("0aefeaa0335502ec")
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, QValueEnsemble: 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 Callable

# Define the required classes and functions that are referenced in QValueEnsemble

def default_init(scale: float = jnp.sqrt(2)):
    return nn.initializers.orthogonal(scale)

class BronetBlock(nn.Module):
    hidden_dims: int
    activations: Callable[[jnp.ndarray], jnp.ndarray]

    @nn.compact
    def __call__(self, x: jnp.ndarray):
        res = nn.Dense(self.hidden_dims, kernel_init=default_init())(x)
        res = nn.LayerNorm()(res)
        res = self.activations(res)
        res = nn.Dense(self.hidden_dims, kernel_init=default_init())(res)
        res = nn.LayerNorm()(res)
        return res + x

class BroNet(nn.Module):
    hidden_dims: int
    depth: int
    add_final_layer: bool = False
    output_nodes: int = 101
    activations: Callable[[jnp.ndarray], jnp.ndarray] = nn.relu

    @nn.compact
    def __call__(self, x: jnp.ndarray):
        x = nn.Dense(self.hidden_dims, kernel_init=default_init())(x)
        x = nn.LayerNorm()(x)
        x = self.activations(x)
        for i in range(self.depth):
            x = BronetBlock(self.hidden_dims, self.activations)(x)
        if self.add_final_layer:
            x = nn.Dense(self.output_nodes, kernel_init=default_init())(x)
        return x

class QValue(nn.Module):
    hidden_dims: int = 512
    depth: int = 2
    activations: Callable[[jnp.ndarray], jnp.ndarray] = nn.relu
    output_nodes: int = 101

    def setup(self):
        self.critic = BroNet(hidden_dims=self.hidden_dims, depth=self.depth, activations=self.activations, add_final_layer=True, output_nodes=self.output_nodes)

    def __call__(self, inputs: jnp.ndarray):
        q_value = self.critic(inputs)
        return q_value

class QValueEnsemble(nn.Module):
    ensemble_size: int = 2
    hidden_dims: int = 512
    depth: int = 2
    activations: Callable[[jnp.ndarray], jnp.ndarray] = nn.relu
    output_nodes: int = 101

    def setup(self):
        VmapCritic = nn.vmap(QValue, variable_axes={'params': 0}, split_rngs={'params': True}, in_axes=None, out_axes=0, axis_size=self.ensemble_size)
        self.q_value_ensemble = VmapCritic(hidden_dims=self.hidden_dims, depth=self.depth, activations=self.activations, output_nodes=self.output_nodes)

    def __call__(self, inputs: jnp.ndarray):
        q_values = self.q_value_ensemble(inputs)
        return q_values

# Build ARGS dictionary with valid arguments for QValueEnsemble
ARGS = {
    'ensemble_size': 2,
    'hidden_dims': 64,
    'depth': 2,
    'activations': nn.relu,
    'output_nodes': 101
}
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 QValueEnsemble — 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 QValueEnsemble 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
SaveState Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("c1a7eb383274fee4")
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, SaveState: 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 torch
import numpy as np
import flax
import optax
from typing import Any, Optional

torch.manual_seed(0)

# Define Params type as a FrozenDict
Params = flax.core.FrozenDict[str, Any]

# Create minimal params - a simple frozen dict with some parameter tensors
params_dict = {
    "layer1": np.array([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]], dtype=np.float32),
    "layer2": np.array([[0.1, 0.2], [0.3, 0.4], [0.5, 0.6], [0.7, 0.8]], dtype=np.float32),
}
params = flax.core.FrozenDict(params_dict)

# Create a minimal opt_state using optax
# Adam optimizer creates a state with count, mu, and nu for each parameter
optimizer = optax.adam(learning_rate=0.001)
opt_state = optimizer.init(params)

# Construct the ARGS dictionary
ARGS = {
    "params": params,
    "opt_state": opt_state,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, numpy, optax, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type SaveState — 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 SaveState 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
TaskEmbedding Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("aed2ac4c0802bf40")
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, TaskEmbedding: 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 the arguments for TaskEmbedding
ARGS = {
    "num_tasks": 8,
    "embedding_size": 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 TaskEmbedding — 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 TaskEmbedding 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
Temperature Ran
ran on a synthesized input · constructed an object; no result was computed
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("610fb1dffadf66a3")
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, Temperature: 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

# The Temperature class is a Flax nn.Module with one parameter:
# - initial_temperature: float (default 1.0)

# The __call__ method doesn't take any arguments and returns a jnp.ndarray
# It uses initial_temperature to initialize a log_temp parameter

ARGS = {
    'initial_temperature': 2.5
}
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 Temperature — 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 Temperature 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
_sample_actions Ran
ran; our contract draft was wrong, not the code
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("fda9e4a1fa4ace30")
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
import numpy as np
import optax
import flax
import flax.linen as nn
from typing import Any, Optional, Sequence

PRNGKey = Any

Params = flax.core.FrozenDict[str, Any]

