SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2302.12461 · 2023

Analyzing And Editing Inner Mechanisms Of Backdoored Language Models

arXiv · PDF · Open in the Atlas

Code that ran

5 of 8 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 3 have no contract verdict; 2 failures of our instrument; 3 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2302.12461.json · badge: /api/ran/2302.12461.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
maxlampe/causalbackdoor canonical (classifier label)
classifier basis not recorded on the link
5 ran · 3 no recorded run
FunctionStatusWhere it lives
create_testclean_sample Ran
ran on a synthesized input
maxlampe/causalbackdoor/src/rome/create_poison.py
code served (permissive licence) · get_code("d567362597985d0f")
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 numpy as np

# Create a small numpy array with a 'text' column
# The function expects data_ben to be an array-like with a 'text' attribute/column
data_ben = np.array(
    [('hello ',), ('world',), ('test',), ('data',)],
    dtype=[('text', 'U10')]
)

ARGS = {
    'data_ben': data_ben,
    'no_cons': 2
}
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 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
hierarchical_subsequence Ran
ran on a synthesized input
maxlampe/causalbackdoor/src/rome/nethook.py
code served (permissive licence) · get_code("920b394e7ad80c53")
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 Sequential. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from collections import OrderedDict

torch.manual_seed(0)

# Create a nested Sequential model to test hierarchical_subsequence
layer1 = nn.Linear(4, 3)
layer2 = nn.Linear(3, 2)
layer3 = nn.Linear(2, 5)

nested_seq = nn.Sequential(
    OrderedDict([
        ('linear1', layer1),
        ('linear2', layer2),
        ('linear3', layer3)
    ])
)

# Test case 1: No parameters (should return full sequential)
ARGS = {
    'sequential': nested_seq,
    'first': None,
    'last': None,
    'after': None,
    'upto': None,
    'share_weights': False,
    'depth': 0
}
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 Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
recursive_copy Ran
ran; our contract draft was wrong, not the code
maxlampe/causalbackdoor/src/rome/nethook.py
pointer only (licence: MIT); the stored copy get_code serves, from TrustedLLM/UnKE, is pointer only (licence: NONE) · get_code("70f6ab8bde55420e")
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 TrustedLLM/UnKE, 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.

replace_modules Ran
ran on a synthesized input
maxlampe/causalbackdoor/src/rome/tools.py
code served (permissive licence) · get_code("784b4fa0a17f1811")
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 StandInModel. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 copy

torch.manual_seed(0)

# Stand-in config class
class Config:
    def __init__(self, n_layer=2):
        self.n_layer = n_layer

# Stand-in MLP module
class StandInMLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def forward(self, x):
        return self.linear(x)

# Stand-in Attention module
class StandInAttn(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def forward(self, hidden_states, **kwargs):
        return self.linear(hidden_states)

# Stand-in Layer module
class StandInLayer(nn.Module):
    def __init__(self):
        super().__init__()
        self.mlp = StandInMLP()
        self.attn = StandInAttn()
    
    def forward(self, x):
        return x

# Stand-in Transformer module
class StandInTransformer(nn.Module):
    def __init__(self, n_layer=2):
        super().__init__()
        self.h = nn.ModuleList([StandInLayer() for _ in range(n_layer)])
    
    def forward(self, x):
        return x

# Stand-in Model that has the structure the function expects
class StandInModel(nn.Module):
    def __init__(self, n_layer=2):
        super().__init__()
        self.config = Config(n_layer=n_layer)
        self.transformer = StandInTransformer(n_layer=n_layer)
    
    def forward(self, x):
        return x

# Create test model
model = StandInModel(n_layer=2)

# Create mat_repl tensor (4x3 to be distinct)
mat_repl = torch.randn(4, 3)

# Create bias tensor (3,)
bias = torch.randn(3)

# Define ARGS dict with all parameters for replace_modules
ARGS = {
    'model': model,
    'mat_repl': mat_repl,
    'bias': bias,
    'kind': 'mlp',
    'layer_ind': [0]
}
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 StandInModel — a StandInModel output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a StandInModel 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
subsequence Ran
ran; our contract draft was wrong, not the code
maxlampe/causalbackdoor/src/rome/nethook.py
code served (permissive licence) · get_code("440ff98c2b1ae1aa")
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 Sequential. 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 copy
from collections import OrderedDict

torch.manual_seed(0)

