SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.13356 · ICLR · 2025

UNLEARNING OR OBFUSCATING? JOGGING THE MEM-ORY OF UNLEARNED LLMS VIA BENIGN RELEARNING CONTENT WARNING: THIS MANUSCRIPT CONTAINS EXAMPLES OF HARMFUL/HAZARDOUS TEXT

Virginia Smith, Shengyuan Hu, Zhiwei Steven, Yiwei Fu

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; 4 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 9 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/2406.13356.json · badge: /api/ran/2406.13356.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
s-huu/jog_llm_memory canonical (classifier label)
classifier basis: census
13 ran · 3 no recorded run
FunctionStatusWhere it lives
collate_fn Ran
ran, honoured the contract we drafted
s-huu/jog_llm_memory/synthetic/data_module.py
code served (permissive licence) · get_code("cdd171f20cd6f53d")
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 3, 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 torch
from torch.nn.utils.rnn import pad_sequence

torch.manual_seed(0)

# Create sample batch data
# Each item in batch is a tuple of (input_ids, attention_mask)
# input_ids and attention_masks should be 1D tensors of different lengths

batch = [
    (torch.tensor([1, 2, 3], dtype=torch.long), torch.tensor([1, 1, 1], dtype=torch.long)),
    (torch.tensor([4, 5], dtype=torch.long), torch.tensor([1, 1], dtype=torch.long)),
]

ARGS = {
    "batch": batch
}
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 3, 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
compute_metrics Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/synthetic/dataloader.py
code served (permissive licence) · get_code("88a4f13abd54a00c")
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 torch
import numpy as np

torch.manual_seed(0)

# Create a simple pred object with predictions and label_ids attributes
class PredObject:
    def __init__(self, predictions, label_ids):
        self.predictions = predictions
        self.label_ids = label_ids

# Create small tensors: batch_size=2, vocab_size=4, seq_length=3
batch_size = 2
vocab_size = 4
seq_length = 3

# predictions shape: (batch_size, seq_length, vocab_size) - logits
predictions = np.random.randn(batch_size, seq_length, vocab_size).astype(np.float32)

# label_ids shape: (batch_size, seq_length) - token indices
# Use values in range [0, vocab_size) or -100 for ignore_index
label_ids = np.array([
    [1, 2, 0],
    [3, 1, -100]
], dtype=np.int64)

# Create pred object
pred = PredObject(predictions, label_ids)

ARGS = {
    'pred': pred
}
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 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
convert_data_to_model_format Ran
ran, honoured the contract we drafted
s-huu/jog_llm_memory/synthetic/data_module.py
code served (permissive licence) · get_code("b4caf62fbc9c93a4")
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 2, 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 torch

torch.manual_seed(0)

class MockTokenizer:
    def __init__(self):
        self.eos_token_id = 2
    
    def __call__(self, text, add_special_tokens=True, max_length=5000, truncation=True):
        # Return a mock encoded object with input_ids and attention_mask
        # Create a small mock result
        mock_encoded = {
            'input_ids': [1, 2, 3, 4],
            'attention_mask': [1, 1, 1, 1]
        }
        # Create an object that supports dict-like access
        class EncodedResult:
            def __init__(self, data):
                self.data = data
                self.input_ids = data['input_ids']
                self.attention_mask = data['attention_mask']
            
            def __getitem__(self, key):
                return self.data[key]
        
        return EncodedResult(mock_encoded)

tokenizer = MockTokenizer()
data = ["This is a sample text for tokenization"]
max_length = 500

ARGS = {
    'data': data,
    'tokenizer': tokenizer,
    'max_length': max_length
}
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 2, 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
custom_data_collator_forget Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
s-huu/jog_llm_memory/synthetic/dataloader.py
code served (permissive licence) · get_code("e951dedc21c69ac1")
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 4, 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 torch

torch.manual_seed(0)

# Construct samples for the data collator
# Each sample is a tuple of (forget_data, retain_data)
# Where each of those is a tuple of (input_ids, labels, attention_mask)

# Create small tensors for forget and retain samples
batch_size = 2
seq_length = 4
vocab_size = 3

forget_sample_1 = (
    torch.randint(0, vocab_size, (seq_length,)),  # input_ids
    torch.randint(0, vocab_size, (seq_length,)),  # labels
    torch.ones(seq_length, dtype=torch.long)      # attention_mask
)

forget_sample_2 = (
    torch.randint(0, vocab_size, (seq_length,)),
    torch.randint(0, vocab_size, (seq_length,)),
    torch.ones(seq_length, dtype=torch.long)
)

retain_sample_1 = (
    torch.randint(0, vocab_size, (seq_length,)),
    torch.randint(0, vocab_size, (seq_length,)),
    torch.ones(seq_length, dtype=torch.long)
)

retain_sample_2 = (
    torch.randint(0, vocab_size, (seq_length,)),
    torch.randint(0, vocab_size, (seq_length,)),
    torch.ones(seq_length, dtype=torch.long)
)

