SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.02167 · ICML · 2026

Manifold-Aligned Guided Integrated Gradients for Reliable Feature Attribution

Kyowoon Lee, Jaesik Choi, Soyeon Kim, Seongwoo Lim

arXiv · PDF · Open in the Atlas

Code that ran

10 of 11 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; 2 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 8 failures of our instrument; 1 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/2605.02167.json · badge: /api/ran/2605.02167.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
leekwoon/ma-gig canonical (classifier label)
classifier basis: census
10 ran · 1 no recorded run
FunctionStatusWhere it lives
abs_grayscale_norm Ran
ran, honoured the contract we drafted
leekwoon/ma-gig/cleanig/plot_utils.py
code served (permissive licence) · get_code("18854f664dd257b3")
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 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 numpy as np

# Test case 1: 2D grayscale image
img_2d = np.random.rand(4, 3).astype(np.float32) * 255

# Test case 2: 3D color image (height, width, channels)
img_3d = np.random.rand(5, 6, 3).astype(np.float32) * 255

ARGS = {
    "img": img_2d
}
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 ndarray, 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
adopt_weight Ran
ran, honoured the contract we drafted
leekwoon/ma-gig/cleanig/mar_vae/loss.py
pointer only (licence: MIT); the stored copy get_code serves, from joanrod/ocr-vqgan, is pointer only (licence: NONE) · get_code("7439a9d4c6be04b0")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from joanrod/ocr-vqgan, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_classifier Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/classifier/utils.py
code served (permissive licence) · get_code("b82506d701961b39")
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 ResNet. 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
import torchvision.models as models

torch.manual_seed(0)

ARGS = {
    "model_name": "resnet18",
    "dataset_name": "oxfordflower",
    "image_size": 224,
    "num_classes": 102,
    "pretrained": False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResNet — a ResNet output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ResNet 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
get_sample_batch Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/utils.py
code served (permissive licence) · get_code("72d03510affd7b7c")
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 torch

torch.manual_seed(0)

# Create a simple dataloader as a list of (images, labels) tuples
batch_images_1 = torch.randn(2, 3, 4, 4)  # 2 samples, 3 channels, 4x4 spatial
batch_labels_1 = torch.tensor([0, 1])

batch_images_2 = torch.randn(2, 3, 4, 4)
batch_labels_2 = torch.tensor([2, 3])

dataloader = [
    (batch_images_1, batch_labels_1),
    (batch_images_2, batch_labels_2),
]

ARGS = {
    "dataloader": dataloader,
    "n_samples": 3,
    "return_labels": True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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
hinge_d_loss Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/mar_vae/loss.py
pointer only (licence: MIT); the stored copy get_code serves, from joanrod/ocr-vqgan, is pointer only (licence: NONE) · get_code("22a83df891d750c6")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from joanrod/ocr-vqgan, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

load_imagenet_datasets Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/dataset/imagenet_dataset.py
code served (permissive licence) · get_code("a557ab81c42a4832")
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 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 os
import tempfile
import shutil
from PIL import Image
import torch
from torch.utils.data import DataLoader
from torchvision import datasets, transforms

torch.manual_seed(0)

# Create a temporary directory structure mimicking ImageNet layout
temp_dir = tempfile.mkdtemp()
train_dir = os.path.join(temp_dir, 'train')
val_dir = os.path.join(temp_dir, 'val')

os.makedirs(train_dir, exist_ok=True)
os.makedirs(val_dir, exist_ok=True)

# Create two dummy classes with a few dummy images each
for split_dir in [train_dir, val_dir]:
    for class_idx in range(2):
        class_dir = os.path.join(split_dir, f'class_{class_idx}')
        os.makedirs(class_dir, exist_ok=True)
        
        # Create a few tiny dummy images
        for img_idx in range(2):
            img = Image.new('RGB', (64, 64), color=(73 + class_idx * 10, 109 + img_idx * 5, 137))
            img.save(os.path.join(class_dir, f'image_{img_idx}.jpg'))

ARGS = {
    'dataset_path': temp_dir,
    'test_split': 0.05,
    'image_size': 256,
    'batch_size': 2,
    'num_workers': 0,
    'mean': [0.485, 0.456, 0.406],
    'std': [0.229, 0.224, 0.225],
    'random_flip': False,
    'val_only': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, torch, torchvision — 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 float — 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
load_oxfordpet_datasets Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/dataset/oxfordpet_dataset.py
code served (permissive licence) · get_code("14042876bcf0ef07")
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 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 os
import tempfile
from PIL import Image
import numpy as np

# Create a temporary directory with fake Oxford Pet images
temp_dir = tempfile.mkdtemp()

# Create a few fake .jpg files with pet-like names
pet_names = ['cat_1', 'cat_2', 'dog_1', 'dog_2']
for pet_name in pet_names:
    # Create a small fake image file
    img = Image.new('RGB', (64, 64), color=(73, 109, 137))
    img.save(os.path.join(temp_dir, f'{pet_name}.jpg'))

