SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2211.09760 · 2022

VeLO: Training Versatile Learned Optimizers by Scaling Up

arXiv · PDF · Open in the Atlas

Code that ran

15 of 20 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 1 violated it and 12 have no contract verdict; 1 failure of our instrument; 5 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/2211.09760.json · badge: /api/ran/2211.09760.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
google/learned_optimization canonical (classifier label)
classifier basis not recorded on the link
8 ran · 3 no recorded run
amoudgl/celo community (classifier label)
classifier basis not recorded on the link
7 ran · 2 no recorded run
FunctionStatusWhere it lives
colors_for_num Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/notebook_utils.py
code served (permissive licence) · get_code("dc03aa48546537e4")
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 list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from matplotlib import pylab as plt
import seaborn as sns

ARGS = {
    'num': 5,
    'rep': 2,
    'stride': 3,
    'categorical': True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy, seaborn, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
decay_to_param Ran
ran, violated the contract we drafted
amoudgl/celo/celo/optimizers/adafac_mlp_lopt.py
code served (permissive licence) · get_code("519c4737cd569640")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp
import numpy as np

def decay_to_param(x):
    return jnp.log(1 - x) / 10.0

# Create a small valid input for x
# The function computes log(1 - x), so x should be in range (0, 1)
x_value = jnp.array([0.1, 0.2, 0.3, 0.4])

ARGS = {
    "x": x_value
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ema Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/notebook_utils.py
code served (permissive licence) · get_code("e6c862d84c69f94c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (5,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import chex

# Create test data for the ema function
data = np.array([1.0, 2.0, 3.0, 4.0, 5.0])

# Define alpha parameter (decay rate, typically between 0 and 1)
alpha = 0.3

# Define ignore_nan parameter
ignore_nan = False

ARGS = {
    "data": data,
    "alpha": alpha,
    "ignore_nan": ignore_nan,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
chex, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (5,) — shape (5,) 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
factored_dims Ran
ran on a synthesized input
amoudgl/celo/celo/optimizers/celo.py
code served (permissive licence) · get_code("048bb21eeedc76ed")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as onp

# Test case 1: 4D shape - should return tuple of two largest axes
shape_4d = (2, 5, 3, 7)

# Test case 2: 3D shape - should return tuple of two largest axes
shape_3d = (4, 6, 2)

# Test case 3: 2D shape - should return tuple of two largest axes
shape_2d = (3, 8)

# Test case 4: 1D shape - should return None (less than 2 dimensions)
shape_1d = (5,)

# Test case 5: Empty-like shape with 0 elements - still 2D, should work
shape_with_zero = (0, 5)

ARGS = {
    "shape": shape_4d
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
factored_dims Ran
ran on a synthesized input
amoudgl/celo/celo/optimizers/celo2_optax.py
code served (permissive licence) · get_code("f1a9b2bb9feb2fa3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as onp
import torch

# The function takes a 'shape' parameter which should be a tuple or array-like
# of integers representing dimensions. The function:
# 1. Checks if len(shape) < 2, returns None if so
# 2. Uses onp.argsort on the shape to get sorted dimension indices
# 3. Returns the last two indices from the sorted array

# Test case 1: A shape with at least 2 dimensions
shape_example = (5, 3, 7, 2)

ARGS = {
    "shape": shape_example
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
file_open Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/filesystem.py
code served (permissive licence) · get_code("95b8c213917f58e2")
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 TextIOWrapper. 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 tensorflow as tf
import tempfile
import os

# Create a temporary file for testing
temp_dir = tempfile.mkdtemp()
temp_file_path = os.path.join(temp_dir, "test_file.txt")

# Write some test content to the temporary file
with open(temp_file_path, 'w') as f:
    f.write("test content")

# Define ARGS with valid parameters for file_open
ARGS = {
    'path': temp_file_path,
    'mode': 'r'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TextIOWrapper — a TextIOWrapper 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
load_state Ran
ran on a synthesized input
amoudgl/celo/celo/utils.py
code served (permissive licence) · get_code("d69eb755b5da1e5c")
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 tempfile
import os
import jax
from flax import serialization

# Create a temporary directory and file for testing
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "state.pkl")

# Create a sample pytree state structure
sample_state = {
    'layer1': jax.numpy.array([1.0, 2.0, 3.0, 4.0]),
    'layer2': {
        'weights': jax.numpy.array([[1.0, 2.0], [3.0, 4.0]]),
        'bias': jax.numpy.array([0.5, 0.5])
    }
}

# Serialize and save the state to the temporary file
state_bytes = serialization.to_bytes(sample_state)
with open(temp_file, 'wb') as f:
    f.write(state_bytes)

