SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2307.13883 · ICLR · 2024

ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis

U Berkeley, Charles Sutton, Manzil Zaheer, Joey Hong, Pengcheng Yin, Kensen Shi, Yinlin Deng

arXiv · PDF · Open in the Atlas

Code that ran

13 of 16 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, 2 violated it and 0 have no contract verdict; 10 failures of our instrument; 3 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/2307.13883.json · badge: /api/ran/2307.13883.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-deepmind/exedec canonical (classifier label)
classifier basis: census
13 ran · 3 no recorded run
FunctionStatusWhere it lives
add_beam_dim Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/decode.py
code served (permissive licence) · get_code("323ec9d2adf0a81f")
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 3, dtype family float, shape (2, 3, 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

# Create test arguments for add_beam_dim
# The function expects:
# - x: an array (scalar or non-scalar)
# - beam_size: an integer for tiling

# Test case 1: non-scalar array
x_non_scalar = jnp.ones((2, 4))  # batch_size=2, feature_dim=4

# Test case 2: scalar array
x_scalar = jnp.array(5.0)

# beam_size parameter
beam_size = 3

# Primary ARGS dict using non-scalar input
ARGS = {
    "x": x_non_scalar,
    "beam_size": beam_size
}
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 3, dtype family float, shape (2, 3, 4) — shape (2, 3, 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
brevity_penalty Ran
ran, honoured the contract we drafted
google-deepmind/exedec/spec_decomposition/decode.py
code served (permissive licence) · get_code("075fd59d804e340b")
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.numpy as jnp

# Construct arguments for brevity_penalty function
alpha = 0.5  # float: brevity-penalty scaling parameter
length = 10  # int: length of considered sequence

ARGS = {
    "alpha": alpha,
    "length": length,
}
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
create_dataset Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/llm_utils.py
code served (permissive licence) · get_code("4b1a9cf5feab44ca")
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 _MapDataset. 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 directory for test TFRecord files
temp_dir = tempfile.mkdtemp()

def create_tfrecord_file(filepath, num_examples=2):
    """Helper to create a minimal valid TFRecord file."""
    writer = tf.io.TFRecordWriter(filepath)
    
    # Create a few example records
    for _ in range(2):
        inputs = ['input_' + str(i) for i in range(num_examples)]
        outputs = ['output_' + str(i) for i in range(num_examples)]
        program = 'def program(): pass'
        
        feature = {
            'inputs': tf.train.Feature(bytes_list=tf.train.BytesList(
                value=[inp.encode('utf-8') for inp in inputs]
            )),
            'outputs': tf.train.Feature(bytes_list=tf.train.BytesList(
                value=[out.encode('utf-8') for out in outputs]
            )),
            'program': tf.train.Feature(bytes_list=tf.train.BytesList(
                value=[program.encode('utf-8')]
            )),
        }
        
        example = tf.train.Example(features=tf.train.Features(feature=feature))
        writer.write(example.SerializeToString())
    
    writer.close()

# Create test TFRecord files
num_examples = 3
test_file_1 = os.path.join(temp_dir, 'test_0.tfrecord')
test_file_2 = os.path.join(temp_dir, 'test_1.tfrecord')
create_tfrecord_file(test_file_1, num_examples)
create_tfrecord_file(test_file_2, num_examples)

# Create file pattern that matches the test files
file_pattern = os.path.join(temp_dir, 'test_*.tfrecord')

ARGS = {
    'file_pattern': file_pattern,
    'num_examples': num_examples,
}
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 _MapDataset — a _MapDataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a _MapDataset 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
flatten_beam_dim Ran
ran, violated the contract we drafted
google-deepmind/exedec/spec_decomposition/decode.py
code served (permissive licence) · get_code("758199a57cc4780a")
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 Tensor, ndim 2, dtype family float, shape (6, 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 numpy as np

torch.manual_seed(0)

# The function flatten_beam_dim takes an array-like object (tensor or numpy array)
# and flattens its first two dimensions while preserving the rest.
# We need to test with:
# 1. A scalar (ndim == 0) - should return unchanged
# 2. A multi-dimensional array - should flatten first two dims