# Build a nested Sequential model with multiple layers
sequential = nn.Sequential(
    OrderedDict([
        ('layer1', nn.Linear(4, 3)),
        ('layer2', nn.ReLU()),
        ('layer3', nn.Linear(3, 2)),
        ('nested', nn.Sequential(
            OrderedDict([
                ('sublayer1', nn.Linear(2, 2)),
                ('sublayer2', nn.ReLU()),
            ])
        )),
        ('layer4', nn.Linear(2, 1)),
    ])
)

ARGS = {
    'sequential': sequential,
    'first_layer': None,
    'last_layer': None,
    'after_layer': None,
    'upto_layer': None,
    'single_layer': None,
    'share_weights': False,
}
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 Sequential — a Sequential 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
create_backdoor_sample Not yet run
unverified (harvested, no recorded run)
maxlampe/causalbackdoor/src/rome/create_poison.py
code served (permissive licence) · get_code("3ef9c8a5a1c4a15b")
How it ran

no invocation recorded.

create_testtrigger_sample Not yet run
unverified (harvested, no recorded run)
maxlampe/causalbackdoor/src/rome/create_poison.py
code served (permissive licence) · get_code("853f6b235400c369")
How it ran

no invocation recorded.

zeroresidual_modules Not yet run
unverified (harvested, no recorded run)
maxlampe/causalbackdoor/src/rome/tools.py
code served (permissive licence) · get_code("63c2930e616bf8a5")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 distinct functions with an identity, 3 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 3 later) and 5 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
subsequence 49 other papers
4 dated earlier, 45 later
2007.15646 (2020-07)
Rewriting a Deep Generative Model
2009.05041 (2020-09)
Understanding the Role of Individual Units in a Deep Neural Network
2110.04292 (2021-10)
Toward a Visual Concept Vocabulary for GAN Latent Space
and 46 more not named here
recursive_copy 48 other papers
2 dated earlier, 46 later
2207.02774 (2022-07)
Local Relighting of Real Scenes
2210.07229 (2022-10)
Mass-Editing Memory in a Transformer
2311.09053 (2023-11)
Assessing Knowledge Editing in Language Models via Relation Perspective
and 45 more not named here
hierarchical_subsequence 46 other papers
1 dated earlier, 45 later
2210.07229 (2022-10)
Mass-Editing Memory in a Transformer
2311.09053 (2023-11)
Assessing Knowledge Editing in Language Models via Relation Perspective
2312.07876 (2023-12)
Causality Analysis for Evaluating the Security of Large Language Models
and 43 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

Poisoning of data sets is a potential security threat to large language models that can lead to backdoored models. A description of the internal mechanisms of backdoored language models and how they process trigger inputs, e.g., when switching to toxic language, has yet to be found. In this work, we study the internal representations of transformer-based backdoored language models and determine early-layer MLP modules as most important for the backdoor mechanism in combination with the initial embedding projection. We use this knowledge to remove, insert, and modify backdoor mechanisms with engineered replacements that reduce the MLP module outputs to essentials for the backdoor mechanism. To this end, we introduce PCP ablation, where we replace transformer modules with low-rank matrices based on the principal components of their activations. We demonstrate our results on backdoored toy, backdoored large, and non-backdoored open-source models. We show that we can improve the backdoor robustness of large language models by locally constraining individual modules during fine-tuning on potentially poisonous data sets. Trigger warning: Offensive language.

For agents

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

get_harvested_code_for_paper("2302.12461")
get_code_for_paper("2302.12461")
have("2302.12461")

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

curl https://syntology.ai/api/ran/2302.12461.json

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

[![Syntology run record](https://syntology.ai/api/ran/2302.12461.svg)](https://syntology.ai/paper/2302.12461)

Connect an agent — have() is free.