@flax.struct.dataclass
class SaveState:
    params: Params
    opt_state: Optional[optax.OptState] = None

def tree_norm(tree):
    return jnp.sqrt(sum(((x ** 2).sum() for x in jax.tree_util.tree_leaves(tree))))

@flax.struct.dataclass
class Model:
    step: int
    apply_fn: flax.linen.Module = flax.struct.field(pytree_node=False)
    params: Params
    tx: Optional[optax.GradientTransformation] = flax.struct.field(pytree_node=False)
    opt_state: Optional[optax.OptState] = None

    @classmethod
    def create(cls, model_def: flax.linen.Module, inputs: Sequence[jnp.ndarray], tx: Optional[optax.GradientTransformation]=None):
        variables = model_def.init(*inputs)
        params = variables.pop('params')
        if tx is not None:
            opt_state = tx.init(params)
        else:
            opt_state = None
        return cls(step=1, apply_fn=model_def, params=params, tx=tx, opt_state=opt_state)

    def __call__(self, *args, **kwargs):
        return self.apply_fn.apply({'params': self.params}, *args, **kwargs)

    def apply(self, *args, **kwargs):
        return self.apply_fn.apply(*args, **kwargs)

    def apply_gradient(self, loss_fn):
        grad_fn = jax.grad(loss_fn, has_aux=True)
        grads, info = grad_fn(self.params)
        grad_norm = tree_norm(grads)
        info['grad_norm'] = grad_norm
        updates, new_opt_state = self.tx.update(grads, self.opt_state, self.params)
        new_params = optax.apply_updates(self.params, updates)
        return (self.replace(step=self.step + 1, params=new_params, opt_state=new_opt_state), info)

    def get_gradient(self, loss_fn):
        grad_fn = jax.grad(loss_fn, has_aux=True)
        grads, info = grad_fn(self.params)
        return grads

    def save(self, save_path: str):
        import os
        os.makedirs(os.path.dirname(save_path), exist_ok=True)
        with open(save_path, 'wb') as f:
            f.write(flax.serialization.to_bytes(SaveState(params=self.params, opt_state=self.opt_state)))

    def load(self, load_path: str):
        with open(load_path, 'rb') as f:
            contents = f.read()
            saved_state = flax.serialization.from_bytes(SaveState(params=self.params, opt_state=self.opt_state), contents)
        return self.replace(params=saved_state.params, opt_state=saved_state.opt_state)

@jax.jit
def _sample_actions(rng: PRNGKey, actor: Model, inputs: np.ndarray, temperature: float=1.0):
    dist = actor(inputs, temperature)
    rng, key = jax.random.split(rng)
    actions = dist.sample(seed=key)
    return (rng, actions)

class SimpleDistribution:
    """Mock distribution with sample method"""
    def __init__(self, shape=(2, 4)):
        self.shape = shape
    
    def sample(self, seed):
        return jax.random.normal(seed, self.shape)

class SimpleActor(nn.Module):
    """Minimal actor module that returns a distribution"""
    
    @nn.compact
    def __call__(self, inputs, temperature):
        x = nn.Dense(8)(inputs)
        x = nn.relu(x)
        x = nn.Dense(4)(x)
        return SimpleDistribution(shape=x.shape)

# Create the actor model
actor_def = SimpleActor()
inputs_for_init = [jax.random.PRNGKey(0), jnp.ones((2, 4)), 1.0]
actor_model = Model.create(actor_def, inputs_for_init, tx=optax.adam(1e-3))

# Create inputs for the function
rng_key = jax.random.PRNGKey(42)
input_array = np.ones((2, 4), dtype=np.float32)

ARGS = {
    'rng': rng_key,
    'actor': actor_model,
    'inputs': input_array,
    'temperature': 1.0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, numpy, optax — 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
categorical_td_loss Ran
ran; our contract draft was wrong, not the code
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("c5c64da0a5960795")
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

# Set random seed for reproducibility
jax.config.update('jax_platform_name', 'cpu')

batch_size = 2
num_bins = 4

# Create small valid arguments
pred_log_probs = jnp.log(jnp.array([[0.25, 0.25, 0.25, 0.25],
                                      [0.2, 0.3, 0.3, 0.2]]))
target_log_probs = jnp.log(jnp.array([[0.3, 0.2, 0.3, 0.2],
                                       [0.25, 0.25, 0.25, 0.25]]))
reward = jnp.array([1.0, -0.5])
done = jnp.array([0.0, 1.0])
actor_entropy = jnp.array([0.5, 0.3])
gamma = 0.99
min_v = -10.0
max_v = 10.0

ARGS = {
    "pred_log_probs": pred_log_probs,
    "target_log_probs": target_log_probs,
    "reward": reward,
    "done": done,
    "actor_entropy": actor_entropy,
    "gamma": gamma,
    "num_bins": num_bins,
    "min_v": min_v,
    "max_v": max_v,
}
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
clip Ran
ran, honoured the contract we drafted
sukhijab/maxinforl_jax/maxinforl_jax/agents/drqv2/actor.py
pointer only (licence: NOASSERTION) · get_code("21c74a14da22835c")
How it ran