ARGS = {
    'dataset_path': temp_dir,
    'test_split': 0.25,
    'image_size': 64,
    'batch_size': 2,
    'num_workers': 0,
    'mean': [0.485, 0.456, 0.406],
    'std': [0.229, 0.224, 0.225],
    'random_flip': False,
    'val_only': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, sklearn, torch, torchvision — 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 float — 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
preprocess Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/utils.py
code served (permissive licence) · get_code("7a30c66aa0eb3b7e")
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

torch.manual_seed(0)

# Create a small batch of image data (batch_size=2, channels=3, height=4, width=4)
x = torch.randn(2, 3, 4, 4)

# Create mean and std for 3 channels
mean = torch.tensor([0.5, 0.4, 0.3])
std = torch.tensor([0.2, 0.25, 0.15])

ARGS = {
    "x": x,
    "mean": mean,
    "std": std
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
undo_preprocess Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/utils.py
code served (permissive licence) · get_code("bb3251b543cfa179")
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, 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

torch.manual_seed(0)

# Construct arguments for undo_preprocess
# x: tensor of shape (batch, channels, height, width) with channels=3
x = torch.randn(2, 3, 4, 5)

# mean: tensor or list of 3 values (one per channel)
mean = torch.tensor([0.5, 0.4, 0.3])

# std: tensor or list of 3 values (one per channel)
std = torch.tensor([0.2, 0.25, 0.15])

ARGS = {
    "x": x,
    "mean": mean,
    "std": std,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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, 5) — shape (2, 3, 4, 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
vanilla_d_loss Ran
ran; our contract draft was wrong, not the code
leekwoon/ma-gig/cleanig/mar_vae/loss.py
pointer only (licence: MIT); the stored copy get_code serves, from joanrod/ocr-vqgan, is pointer only (licence: NONE) · get_code("263c54290c553ddd")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from joanrod/ocr-vqgan, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

load_oxfordflower_datasets Not yet run
unverified (harvested, no recorded run)
leekwoon/ma-gig/cleanig/dataset/oxfordflower_dataset.py
code served (permissive licence) · get_code("f2adec6ae07da7dc")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 5 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 0 the same month, 0 later) and 6 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
hinge_d_loss 15 other papers
15 dated earlier
2012.09841 (2020-12)
Taming Transformers for High-Resolution Image Synthesis
2112.00384 (2021-12)
Exploration into Translation-Equivariant Image Quantization
2211.13221 (2022-11)
Latent Video Diffusion Models for High-Fidelity Long Video Generation
and 12 more not named here
vanilla_d_loss 11 other papers
11 dated earlier
2012.09841 (2020-12)
Taming Transformers for High-Resolution Image Synthesis
2112.00384 (2021-12)
Exploration into Translation-Equivariant Image Quantization
2211.13221 (2022-11)
Latent Video Diffusion Models for High-Fidelity Long Video Generation
and 8 more not named here
adopt_weight 9 other papers
9 dated earlier
2004.13166 (2020-04)
A Disentangling Invertible Interpretation Network for Explaining Latent Representations
2012.09841 (2020-12)
Taming Transformers for High-Resolution Image Synthesis
2112.00384 (2021-12)
Exploration into Translation-Equivariant Image Quantization
and 6 more not named here
preprocess 2 other papers
2 dated earlier
1606.00915 (2016-06)
DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs
2410.20722 (2024-10)
Interpretable Image Classification with Adaptive Prototype-based Vision Transformers
undo_preprocess 2 other papers
2 dated earlier
1606.00915 (2016-06)
DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs
2410.20722 (2024-10)
Interpretable Image Classification with Adaptive Prototype-based Vision Transformers

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

Feature attribution is central to diagnosing and trusting deep neural networks, and Integrated Gradients (IG) is widely used due to its axiomatic properties. However, IG can yield unreliable explanations when the integration path between a baseline and the input passes through regions with noisy gradients. While Guided Integrated Gradients reduces this sensitivity by adaptively updating low-gradient-magnitude features, inputspace guidance still produces intermediate inputs that deviate from the data manifold. To address this limitation, we propose Manifold-Aligned Guided Integrated Gradients (MA-GIG), which constructs attribution paths in the latent space of a pre-trained variational autoencoder. By decoding intermediate latent states, MA-GIG biases the path toward the learned generative manifold and reduces exposure to implausible inputspace regions. Through qualitative and quantitative evaluations, we demonstrate that MA-GIG produces faithful explanations by aggregating gradients on path features proximal to the input. Consequently, our method reduces offmanifold noise and outperforms prior path-based attribution methods across multiple datasets and classifiers. Our code is available at https: //github.com/leekwoon/ma-gig/ † Equal advising.

For agents

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

get_harvested_code_for_paper("2605.02167")
get_code_for_paper("2605.02167")
have("2605.02167")

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

curl https://syntology.ai/api/ran/2605.02167.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.02167.svg)](https://syntology.ai/paper/2605.02167)

Connect an agent — have() is free.