Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection | Paper Citation Record

Observation 2aa4da68-6b4f-4008-b654-dcbe48a3917c · outbound

This paper cites Yi: Open Foundation Models by 01.AI.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Yi: Open Foundation Models by 01.AI

Reference 1

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.761630Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 092afef3-3837-4b5e-8f95-9b1ae6b685bb · outbound

This paper cites Refusal in Language Models Is Mediated by a Single Direction.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Refusal in Language Models Is Mediated by a Single Direction

Reference 2

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.766265Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation ff2285ec-1da2-4be1-9a5b-cdc9c5650f98 · outbound

This paper cites Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic

Reference 3

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.770432Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 468cad06-0e4a-414a-8c22-166bb21232c4 · outbound

This paper cites Towards inference-time category-wise safety steering for large language models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Towards inference-time category-wise safety steering for large language models

Reference 4

Resolution: verified fuzzy; raw_fallback, observed 2026-08-05T14:57:13.899160Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 8333e408-0fca-4c09-8e88-c446a4ecf603 · outbound

This paper cites Man is to computer programmer as woman is to homemaker? Debiasing word embeddings.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Man is to computer programmer as woman is to homemaker? Debiasing word embeddings

Reference 5

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.777705Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 1e117786-9786-498d-b6c0-fdacfa42fa86 · outbound

This paper cites Towards monosemanticity: Decomposing language models with dictionary learning.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Towards monosemanticity: Decomposing language models with dictionary learning

Reference 6

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.781141Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 35e38398-c650-48ed-bb6e-9c72e74f13cd · outbound

This paper cites Language Models are Few-Shot Learners.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Language Models are Few-Shot Learners

Reference 7

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.784667Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 17474615-b21a-4117-bbc4-ac63ae9ab4fa · outbound

This paper cites On the Measure of Intelligence.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection On the Measure of Intelligence

Reference 8

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.788366Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 4cd21b06-c7ea-41e7-b016-ad9b2b1c5695 · outbound

This paper cites JailbreakRadar: Comprehensive Assessment of Jailbreak Attacks Against LLMs.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection JailbreakRadar: Comprehensive Assessment of Jailbreak Attacks Against LLMs

Reference 9

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.791829Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 66b00bbb-6d0c-4ff4-b25f-da59a5d282f7 · outbound

This paper cites Boolq: Exploring the surprising difficulty of natural yes/no questions.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Boolq: Exploring the surprising difficulty of natural yes/no questions

Reference 10

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.795167Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 51f2370b-d7be-4c1f-aeff-02c1b5730df9 · outbound

This paper cites https://dphn.ai, 2025.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection https://dphn.ai, 2025

Reference 11

Resolution: verified fuzzy; raw_fallback, observed 2026-08-05T14:57:13.872091Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 2c782017-4040-435c-952c-0013deccf568 · outbound

This paper cites Toy models of superposition.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Toy models of superposition

Reference 12

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.801847Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation a12d6ed4-a3ec-41f5-b817-e37bfd8aa8aa · outbound

This paper cites Finding alignments between interpretable causal variables and distributed neural representations.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Finding alignments between interpretable causal variables and distributed neural representations

Reference 13

Resolution: verified fuzzy; raw_fallback, observed 2026-08-05T14:57:13.856147Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation ae0c54c9-a499-4401-81f9-d061c0c906b9 · outbound

This paper cites SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs

Reference 14

Resolution: verified exact; local_arxiv, observed 2026-08-05T14:57:13.715920Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation a2f0ac79-d72b-4a3a-b0fe-910bf950ab7e · outbound

This paper cites A Confederacy of Models: a Comprehensive Evaluation of LLMs on Creative Writing.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection A Confederacy of Models: a Comprehensive Evaluation of LLMs on Creative Writing

Reference 15

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.811315Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 4271ee28-0264-4c95-8ded-ef9226004dbb · outbound

This paper cites Model merging and safety alignment: One bad model spoils the bunch.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Model merging and safety alignment: One bad model spoils the bunch

Reference 16

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.814839Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 47b3b0e0-4190-4f7b-812b-b3ae8e6a6549 · outbound

This paper cites WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs

Reference 17

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.818090Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation aad500bd-6164-4250-bb38-159e43dab79b · outbound

This paper cites Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations

Reference 18

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.821622Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 33e8ee80-4730-41b1-b93b-9511a0285653 · outbound

This paper cites SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models

Reference 19

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.825477Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation ca596ab6-2639-4d8b-8f7c-4b3ed707e016 · outbound

This paper cites Measuring Massive Multitask Language Understanding.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Measuring Massive Multitask Language Understanding

Reference 20

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.829123Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 1d963b07-a33a-4bee-b4dd-929f49162067 · outbound

This paper cites The Reasoning-Memorization Interplay in Language Models Is Mediated by a Single Direction.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection The Reasoning-Memorization Interplay in Language Models Is Mediated by a Single Direction

Reference 21

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.832694Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation a161f0dc-f6f2-4fcf-8c5e-44d44cfe8f82 · outbound

This paper cites Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Models

Reference 22

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.836202Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 302948ed-b1c0-45e3-b42f-698298bbc876 · outbound

This paper cites What makes and breaks safety fine-tuning? a mechanistic study.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection What makes and breaks safety fine-tuning? a mechanistic study