The body of this sample is withheld: GitHub could not classify the licence of this paper's repository, sukhijab/maxinforl_jax (NOASSERTION). 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
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("7e1b1dbdf882411b")
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
import flax.linen as nn

ARGS = {
    "scale": jnp.sqrt(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 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
l2normalize_layer Ran
ran; our contract draft was wrong, not the code
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("23df276ac0da323e")
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 small tree structure with 'kernel' that will be processed by l2normalize_layer
# The function checks len(tree['kernel'].shape) to determine axis

# Case 1: 2D kernel (axis=0)
tree_2d = {
    'kernel': jnp.array([[1.0, 2.0, 3.0, 4.0],
                         [5.0, 6.0, 7.0, 8.0]]),
    'bias': jnp.array([0.1, 0.2, 0.3, 0.4])
}

ARGS = {
    'tree': tree_2d
}
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
BRC Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("78dfe812847082e4")
How it ran

no invocation recorded.

NormalTanhPolicy Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("26d8c136c86d1253")
How it ran

no invocation recorded.

_do_multiple_updates Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("96a825f1c022acb2")
How it ran

no invocation recorded.

_get_infos Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("935d551cc3f84f79")
How it ran

no invocation recorded.

_get_temperature Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("25af93d5355ed0b1")
How it ran

no invocation recorded.

_update Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("1c86b0474645b17c")
How it ran

no invocation recorded.

build_actor_input Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("4098c0fa523794b2")
How it ran

no invocation recorded.

l2normalize_network Not yet run
unverified (harvested, no recorded run)
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("c0a8dc3a4cf1fc77")
How it ran

no invocation recorded.

tree_map_until_match Not yet run
unverified (harvested, no recorded run)
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("a87910580e169679")
How it ran

no invocation recorded.

update_actor Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("c6bff04ee9e498a2")
How it ran

no invocation recorded.

update_critic Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("19351a39d87660f8")
How it ran

no invocation recorded.

update_critic Not yet run
unverified (harvested, no recorded run)
DAVIAN-Robotics/SimbaV2/scale_rl/agents/simbaV2/simbaV2_update.py
code served (permissive licence) · get_code("715e080110f0be4c")
How it ran

no invocation recorded.

update_target_critic Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("739f5696cfb832fc")
How it ran

no invocation recorded.

update_temperature Not yet run
unverified (harvested, no recorded run)
naumix/BiggerRegularizedCategorical/jaxrl/agent/brc_learner.py
code served (permissive licence) · get_code("359b630f0fc733a0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 29 samples lifted from this paper's linked repositories, 29 carry a function identity and 0 do not. Of the 29 distinct functions with an identity, 9 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 8 later) and 20 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
SaveState
brick of this paper's assembly: helper
4 other papers
3 dated earlier, 1 later · 2 only as an assembly brick
2405.16158 (2024-05)
Bigger, Regularized, Optimistic: scaling for compute and sample-efficient continuous control
2502.15280 (2025-02)
Hyperspherical Normalization for Scalable Deep Reinforcement Learning
2506.05968 (2025-06)
Gradual Transition from Bellman Optimality Operator to Bellman Operator in Online Reinforcement Learning
and 1 more not named here
BroNet
brick of this paper's assembly: helper
1 other paper
1 later
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
BronetBlock
brick of this paper's assembly: helper
1 other paper
1 later
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
Critic
brick of this paper's assembly: model
1 other paper
1 later
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
QValue
brick of this paper's assembly: helper
1 other paper
1 later
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
QValueEnsemble
brick of this paper's assembly: model
1 other paper
1 later
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
TaskEmbedding
brick of this paper's assembly: model
1 other paper
1 later · 1 only as an assembly brick
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control
Temperature
brick of this paper's assembly: model
1 other paper
1 dated earlier
2310.17966 (2023-10)
Train Once, Get a Family: State-Adaptive Balances for Offline-to-Online Reinforcement Learning
default_init
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2605.16692 (2026-05)
EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control

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

Single-task RL agents are typically trained under a fixed reward function, which limits their robustness to reward misspecification and their ability to adapt to changing preferences. We introduce Reward-Conditioned Reinforcement Learning (RCRL), an off-policy method that conditions agents on reward parameterizations while collecting experience under a single nominal objective. By recomputing counterfactual rewards from shared replay data, RCRL exposes the agent to multiple reward objectives without additional environment interaction, connecting single-task RL with ideas from multi-objective and multi-task learning. Across single-task, multi-task, and vision-based benchmarks, RCRL improves sample efficiency under the nominal reward parameterization, enables efficient adaptation to new parameterizations, and supports zero-shot behavioral adjustment at deployment. Our results show that RCRL provides a scalable mechanism for learning robust, steerable policies without sacrificing the simplicity of single-task training.

For agents

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

get_harvested_code_for_paper("2603.05066")
get_code_for_paper("2603.05066")
have("2603.05066")

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

curl https://syntology.ai/api/ran/2603.05066.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.05066.svg)](https://syntology.ai/paper/2603.05066)

Connect an agent — have() is free.