SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2501.19403 · ICML · 2026

Tackling Fake Forgetting through Uncertainty Quantification

Sijia Liu, Kaize Ding, Ren Wang, Yingdan Shi

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 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; 0 honoured the contract we drafted, 0 violated it and 9 have no contract verdict; 1 failure 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/2501.19403.json · badge: /api/ran/2501.19403.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
TIML-Group/Conformal-Prediction-Unlearning canonical (classifier label)
classifier basis not recorded on the link
10 ran · 3 no recorded run
FunctionStatusWhere it lives
ResNet18 Ran
ran; our contract draft was wrong, not the code
TIML-Group/Conformal-Prediction-Unlearning/models/resnet.py
code served (permissive licence) · get_code("f58940e16113096b")
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

torch.manual_seed(0)

ARGS = {
    "num_classes": 10
}
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 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
accuracy Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/utils.py
code served (permissive licence) · get_code("dc509162e5147173")
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 torch

torch.manual_seed(0)

# Create sample outputs and labels for the accuracy function
# outputs: logits/probabilities with shape (batch_size, num_classes)
# labels: ground truth class indices with shape (batch_size,)

batch_size = 2
num_classes = 4

outputs = torch.randn(batch_size, num_classes)
labels = torch.tensor([0, 2])

ARGS = {
    "outputs": outputs,
    "labels": labels
}
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, 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
calc_sparsity Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/SFRon.py
code served (permissive licence) · get_code("0901b9d51742b200")
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 torch

torch.manual_seed(0)

# Create a small tensor with some zero elements for sparsity calculation
tensor = torch.tensor([
    [1.0, 0.0, 2.0, 0.0],
    [0.0, 3.0, 0.0, 4.0]
], dtype=torch.float32)

ARGS = {
    "tensor": tensor
}
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, 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
expdecay_lr_scheduler Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/SFRon.py
code served (permissive licence) · get_code("6ae0abac6209c398")
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 float, 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 torch
import numpy as np

torch.manual_seed(0)

# Analyze the function signature and body:
# def expdecay_lr_scheduler(base_lr, current_epoch, T_max, base=2):
#     return base_lr * (1 - current_epoch / T_max) ** base
#
# Parameters:
# - base_lr: float, used in multiplication
# - current_epoch: int/float, used in arithmetic
# - T_max: int/float, used in division
# - base: int/float with default=2, used as exponent

ARGS = {
    "base_lr": 0.1,
    "current_epoch": 5,
    "T_max": 100,
    "base": 2
}
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 float, 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
find_quantile Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/unlearn_cpu.py
code served (permissive licence) · get_code("a0ecfc58ff52a44e")
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 float64, ndim 0, 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 numpy as np

# Create calibration scores - a simple numpy array
cal_scores = np.array([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0])

# Set n to the number of calibration samples
n = 10

# Set alpha (significance level, typically between 0 and 1)
alpha = 0.1

ARGS = {
    "cal_scores": cal_scores,
    "n": n,
    "alpha": alpha
}
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 float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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_cpu_loss Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/unlearn_cpu.py
code served (permissive licence) · get_code("b36e593eed15201e")
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 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 torch

torch.manual_seed(0)

# Construct arguments for get_cpu_loss
batch_size = 2
num_classes = 4

# out_softmax: batch_size x num_classes tensor
# The function accesses out_softmax[:, clabels], so we need softmax probabilities
out_softmax = torch.rand(batch_size, num_classes)

# clabels: class labels for each sample in batch
# Must be valid indices into the second dimension of out_softmax
clabels = torch.tensor([0, 2])

# q_hat: scalar or tensor representing some threshold/target
# Used in: torch.clamp(q_hat - score + delta, min=0)
q_hat = torch.tensor(0.8)

# delta: scalar or tensor, added to the loss computation
# Used in: torch.clamp(q_hat - score + delta, min=0)
delta = torch.tensor(0.1)

# device: string specifying device
device = 'cpu'

# unlearn_type: string indicating unlearn method (not used in function body)
unlearn_type = 'gradient_ascent'

ARGS = {
    'out_sofmax': out_softmax,
    'clabels': clabels,
    'q_hat': q_hat,
    'delta': delta,
    'device': device,
    'unlearn_type': unlearn_type
}
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 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
linear_lr_scheduler Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/SFRon.py
code served (permissive licence) · get_code("e7e3ea7a9ca943ac")
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 float, 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 torch
import numpy as np

torch.manual_seed(0)