# Test case: a 3D tensor with shape (2, 3, 4)
# After flattening first two dims: (6, 4)
x = torch.randn(2, 3, 4)

ARGS = {
    "x": x
}
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 Tensor, ndim 2, dtype family float, shape (6, 4) — shape (6, 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
flatten_num_io_dim Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/models/base_models.py
code served (permissive licence) · get_code("8ba7238984c6122b")
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 Tensor, ndim 3, dtype family float, shape (2, 12, 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 torch
import numpy as np

torch.manual_seed(0)

# The function takes a tensor/array x and an optional axis parameter
# It reshapes the tensor by flattening dimensions at and after the axis
# into a single dimension

# Create a test tensor with distinct dimensions
# Shape: (batch=2, num_io_1=4, num_io_2=3, feature=5)
# This allows testing the flattening of dimensions at axis
x = torch.randn(2, 4, 3, 5)

# The axis parameter specifies where to start flattening
# Default is 1, which will flatten dimensions 1 and 2 together
axis = 1

ARGS = {
    "x": x,
    "axis": axis,
}
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 Tensor, ndim 3, dtype family float, shape (2, 12, 5) — shape (2, 12, 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
get_boolean_options Ran
ran, violated the contract we drafted
google-deepmind/exedec/spec_decomposition/launch_end_to_end_predict.py
code served (permissive licence) · get_code("6521c64803ccbc0e")
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.

from enum import Enum

class Flag(Enum):
    TRUE = 'true'
    FALSE = 'false'
    BOTH = 'both'

ARGS = {
    'flag': Flag.TRUE
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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
make_aligned_relative_position Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/decomposition_models.py
code served (permissive licence) · get_code("195e13318aaa4b57")
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 3, dtype family int, shape (2, 5, 6) 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

# Construct small test tensors for the function
# targets: shape (batch_size, sequence_length) - contains token IDs including separator tokens
batch_size = 2
seq_length = 5
separator_token = 0

targets = jnp.array([
    [1, 2, 0, 3, 4],
    [5, 0, 6, 7, 0]
], dtype=jnp.int32)

# flat_encoded: shape (batch_size, flat_encoded_length, embedding_dim)
# The function uses flat_encoded.shape[-2] to get the sequence length
flat_encoded_length = 6
flat_encoded = jnp.ones((batch_size, flat_encoded_length, 8), dtype=jnp.float32)

# max_input_length: scalar value used for computing aligned positions
max_input_length = 10

# dtype: output dtype
dtype = jnp.int32

ARGS = {
    "targets": targets,
    "flat_encoded": flat_encoded,
    "max_input_length": max_input_length,
    "separator_token": separator_token,
    "dtype": dtype,
}
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 3, dtype family int, shape (2, 5, 6) — shape (2, 5, 6) 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
multinomial Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/old_decode.py
code served (permissive licence) · get_code("ee2a82c6ca375fca")
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 int, shape (2,) 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 seed for reproducibility
jax.config.update("jax_default_prng_impl", "threefry2x32")

# Create a PRNG key
rng = jax.random.PRNGKey(0)

# Create logits array of shape (..., num_categories)
# Using shape (2, 4) where batch_size=2, num_categories=4
logits = jnp.array([
    [1.0, 2.0, 0.5, 1.5],
    [0.3, 1.2, 2.1, 0.8]
])

ARGS = {
    "rng": rng,
    "logits": logits,
}
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 int, shape (2,) — shape (2,) 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
product_sweep Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/launch_end_to_end_predict.py
code served (permissive licence) · get_code("914b193e68f5628d")
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 itertools
from typing import Any

def product_sweep(all_name_and_values: list[tuple[str, list[Any]]]) -> dict[str, dict[str, Any]]:
    names = []
    value_lists = []
    for name, value_list in all_name_and_values:
        names.append(name)
        value_lists.append(value_list)
    return [dict(zip(names, choices)) for choices in itertools.product(*value_lists)]


# Setup code to construct valid arguments for product_sweep
all_name_and_values = [
    ("learning_rate", [0.001, 0.01, 0.1]),
    ("batch_size", [16, 32]),
    ("activation", ["relu", "tanh"]),
]

ARGS = {
    "all_name_and_values": all_name_and_values,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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
shift_right Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google-deepmind/exedec/models/base_models.py
code served (permissive licence) · get_code("8705b2614e592b74")
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 int, 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