# Create a template state with the same structure but different values
template_state = {
    'layer1': jax.numpy.array([0.0, 0.0, 0.0, 0.0]),
    'layer2': {
        'weights': jax.numpy.array([[0.0, 0.0], [0.0, 0.0]]),
        'bias': jax.numpy.array([0.0, 0.0])
    }
}

ARGS = {
    'path': temp_file,
    'state': template_state
}
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 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
maybe_do Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/jax_utils.py
code served (permissive licence) · get_code("84bfa52883247bc1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp

def maybe_do(pred, do_fn, operand):
    def body_fn(_, operand):
        return do_fn(operand)
    return jax.lax.fori_loop(0, jnp.asarray(pred, jnp.int32), body_fn, operand)


# Define a simple function that takes an operand and returns a modified version
def simple_do_fn(operand):
    """Simple function that increments the operand by 1"""
    return operand + 1.0


# Create the arguments
pred = 3  # Number of iterations to run
do_fn = simple_do_fn  # Function to apply in the loop
operand = jnp.array([1.0, 2.0, 3.0, 4.0])  # Initial operand (4-element array)

ARGS = {
    "pred": pred,
    "do_fn": do_fn,
    "operand": operand,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
maybe_static_cond Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/jax_utils.py
code served (permissive licence) · get_code("d2feb266da7afc46")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import numpy as onp

# Define simple functions that will be used as true_fn and false_fn
def true_fn(val):
    """Function to execute when predicate is true."""
    return val * 2

def false_fn(val):
    """Function to execute when predicate is false."""
    return val + 1

# Create test values
test_val = jnp.array([1.0, 2.0, 3.0, 4.0])

# Test with numpy array predicate (compile-time determinable)
pred_numpy = onp.array(True)

# Test with boolean predicate (compile-time determinable)
pred_bool = True

# Test with JAX array predicate (runtime-dependent, will use jax.lax.cond)
pred_jax = jnp.array(True)

# Build ARGS dict with all parameters
ARGS = {
    "pred": pred_bool,
    "true_fn": true_fn,
    "false_fn": false_fn,
    "val": test_val
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
orthogonalize_via_newton_schulz Ran
ran on a synthesized input
amoudgl/celo/celo/optimizers/celo2_optax.py
code served (permissive licence) · get_code("66d025f06569a161")
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 (4, 3) 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')

# Create a small matrix for orthogonalization
# Using distinct dimensions: 4 rows, 3 columns to make shapes unambiguous
x = jnp.array([
    [1.0, 2.0, 3.0],
    [4.0, 5.0, 6.0],
    [7.0, 8.0, 9.0],
    [10.0, 11.0, 12.0]
], dtype=jnp.float32)

# ns_coeffs are the coefficients used in Newton-Schulz iteration
# These are typical coefficients for the Newton-Schulz method
ns_coeffs = jnp.array([3.0, -3.0, 1.0], dtype=jnp.float32)

# Number of Newton-Schulz iteration steps
ns_steps = 5

# Small epsilon value for numerical stability
eps = 1e-8

ARGS = {
    'x': x,
    'ns_coeffs': ns_coeffs,
    'ns_steps': ns_steps,
    '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 (4, 3) — shape (4, 3) 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
second_moment_normalizer Ran
ran; our contract draft was wrong, not the code
amoudgl/celo/celo/optimizers/adafac_mlp_lopt.py
code served (permissive licence) · get_code("d18fac6308869679")
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.numpy as jnp
import jax

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

# Create a small test array
x = jnp.array([[1.0, 2.0, 3.0, 4.0],
               [5.0, 6.0, 7.0, 8.0]])

# axis parameter - normalize over the last dimension
axis = 1

# eps parameter - small value for numerical stability
eps = 1e-05

ARGS = {
    'x': x,
    'axis': axis,
    '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
tanh_embedding Ran
ran, honoured the contract we drafted
amoudgl/celo/celo/optimizers/adafac_mlp_lopt.py
code served (permissive licence) · get_code("888ee6dfa73adb32")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (11,) 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

# The function tanh_embedding takes a single parameter x
# Looking at the function body:
# - x is used in the expression: jnp.tanh(x / f32(timescale) - 1.0)
# - This means x should be a JAX array (scalar or array-like)
# - The function will broadcast x across the vmap over timescales

# Create a small JAX array as input
x = jnp.array(5.0, dtype=jnp.float32)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (11,) — shape (11,) 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
threaded_tqdm_map Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/notebook_utils.py
code served (permissive licence) · get_code("56f6fcb69aa5d0e8")
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 list. 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 torch.nn as nn
from concurrent import futures
from typing import Any, Callable, Sequence, Tuple, Optional
import tqdm

torch.manual_seed(0)