# Analyze the function signature and body:
# - base_lr: a learning rate (float)
# - current_epoch: current epoch number (int or float)
# - T_max: maximum number of epochs (int or float)
# - base: exponent parameter (float, default=1)
# The function computes: base_lr * (1 - current_epoch / T_max) ** base
# All parameters are numeric types used in arithmetic operations

ARGS = {
    "base_lr": 0.1,
    "current_epoch": 5,
    "T_max": 100,
    "base": 1
}
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 float, 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
test_loop Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/main_original_model.py
code served (permissive licence) · get_code("518b89359b3be8bc")
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 torch
import torch.nn as nn

torch.manual_seed(0)

# Define a small model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)  # 4 input dims, 3 output classes
    
    def forward(self, x):
        return self.linear(x)

# Create model
model = SimpleModel()

# Create a simple dataloader (list of (x, y) tuples)
# Batch size 2, input dim 4, 3 classes
dataloader_data = [
    (torch.randn(2, 4), torch.randint(0, 3, (2,))),
    (torch.randn(2, 4), torch.randint(0, 3, (2,))),
]

class SimpleDataLoader:
    def __init__(self, data):
        self.data = data
        self.dataset = type('obj', (object,), {'__len__': lambda self: sum(len(batch[1]) for batch in dataloader_data)})()
    
    def __iter__(self):
        return iter(self.data)
    
    def __len__(self):
        return len(self.data)

dataloader = SimpleDataLoader(dataloader_data)

# Define loss function
loss_fn = nn.CrossEntropyLoss()

# Device
device = 'cpu'

ARGS = {
    'dataloader': dataloader,
    'model': model,
    'loss_fn': loss_fn,
    'device': device
}
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 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
train_loop Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/main_original_model.py
code served (permissive licence) · get_code("28bf56b45f8ff355")
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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a simple model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.fc(x)

model = SimpleModel()

# Create a simple dataloader (list of (x, labels) tuples)
batch_size = 2
num_batches = 3
dataloader = []
for _ in range(num_batches):
    x = torch.randn(batch_size, 4)
    labels = torch.randint(0, 3, (batch_size,))
    dataloader.append((x, labels))

# Add a dataset attribute to the dataloader
class SimpleDataLoader(list):
    def __init__(self, data):
        super().__init__(data)
        self.dataset = self
    
    def __len__(self):
        return sum(len(batch[0]) for batch in self)

dataloader = SimpleDataLoader(dataloader)

# Create loss function
loss_fn = nn.CrossEntropyLoss()

# Create optimizer
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# Device
device = torch.device('cpu')

ARGS = {
    'dataloader': dataloader,
    'model': model,
    'loss_fn': loss_fn,
    'optimizer': optimizer,
    'device': device
}
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 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
training_step Ran
ran on a synthesized input
TIML-Group/Conformal-Prediction-Unlearning/unlearn.py
code served (permissive licence) · get_code("f27cb338aad835ff")
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
import torch.nn as nn

torch.manual_seed(0)

# Define a minimal model that takes images and outputs class logits
class SimpleClassifier(nn.Module):
    def __init__(self, input_dim=4, num_classes=3):
        super().__init__()
        self.fc = nn.Linear(input_dim, num_classes)
    
    def forward(self, x):
        return self.fc(x)

# Create model
model = SimpleClassifier(input_dim=4, num_classes=3)

# Create batch: (images, class_labels)
# images: batch_size=2, feature_dim=4
# clabels: batch_size=2, with class indices in range [0, 2]
images = torch.randn(2, 4)
clabels = torch.tensor([0, 2])
batch = (images, clabels)

# Create criterion (CrossEntropyLoss expects logits and class indices)
criterion = nn.CrossEntropyLoss()

# Device (CPU only)
device = 'cpu'

ARGS = {
    'model': model,
    'batch': batch,
    'criterion': criterion,
    'device': device
}
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
generate_dataset Not yet run
unverified (harvested, no recorded run)
TIML-Group/Conformal-Prediction-Unlearning/utils.py
code served (permissive licence) · get_code("aa200986af8daf4a")
How it ran

no invocation recorded.

get_calibration Not yet run
unverified (harvested, no recorded run)
TIML-Group/Conformal-Prediction-Unlearning/unlearn_cpu.py
code served (permissive licence) · get_code("ea45d6b3c8324c43")
How it ran

no invocation recorded.

