SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.09329 · NeurIPS · 2024

Is Value Learning Really the Main Bottleneck in Offline RL?

Sergey Levine, Aviral Kumar, Seohong Park, Kevin Frans

arXiv · PDF · Open in the Atlas

Code that ran

7 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (6 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 7 have no contract verdict; 0 failures of our instrument; 3 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/2406.09329.json · badge: /api/ran/2406.09329.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
seohongpark/ogbench community
Papers with Code archive
7 ran · 3 no recorded run
FunctionStatusWhere it lives
GCActor Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("9bc539476ffcf222")
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, GCActor: 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 distrax

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

# Define a minimal GCEncoder if needed
class MinimalGCEncoder(nn.Module):
    """Minimal goal-conditioned encoder."""
    
    @nn.compact
    def __call__(self, observations, goals, goal_encoded=False):
        # Simple concatenation of obs and goals
        if goals is not None:
            return jnp.concatenate([observations, goals], axis=-1)
        return observations

# Define ARGS dict with all parameters for GCActor
ARGS = {
    "hidden_dims": (8, 6),  # Two hidden layers with distinct sizes
    "action_dim": 3,  # Action dimension
    "log_std_min": -5,  # Minimum log std
    "log_std_max": 2,  # Maximum log std
    "tanh_squash": True,  # Use tanh squashing
    "state_dependent_std": False,  # Use constant std
    "const_std": True,  # Constant std
    "final_fc_init_scale": 0.01,  # Final layer init scale
    "gc_encoder": MinimalGCEncoder(),  # Optional encoder
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
distrax, flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GCActor — 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 GCActor 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
GCEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("85c24c360661275c")
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, GCEncoder: 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 flax.linen as nn
import jax.numpy as jnp
import jax

# Define small encoder modules
class SimpleEncoder(nn.Module):
    output_dim: int
    
    @nn.compact
    def __call__(self, x):
        # Simple linear encoder
        x = nn.Dense(self.output_dim)(x)
        return x

# Create encoder instances
state_encoder = SimpleEncoder(output_dim=5)
goal_encoder = SimpleEncoder(output_dim=6)
concat_encoder = SimpleEncoder(output_dim=7)

# Create sample data
batch_size = 2
obs_dim = 4
goal_dim = 3

# JAX arrays for observations and goals
observations = jnp.ones((batch_size, obs_dim))
goals = jnp.ones((batch_size, goal_dim))
goal_encoded_rep = jnp.ones((batch_size, 6))

ARGS = {
    "state_encoder": state_encoder,
    "goal_encoder": goal_encoder,
    "concat_encoder": concat_encoder,
}
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 GCEncoder — 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 GCEncoder 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
GCValue Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("1a0dd970d5292061")
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, GCValue: 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

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

# Create a simple GCEncoder stand-in class
class SimpleGCEncoder(nn.Module):
    """Simple goal-conditioned encoder."""
    
    @nn.compact
    def __call__(self, observations, goals):
        # Concatenate observations and goals, return encoded representation
        concatenated = jnp.concatenate([observations, goals], axis=-1)
        encoded = nn.Dense(4)(concatenated)
        return encoded

# Test data shapes: batch=2, obs_dim=4, goal_dim=3, action_dim=5
batch_size = 2
obs_dim = 4
goal_dim = 3
action_dim = 5

# Create sample observations, goals, and actions
observations = jnp.ones((batch_size, obs_dim))
goals = jnp.ones((batch_size, goal_dim))
actions = jnp.ones((batch_size, action_dim))

# Create the GCEncoder instance
gc_encoder_module = SimpleGCEncoder()

ARGS = {
    'hidden_dims': (8, 6, 1),
    'layer_norm': True,
    'ensemble': True,
    'gc_encoder': gc_encoder_module,
}
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 GCValue — 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 GCValue 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
Identity Ran
ran on a synthesized input
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("55eac5ab51a6832f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) 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 torch
import flax.linen as nn
import jax
import jax.numpy as jnp

class Identity(nn.Module):
    """Identity layer."""

    def __call__(self, x):
        return x

INIT_ARGS = {}

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

# Construct and call
model = Identity()
key = jax.random.PRNGKey(0)
params = model.init(key, ARGS['x'])
output = model.apply(params, ARGS['x'])
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
flax, jax, torch — 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ImpalaEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("503831beac2bf3d9")
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, ImpalaEncoder: 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
from flax import linen as nn
from typing import Sequence

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

# Create a small test input tensor
# Shape: (batch=2, height=4, width=3, channels=1)
test_input = jnp.ones((2, 4, 3, 1), dtype=jnp.uint8) * 128

# Define ARGS mapping all parameters of ImpalaEncoder
ARGS = {
    'width': 1,
    'stack_sizes': (16, 32, 32),
    'num_blocks': 2,
    'dropout_rate': 0.1,
    'mlp_hidden_dims': (512,),
    'layer_norm': False,
}
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 ImpalaEncoder — 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 ImpalaEncoder 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
MLP Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("ccbb1f2bba20397b")
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, MLP: 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, Any

def default_init(scale=1.0):
    """Default kernel initializer."""
    return nn.initializers.variance_scaling(scale, 'fan_avg', 'uniform')

class MLP(nn.Module):
    """Multi-layer perceptron.