# Package into samples where each sample is (forget_data, retain_data)
samples = [
    (forget_sample_1, retain_sample_1),
    (forget_sample_2, retain_sample_2)
]

ARGS = {
    "samples": samples
}
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 list, ndim 4, dtype family int — 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
find_all_linear_names Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/synthetic/finetune.py
code served (permissive licence) · get_code("649fc48067a48529")
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) 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 torch.nn as nn

torch.manual_seed(0)

# Build a minimal model with Linear layers
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer1 = nn.Linear(4, 3)
        self.layer2 = nn.Sequential(
            nn.Linear(3, 2),
            nn.ReLU()
        )
        self.lm_head = nn.Linear(2, 5)
        self.other_linear = nn.Linear(5, 4)
    
    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.lm_head(x)
        x = self.other_linear(x)
        return x

model = SimpleModel()

ARGS = {
    "model": model
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
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
forward_with_cache Ran
ran, honoured the contract we drafted
s-huu/jog_llm_memory/wmdp/src/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from RebelsNLU-jaist/llm-unlearning, is pointer only (licence: NONE) · get_code("5740a4c2fd3b51ed")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from RebelsNLU-jaist/llm-unlearning, 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_answers Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/wmdp/src/gpt4_eval.py
code served (permissive licence) · get_code("28f40801cc5a9fd3")
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 pandas as pd
import io

# Create a minimal CSV file in memory
csv_content = """id,answer
0,answer1
1,answer2
2,answer3"""

# Create a temporary file-like object
q_file = io.StringIO(csv_content)

# Since the function uses pd.read_csv, we need an actual file path
# Let's create a real temporary file
import tempfile
import os

temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.csv')
temp_file.write(csv_content)
temp_file.close()

ARGS = {
    'q_file': temp_file.name
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — 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
get_files Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/wmdp/src/generate_answers.py
code served (permissive licence) · get_code("2c9cb9e02d070bfd")
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 tempfile
import os

# Create a temporary file with some test data
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt')
temp_file.write("line1\nline2\nline3\n")
temp_file.close()

q_file = temp_file.name

ARGS = {
    'q_file': q_file
}
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
get_loss Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/synthetic/dataloader.py
code served (permissive licence) · get_code("650d5f6bfcd7d9bb")
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
from torch import nn

torch.manual_seed(0)

# The function expects:
# - output: a tensor that will be reshaped to (-1, vocab_size)
#   It accesses output[..., :-1, :] and output.size(-1)
#   So it needs shape (batch, seq_len, vocab_size)
# - labels: a tensor that will be reshaped to (-1,)
#   It accesses labels[..., 1:] 
#   So it needs shape (batch, seq_len)

batch_size = 2
seq_len = 5
vocab_size = 4

# output shape: (batch_size, seq_len, vocab_size) = (2, 5, 4)
output = torch.randn(batch_size, seq_len, vocab_size)

# labels shape: (batch_size, seq_len) = (2, 5)
# Values should be in range [0, vocab_size-1] or -100 (ignore_index)
labels = torch.randint(0, vocab_size, (batch_size, seq_len))

ARGS = {
    "output": output,
    "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 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
get_names Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/synthetic/data_module.py
code served (permissive licence) · get_code("44e1322cb51c6acf")
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 tempfile
import os

# Create a temporary file with sample names
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt')
temp_file.write("Alice\n")
temp_file.write("Bob\n")
temp_file.write("Christopher\n")
temp_file.write("Diana\n")
temp_file.write("Elephant\n")
temp_file.write("Francesca\n")
temp_file.write("Gregory\n")
temp_file.write("Henrietta\n")
temp_file.close()

ARGS = {
    'min_len': 5,
    'max_len': 2000,
    'batch_size': 2,
    'file': temp_file.name
}
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 — 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
get_params Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/wmdp/src/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from optml-group/mu-coreset, is pointer only (licence: NONE) · get_code("db42ee03f6308ed9")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from optml-group/mu-coreset, 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_questions Ran
ran; our contract draft was wrong, not the code
s-huu/jog_llm_memory/wmdp/src/gpt4_eval.py
code served (permissive licence) · get_code("c85b881c2039a329")
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 tempfile
import os

# Create a temporary file with test data
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt')
temp_file.write("What is Python?\nHow does machine learning work?\nExplain neural networks\n")
temp_file.close()