get_dataset Not yet run
unverified (harvested, no recorded run)
TIML-Group/Conformal-Prediction-Unlearning/utils.py
code served (permissive licence) · get_code("3ea50daa1a9dcd36")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 12 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
ResNet18 4 other papers
4 dated earlier
2105.00303 (2021-05)
RATT: Leveraging Unlabeled Data to Guarantee Generalization
2109.13398 (2021-09)
Unrolling SGD: Understanding Factors Influencing Machine Unlearning
2308.04142 (2023-08)
Class-level Structural Relation Modelling and Smoothing for Visual Representation Learning
and 1 more not named here

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

Machine unlearning seeks to remove the influence of specified data from a trained model. While the unlearning accuracy is a widely used metric for assessing unlearning performance, it falls short in assessing the reliability of forgetting. In this paper, we find that the forget data points misclassified by unlearning accuracy still have their ground truth labels included in the conformal prediction set from the uncertainty quantification perspective, leading to a phenomenon we term fake forgetting. To address this issue, we propose a novel metric CR, inspired by conformal prediction, that offers a more reliable assessment of forgetting quality. Building on these insights, we further propose an unlearning framework CPU that incorporates conformal prediction into the Carlini & Wagner adversarial attack loss, enabling the ground truth label to be effectively removed from the conformal prediction set. Through extensive experiments on image classification tasks, we demonstrate both the effectiveness of our proposed metric and the superior forgetting quality achieved by our framework. Code is available at https://github.com/TIML-Group/Conformal-Prediction-Unlearning. visualize the important feature of models' prediction by using Grad-CAM [28], further 37 phenomenon. Despite the Finetune method incorrectly predicting the forget data, the 38 maps still focus heavily on the important features of the object itself. 39 Based on the above findings, we design two novel metrics that capture the uncertainty an 40 of unlearning performance more effectively inspired by conformal prediction. Additionall 41 by Carlini & Wagner (C&W) attack [6] and conformal prediction, we propose a genera 42 framework to improve training-based unlearning methods and promote reliable forgetting. 43 maps of our method in Table 1 reveal that once the true label no longer falls within th 44 prediction set, the activation regions shift significantly. To sum up, the contributions of o 45 as follows: 46 • We identify pivotal limitations in current unlearning metrics, as they overlook m 47 data where ground truth labels remain potential candidates under uncertainty qu 48 • We design two novel metrics to address the limitations motivated by conforma 49 • We propose a general unlearning framework for training-based machine unlearn 50 motivated by conformal prediction and C&W loss. 51 • Extensive experiments demonstrate the effectiveness of novel metrics and ou 52 framework. 53 2 Enhancing Metrics for Machine Unlearning Based on Conformal P 54 2.1 Preliminaries and Notations 55 Machine Unlearning. Machine unlearning is the targeted removal of certain training 56 from a machine learning model. In our work, two different forgetting scenarios are co 57 random data forgetting focuses on randomly forgetting specific instances within the traini 58 (ii) class-wise forgetting aims to remove all information associated with an entire class 59 denote the original training data used to obtain an original model ✓ o . We split the wh 60 data D train into two subsets, forget data D f and retain data D r = D train \ D f . In r 61 forgetting, D test represents test data. In class-wise forgetting, D tf corresponds to the tes 62 Tackling Fake Forgetting through Uncertainty Quantification forgetting focuses on randomly forgetting specific data instances within the training data, and (ii) class-wise forgetting aims to remove all data information associated with an entire class. We also report the results of the worst-case and subclass-wise forgetting scenarios in Tables 11 and12 in Appendix G. Let D train denote the original training data used to obtain an original model θ o . We split the whole training data D train into two subsets, forget data D f and retain data D r = D train \ D f . Let D test represent test data. θ u denotes the model after the unlearning process. Conformal Prediction. Conformal prediction (CP) is proposed to quantify uncertainty, providing prediction sets that contain the ground truth label with a theoretically guaranteed probability (Angelopoulos & Bates, 2021). Among the various types of conformal prediction, this work mainly focuses on split conformal prediction (SCP) 1 since it is the most straightforward and easy-to-implement approach. We also report results of other conformal prediction techniques in Appendix F.1. To construct a conformal prediction set, SCP involves four steps on the unlearning model:

For agents

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

get_harvested_code_for_paper("2501.19403")
get_code_for_paper("2501.19403")
have("2501.19403")

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

curl https://syntology.ai/api/ran/2501.19403.json

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

[![Syntology run record](https://syntology.ai/api/ran/2501.19403.svg)](https://syntology.ai/paper/2501.19403)

Connect an agent — have() is free.