# Define a simple function that takes a value and returns a processed value
def simple_process_func(x: Any) -> Any:
    """A simple function that processes input data."""
    if isinstance(x, torch.Tensor):
        return x * 2
    elif isinstance(x, int):
        return x + 1
    else:
        return x

# Create a small dataset with various data types
data_list = [
    torch.tensor([1.0, 2.0, 3.0, 4.0]),
    torch.tensor([5.0, 6.0, 7.0, 8.0]),
    42,
    "test",
    {"key": "value"}
]

ARGS = {
    "threads": 2,
    "func": simple_process_func,
    "data": data_list
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list 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
threaded_tqdm_map Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/py_utils.py
code served (permissive licence) · get_code("2a3363443caf8537")
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 list. 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
from concurrent import futures
from typing import Any, Callable, Sequence
import tqdm

torch.manual_seed(0)

# Define a simple function that processes data
def simple_process_func(x):
    """A simple function that takes input and returns processed output"""
    if isinstance(x, dict):
        return {k: v * 2 for k, v in x.items()}
    elif isinstance(x, (list, tuple)):
        return [item * 2 for item in x]
    else:
        return x * 2

# Create sample data
data = [
    {"value": 1, "count": 5},
    {"value": 2, "count": 10},
    {"value": 3, "count": 15},
    {"value": 4, "count": 20},
    [1, 2, 3],
    [4, 5, 6],
    [7, 8, 9],
    [10, 11, 12],
]

ARGS = {
    "threads": 2,
    "func": simple_process_func,
    "data": data,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list 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
uniquify_server_name Ran
ran on a synthesized input
google/learned_optimization/learned_optimization/distributed.py
code served (permissive licence) · get_code("c5f32a000fc8d76e")
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 str. 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 hashlib
from absl import logging

ARGS = {
    "server_name": "worker_0",
    "experiment_name": "my_training_experiment_v1"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
cached_jit Not yet run
unverified (harvested, no recorded run)
google/learned_optimization/learned_optimization/jax_utils.py
code served (permissive licence) · get_code("4a6a734cd7a980fb")
How it ran

no invocation recorded.

copy Not yet run
unverified (harvested, no recorded run)
google/learned_optimization/learned_optimization/filesystem.py
code served (permissive licence) · get_code("7b798c49a45c7bd4")
How it ran

no invocation recorded.

load_checkpoint Not yet run
unverified (harvested, no recorded run)
amoudgl/celo/celo/optimizers/celo2_optax.py
code served (permissive licence) · get_code("ffce6a5291c41f40")
How it ran

no invocation recorded.

make_dirs Not yet run
unverified (harvested, no recorded run)
google/learned_optimization/learned_optimization/filesystem.py
code served (permissive licence) · get_code("acfa17a3ae480c74")
How it ran

no invocation recorded.

print_param_routing Not yet run
unverified (harvested, no recorded run)
amoudgl/celo/celo/optimizers/celo2.py
code served (permissive licence) · get_code("cab026933643448f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 20 samples lifted from this paper's linked repositories, 20 carry a function identity and 0 do not. Of the 19 distinct functions with an identity, 3 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 3 later) and 16 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
decay_to_param 1 other paper
1 later
2607.06772 (2026-07)
Efficient Long-Horizon Learning for Learned Optimization
second_moment_normalizer 1 other paper
1 later
2607.06772 (2026-07)
Efficient Long-Horizon Learning for Learned Optimization
tanh_embedding 1 other paper
1 later
2607.06772 (2026-07)
Efficient Long-Horizon Learning for Learned Optimization

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 deep learning models have replaced hand-designed features across many domains, these models are still trained with hand-designed optimizers. In this work, we leverage the same scaling approach behind the success of deep learning to learn versatile optimizers. We train an optimizer for deep learning which is itself a small neural network that ingests gradients and outputs parameter updates. Meta-trained with approximately four thousand TPU-months of compute on a wide variety of optimization tasks, our optimizer not only exhibits compelling performance, but optimizes in interesting and unexpected ways. It requires no hyperparameter tuning, instead automatically adapting to the specifics of the problem being optimized. We open source our learned optimizer, meta-training code, the associated train and test data, and an extensive optimizer benchmark suite with baselines at velo-code.github.io.

For agents

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

get_harvested_code_for_paper("2211.09760")
get_code_for_paper("2211.09760")
have("2211.09760")

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

curl https://syntology.ai/api/ran/2211.09760.json

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

[![Syntology run record](https://syntology.ai/api/ran/2211.09760.svg)](https://syntology.ai/paper/2211.09760)

Connect an agent — have() is free.