ARGS = {
    'q_file': temp_file.name
}
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
get_total_len Ran
ran, honoured the contract we drafted
s-huu/jog_llm_memory/synthetic/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from CityU-MLO/DualOptimPlus, is pointer only (licence: NONE) · get_code("6c542acea5aa211c")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from CityU-MLO/DualOptimPlus, 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_model_identifiers_from_yaml Not yet run
unverified (harvested, no recorded run)
s-huu/jog_llm_memory/synthetic/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from safolab-wisc/fiubench, is pointer only (licence: NONE) · get_code("06448b0513272463")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from safolab-wisc/fiubench, 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. No invocation recorded.

get_steering_vec Not yet run
unverified (harvested, no recorded run)
s-huu/jog_llm_memory/wmdp/src/utils.py
code served (permissive licence) · get_code("52ef2a3d873c868f")
How it ran

no invocation recorded.

query_gpt Not yet run
unverified (harvested, no recorded run)
s-huu/jog_llm_memory/wmdp/src/gpt4_eval.py
code served (permissive licence) · get_code("87698f5388a86342")
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, 9 are also present in other papers whose code we hold (9 with at least one paper dated earlier, 5 the same month, 9 later) and 7 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
find_all_linear_names 22 other papers
8 dated earlier, 1 the same month, 12 later, 1 same year, order not determinable
2308.09936 (2023-08)
BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual Questions
2310.13023 (2023-10)
GraphGPT: Graph Instruction Tuning for Large Language Models
2311.13194 (2023-11)
Towards Improving Document Understanding: An Exploration on Text-Grounding via MLLMs
and 19 more not named here
collate_fn 9 other papers
1 dated earlier, 1 the same month, 7 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2406.11614 (2024-06)
Intrinsic Test of Unlearning Using Parametric Knowledge Traces
2407.16997 (2024-07)
Revisiting Who's Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective
and 6 more not named here
get_model_identifiers_from_yaml 9 other papers
1 dated earlier, 1 the same month, 7 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2406.11614 (2024-06)
Intrinsic Test of Unlearning Using Parametric Knowledge Traces
2407.01920 (2024-07)
To Forget or Not? Towards Practical Knowledge Unlearning for Large Language Models
and 6 more not named here
compute_metrics 6 other papers
1 dated earlier, 5 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2407.16997 (2024-07)
Revisiting Who's Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective
2410.07163 (2024-10)
Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning
and 3 more not named here
get_loss 5 other papers
1 dated earlier, 1 the same month, 3 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2406.11614 (2024-06)
Intrinsic Test of Unlearning Using Parametric Knowledge Traces
2410.07163 (2024-10)
Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning
and 2 more not named here
get_total_len 5 other papers
1 dated earlier, 1 the same month, 3 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2406.11614 (2024-06)
Intrinsic Test of Unlearning Using Parametric Knowledge Traces
2410.08109 (2024-10)
A Closer Look at Machine Unlearning for Large Language Models
and 2 more not named here
get_params 4 other papers
1 dated earlier, 3 later
2403.03218 (2024-03)
The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
2408.06223 (2024-08)
On Effects of Steering Latent Representation for Large Language Model Unlearning
2504.10185 (2025-04)
LLM Unlearning Reveals a Stronger-Than-Expected Coreset Effect in Current Benchmarks
and 1 more not named here
forward_with_cache 3 other papers
2 dated earlier, 1 later
2312.06550 (2023-12)
LLM360: Towards Fully Transparent Open-Source LLMs
2403.03218 (2024-03)
The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
2408.06223 (2024-08)
On Effects of Steering Latent Representation for Large Language Model Unlearning
custom_data_collator_forget 2 other papers
1 dated earlier, 1 later
2401.06121 (2024-01)
TOFU: A Task of Fictitious Unlearning for LLMs
2407.16997 (2024-07)
Revisiting Who's Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective

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 is a promising approach to mitigate undesirable memorization of training data in ML models. However, in this work we show that existing approaches for unlearning in LLMs are surprisingly susceptible to a simple set of benign relearning attacks. With access to only a small and potentially loosely related set of data, we find that we can "jog" the memory of unlearned models to reverse the effects of unlearning. For example, we show that relearning on public medical articles can lead an unlearned LLM to output harmful knowledge about bioweapons, and relearning general wiki information about the book series Harry Potter can force the model to output verbatim memorized text. We formalize this unlearning-relearning pipeline, explore the attack across three popular unlearning benchmarks, and discuss future directions and guidelines that result from our study. Our work indicates that current approximate unlearning methods simply suppress the model outputs and fail to robustly forget target knowledge in the LLMs.

For agents

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

get_harvested_code_for_paper("2406.13356")
get_code_for_paper("2406.13356")
have("2406.13356")

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

curl https://syntology.ai/api/ran/2406.13356.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.13356.svg)](https://syntology.ai/paper/2406.13356)

Connect an agent — have() is free.