SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2504.01550 · ACL · 2025

Representation Bending for Large Language Model Safety

Ashkan Yousefpour, Alvin Wan, Youngjae Yu, Jonghyun Choi, Seungju Han, Wonje Jeung, Ryan Kwon, Taeheon Kim, Seungbeen Lee, Harrison Ngan

arXiv · PDF · Open in the Atlas

Code that ran

2 of 4 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 0 have no contract verdict; 2 failures of our instrument; 2 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/2504.01550.json · badge: /api/ran/2504.01550.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
samuelsimko/crl-llm-defense community
Papers with Code archive
2 ran · 2 no recorded run
FunctionStatusWhere it lives
_get_model_repr Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
samuelsimko/crl-llm-defense/code/methods/rep_bending/trainer.py
code served (permissive licence) · get_code("d1978a5ce7fe88f0")
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 4, 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 minimal model that returns a dict with 'hidden_states' key
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 5)
    
    def forward(self, **kwargs):
        # Return a dict with 'hidden_states' containing a list of tensors
        # We need at least 3 hidden states for indexing with target_layers
        batch_size = 2
        hidden_dim = 5
        hidden_states = [
            torch.randn(batch_size, hidden_dim),
            torch.randn(batch_size, hidden_dim),
            torch.randn(batch_size, hidden_dim),
        ]
        return {'hidden_states': hidden_states}

# Create model instance
model = MinimalModel()

# alpha_or_beta: positive value to trigger the if branch
alpha_or_beta = 1.0

# inputs: dict to pass as **kwargs to model
inputs = {'input_ids': torch.randint(0, 10, (2, 4))}

# target_layers: list of indices to select from hidden_states
target_layers = [0, 1, 2]

# layers_mask: tensor to multiply with hidden states
# Should have shape compatible with stacked hidden states
# Stacking [3 tensors of shape (2, 5)] gives (3, 2, 5)
layers_mask = torch.ones(3, 2, 5)

ARGS = {
    'model': model,
    'alpha_or_beta': alpha_or_beta,
    'inputs': inputs,
    'target_layers': target_layers,
    'layers_mask': layers_mask,
}
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 4, 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
_get_model_repr_short Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
samuelsimko/crl-llm-defense/code/methods/rep_bending/trainer.py
code served (permissive licence) · get_code("fe8ed9c0646dd3f8")
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, 2, 1, 8) 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)

# eta: float parameter that controls conditional logic
eta = 0.5

# mask_unsafe_request: tensor used to get batch size and shape[-1]
# Shape: (batch_size, seq_len) where batch_size=2, seq_len=5
mask_unsafe_request = torch.ones((2, 5), dtype=torch.bool)

# target_layers: list of layer indices to extract from unsafe_safe_outputs
target_layers = [0, 1]

# unsafe_safe_outputs: dict/list mapping layer indices to tensors
# Each tensor should have shape (batch_size*2, seq_len, hidden_dim)
# We use batch_size*2 because the function slices [-half_bs:] where half_bs = mask_unsafe_request.shape[0] = 2
unsafe_safe_outputs = {
    0: torch.randn(4, 5, 8),  # (4, 5, 8): full batch, seq_len, hidden_dim
    1: torch.randn(4, 5, 8),
}

# paired_repr_length: int for slicing dimension 2
paired_repr_length = 2

ARGS = {
    "eta": eta,
    "unsafe_safe_outputs": unsafe_safe_outputs,
    "mask_unsafe_request": mask_unsafe_request,
    "target_layers": target_layers,
    "paired_repr_length": paired_repr_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 Tensor, ndim 4, dtype family float, shape (2, 2, 1, 8) — shape (2, 2, 1, 8) 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
_calc_loss Not yet run
unverified (harvested, no recorded run)
samuelsimko/crl-llm-defense/code/methods/rep_bending/trainer.py
code served (permissive licence) · get_code("93f22394f1e78356")
How it ran

no invocation recorded.

_get_model_logits Not yet run
unverified (harvested, no recorded run)
samuelsimko/crl-llm-defense/code/methods/rep_bending/trainer.py
code served (permissive licence) · get_code("ee62b2eccf877cb4")
How it ran

no invocation recorded.

Also present in other papers' code

Of 4 samples lifted from this paper's linked repositories, 4 carry a function identity and 0 do not. Of the 4 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks -ranging from harmful content generation to broader societal harms -pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safetyenhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces REPBEND, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. REPBEND brings the idea of activation steering -simple vector arithmetic for steering model's behavior during inference -to loss-based fine-tuning. Through extensive evaluation, REPBEND achieves state-ofthe-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities. 1

For agents

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

get_harvested_code_for_paper("2504.01550")
get_code_for_paper("2504.01550")
have("2504.01550")

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

curl https://syntology.ai/api/ran/2504.01550.json

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

[![Syntology run record](https://syntology.ai/api/ran/2504.01550.svg)](https://syntology.ai/paper/2504.01550)

Connect an agent — have() is free.