    Attributes:
        hidden_dims: Hidden layer dimensions.
        activations: Activation function.
        activate_final: Whether to apply activation to the final layer.
        kernel_init: Kernel initializer.
        layer_norm: Whether to apply layer normalization.
    """
    hidden_dims: Sequence[int]
    activations: Any = nn.gelu
    activate_final: bool = False
    kernel_init: Any = default_init()
    layer_norm: bool = False

    @nn.compact
    def __call__(self, x):
        for i, size in enumerate(self.hidden_dims):
            x = nn.Dense(size, kernel_init=self.kernel_init)(x)
            if i + 1 < len(self.hidden_dims) or self.activate_final:
                x = self.activations(x)
                if self.layer_norm:
                    x = nn.LayerNorm()(x)
        return x

ARGS = {
    'hidden_dims': (8, 6, 4),
    'activations': nn.gelu,
    'activate_final': False,
    'kernel_init': default_init(1.0),
    'layer_norm': False,
}
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 MLP — 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 MLP 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
ModuleDict Ran
ran on a synthesized input · constructed an object; no result was computed
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("1b4a14ceaf967da5")
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, ModuleDict: 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 flax.linen as nn
from typing import Sequence, Dict, Mapping
import jax
import jax.numpy as jnp

# Define some simple test modules
class SimpleModule1(nn.Module):
    @nn.compact
    def __call__(self, x):
        return nn.Dense(3)(x)

class SimpleModule2(nn.Module):
    @nn.compact
    def __call__(self, x):
        return nn.Dense(5)(x)

class SimpleModule3(nn.Module):
    @nn.compact
    def __call__(self, x, y):
        return nn.Dense(2)(x) + nn.Dense(2)(y)

# Create the modules dictionary
modules_dict = {
    "module_a": SimpleModule1(),
    "module_b": SimpleModule2(),
    "module_c": SimpleModule3(),
}

ARGS = {
    "modules": modules_dict,
}
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 ModuleDict — 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 ModuleDict 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
GCDiscreteActor Not yet run
unverified (harvested, no recorded run)
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("76e656aaad5ce115")
How it ran

no invocation recorded.

HIQLAgent Not yet run
unverified (harvested, no recorded run)
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("3c394762e2a85d6f")
How it ran

no invocation recorded.

LengthNormalize Not yet run
unverified (harvested, no recorded run)
seohongpark/ogbench/impls/agents/hiql.py
code served (permissive licence) · get_code("2c53e405a504fd6d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 7 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 7 later) and 3 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
ModuleDict
brick of this paper's assembly: helper
10 other papers
10 later · 6 only as an assembly brick
2502.02538 (2025-02)
Flow Q-Learning
2506.07744 (2025-06)
Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning
2506.08902 (2025-06)
Intention-Conditioned Flow Occupancy Models
and 7 more not named here
ImpalaEncoder
brick of this paper's assembly: helper
9 other papers
9 later · 5 only as an assembly brick
2502.02538 (2025-02)
Flow Q-Learning
2506.07744 (2025-06)
Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning
2506.21427 (2025-06)
Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning
and 6 more not named here
GCEncoder
brick of this paper's assembly: helper
2 other papers
2 later
2506.07744 (2025-06)
Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning
2605.20878 (2026-05)
CIG: Exploration via Conditional Information Gain
GCActor
brick of this paper's assembly: helper
1 other paper
1 later
2605.20878 (2026-05)
CIG: Exploration via Conditional Information Gain
GCDiscreteActor
brick of this paper's assembly: helper
1 other paper
1 later
2605.20878 (2026-05)
CIG: Exploration via Conditional Information Gain
GCValue
brick of this paper's assembly: helper
1 other paper
1 later
2605.20878 (2026-05)
CIG: Exploration via Conditional Information Gain
MLP
brick of this paper's assembly: helper
1 other paper
1 later
2605.20878 (2026-05)
CIG: Exploration via Conditional Information Gain

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

While imitation learning requires access to high-quality data, offline reinforcement learning (RL) should, in principle, perform similarly or better with substantially lower data quality by using a value function. However, current results indicate that offline RL often performs worse than imitation learning, and it is often unclear what holds back the performance of offline RL. Motivated by this observation, we aim to understand the bottlenecks in current offline RL algorithms. While poor performance of offline RL is typically attributed to an imperfect value function, we ask: is the main bottleneck of offline RL indeed in learning the value function, or something else? To answer this question, we perform a systematic empirical study of (1) value learning, (2) policy extraction, and (3) policy generalization in offline RL problems, analyzing how these components affect performance. We make two surprising observations. First, we find that the choice of a policy extraction algorithm significantly affects the performance and scalability of offline RL, often more so than the value learning objective. For instance, we show that common value-weighted behavioral cloning objectives (e.g., AWR) do not fully leverage the learned value function, and switching to behavior-constrained policy gradient objectives (e.g., DDPG+BC) often leads to substantial improvements in performance and scalability. Second, we find that a big barrier to improving offline RL performance is often imperfect policy generalization on test-time states out of the support of the training data, rather than policy learning on in-distribution states. We then show that the use of suboptimal but high-coverage data or testtime policy training techniques can address this generalization issue in practice. Specifically, we propose two simple test-time policy improvement methods and show that these methods lead to better performance. Project page: https:// seohong.me/projects/offrl-bottlenecks 38th Conference on Neural Information Processing Systems (NeurIPS 2024).

For agents

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

get_harvested_code_for_paper("2406.09329")
get_code_for_paper("2406.09329")
have("2406.09329")

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

curl https://syntology.ai/api/ran/2406.09329.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.09329.svg)](https://syntology.ai/paper/2406.09329)

Connect an agent — have() is free.