# Create a small input tensor for the shift_right function
# The function expects a JAX array and will pad it on the last axis
x = jnp.array([[1, 2, 3, 4],
               [5, 6, 7, 8]], dtype=jnp.int32)

bos_token = 0

ARGS = {
    "x": x,
    "bos_token": bos_token
}
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 int, 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
solver_parallel_run Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/run_llm_experiment.py
code served (permissive licence) · get_code("7484b09be8359551")
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 multiprocessing.dummy
from typing import Any, Callable
import tqdm

torch.manual_seed(0)

def sample_solver(index, problem_data):
    """
    Sample solver function that takes an index and problem data,
    and returns a result dict with 'success' and 'index' keys.
    """
    # Simulate some computation
    result = {
        'success': (index + problem_data['value']) % 2 == 0,
        'index': index,
        'output': problem_data['value'] * 2
    }
    return result

# Create test inputs for solver_parallel_run
# Each input should be a tuple of (index, problem_data) for starmap
inputs = [
    (0, {'value': 10}),
    (1, {'value': 20}),
    (2, {'value': 30}),
]

ARGS = {
    'func': sample_solver,
    'inputs': inputs,
    'num_workers': 2
}
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, 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
to_python_form Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/spec_decomposition/llm_utils.py
code served (permissive licence) · get_code("de375dd399608a9a")
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 re

def to_python_form(io: str) -> str:
    """Convert Deepcoder's "x1 = [ 1 2 ] | x2 = 3" into "x1 = [1, 2], x2 = 3"."""
    io = io.replace(' | ', ', ').replace('[ ', '[').replace(' ]', ']')
    io = re.sub('(?<=\\d) (?=-|\\d)', ', ', io)
    return io

ARGS = {
    "io": "x1 = [ 1 2 ] | x2 = 3"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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
unflatten_num_io_dim Ran
ran; our contract draft was wrong, not the code
google-deepmind/exedec/models/base_models.py
code served (permissive licence) · get_code("7688708073c0de19")
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 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 numpy as np

torch.manual_seed(0)

# Create a test tensor with shape that allows unflattening
# Shape: (batch=2, length=12, features=4)
# This allows unflatten_num_io_dim to work with num_io=3 (12 % 3 == 0)
x = torch.randn(2, 12, 4)

num_io = 3
axis = 1

ARGS = {
    "x": x,
    "num_io": num_io,
    "axis": axis,
}
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 4) — shape (2, 3, 4, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_sweep Not yet run
unverified (harvested, no recorded run)
google-deepmind/exedec/spec_decomposition/launch_train.py
code served (permissive licence) · get_code("6da1245365dfd0a7")
How it ran

no invocation recorded.

make_relative_position_bucket Not yet run
unverified (harvested, no recorded run)
google-deepmind/exedec/models/relative_attention.py
code served (permissive licence) · get_code("bd6bc5a2d5328634")
How it ran

no invocation recorded.

query_llm Not yet run
unverified (harvested, no recorded run)
google-deepmind/exedec/spec_decomposition/cached_llm_access.py
code served (permissive licence) · get_code("d204dbbda2b2669f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 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.

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

When writing programs, people have the ability to tackle a new complex task by decomposing it into smaller and more familiar subtasks. While it is difficult to measure whether neural program synthesis methods have similar capabilities, we can measure whether they compositionally generalize, that is, whether a model that has been trained on the simpler subtasks is subsequently able to solve more complex tasks. In this paper, we characterize several different forms of compositional generalization that are desirable in program synthesis, forming a meta-benchmark which we use to create generalization tasks for two popular datasets, RobustFill and DeepCoder. We then propose ExeDec, a novel decomposition-based synthesis strategy that predicts execution subgoals to solve problems step-by-step informed by program execution at each step. When used with Transformer models trained from scratch, ExeDec has better synthesis performance and greatly improved compositional generalization ability compared to baselines. Finally, we use our benchmarks to demonstrate that LLMs struggle to compositionally generalize when asked to do programming-by-example in a few-shot setting, but an ExeDec-style prompting approach can improve the generalization ability and overall performance.

For agents

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

get_harvested_code_for_paper("2307.13883")
get_code_for_paper("2307.13883")
have("2307.13883")

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

curl https://syntology.ai/api/ran/2307.13883.json

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

[![Syntology run record](https://syntology.ai/api/ran/2307.13883.svg)](https://syntology.ai/paper/2307.13883)

Connect an agent — have() is free.