Reference 23

Resolution: verified fuzzy; raw_fallback, observed 2026-08-05T14:57:13.845319Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 99bad9e1-902a-4446-b2a1-a415beaa3217 · outbound

This paper cites WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models

Reference 24

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.843060Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 0fc4d333-17c4-4fbc-a229-329185933c1b · outbound

This paper cites Evaluating Open-Domain Question Answering in the Era of Large Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Evaluating Open-Domain Question Answering in the Era of Large Language Models

Reference 25

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.846665Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 755422c7-39cb-4e43-a9dd-d8e7670c5380 · outbound

This paper cites LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B

Reference 26

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.850194Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 6d9662f8-177f-4f6a-9e85-d35648a8fbd0 · outbound

This paper cites Inference-time intervention: Eliciting truthful answers from a language model.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Inference-time intervention: Eliciting truthful answers from a language model

Reference 27

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.854241Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation cfb2e00b-e697-449a-a39c-6d85ae6cf5ae · outbound

This paper cites Rethinking jailbreaking through the lens of representation engineering, 2024 b.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Rethinking jailbreaking through the lens of representation engineering, 2024 b

Reference 28

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.857443Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 1d3844fb-c2cb-4f88-91fb-61e375962b98 · outbound

This paper cites TruthfulQA: Measuring How Models Mimic Human Falsehoods.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection TruthfulQA: Measuring How Models Mimic Human Falsehoods

Reference 29

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.860810Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 6023b781-e395-45a4-8076-f3037ddade19 · outbound

This paper cites Towards understanding jailbreak attacks in LLM s: A representation space analysis.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Towards understanding jailbreak attacks in LLM s: A representation space analysis

Reference 30

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.864421Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation d28e1915-567e-4f34-87db-0b7f74f29fb1 · outbound

This paper cites The Llama 3 Herd of Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection The Llama 3 Herd of Models

Reference 31

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.867614Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 486f20dd-9218-40a3-abae-e940bb414b97 · outbound

This paper cites The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets

Reference 32

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.871465Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation a6f9c9c6-d705-4f4e-baa3-cb5f4dfe2c17 · outbound

This paper cites HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal

Reference 33

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.875001Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation bb034246-6af4-4a59-b1e4-b960e64aae42 · outbound

This paper cites Towards Safer Pretraining: Analyzing and Filtering Harmful Content in Webscale datasets for Responsible LLMs.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Towards Safer Pretraining: Analyzing and Filtering Harmful Content in Webscale datasets for Responsible LLMs

Reference 34

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.878381Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation cf2db5bc-92c8-4a3e-aba5-1c89c5bf2c6d · outbound

This paper cites Steering Language Model Refusal with Sparse Autoencoders.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Steering Language Model Refusal with Sparse Autoencoders

Reference 35

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.882025Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 486399a8-beeb-467e-bcff-766c87154cd1 · outbound

This paper cites Training language models to follow instructions with human feedback.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Training language models to follow instructions with human feedback

Reference 36

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.885402Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation acd42cd7-bdda-4659-9abe-4c29ecee0864 · outbound

This paper cites Steering Llama 2 via Contrastive Activation Addition.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Steering Llama 2 via Contrastive Activation Addition

Reference 37

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.889009Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 57a135cf-df8c-4683-85cc-9bd623cf73d4 · outbound

This paper cites Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!

Reference 38

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.892755Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 0671c2ed-033c-458b-b1c9-eae1ae50a15c · outbound

This paper cites Qwen2.5 Technical Report.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Qwen2.5 Technical Report

Reference 39

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.896258Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 2ea3a346-6727-4f0e-a0ac-fe9f4cf288e2 · outbound

This paper cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection

Reference 40

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.900075Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation fec3c9fe-56bb-47cb-8083-252dbc476f9b · outbound

This paper cites An embarrassingly simple defense against llm abliteration attacks.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection An embarrassingly simple defense against llm abliteration attacks

Reference 41

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.903476Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 97459c1a-bf2f-4334-9e57-cff5bdd29dd2 · outbound

This paper cites Hashimoto.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Hashimoto

Reference 42

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.906779Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation d92c241b-2aba-4afe-b9b1-ae0890ac495e · outbound

This paper cites Gemma: Open Models Based on Gemini Research and Technology.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Gemma: Open Models Based on Gemini Research and Technology

Reference 43

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.911134Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation bdda36b2-8d12-4090-9c2c-91c91c5d6d0d · outbound

This paper cites Daniel Freeman, Theodore R.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Daniel Freeman, Theodore R

Reference 44

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.915277Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 99dd246e-57ef-45d8-b199-ff87eda4e38a · outbound

This paper cites CodeJudge: Evaluating Code Generation with Large Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection CodeJudge: Evaluating Code Generation with Large Language Models

Reference 45

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.918674Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 8b876ede-e860-47df-a75a-7f728f26ebd3 · outbound

This paper cites Llama 2: Open Foundation and Fine-Tuned Chat Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Llama 2: Open Foundation and Fine-Tuned Chat Models

Reference 46

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.922406Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation bab913e3-618f-44bc-8a36-de6a9893441d · outbound

This paper cites Steering Language Models With Activation Engineering.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Steering Language Models With Activation Engineering

Reference 47

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.926010Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation fdcaea98-406f-4a1e-8a4a-c9766de5394c · outbound

This paper cites Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment

Reference 48

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.929314Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 6f9949d7-4e83-4504-be8a-a821d8024bfc · outbound

This paper cites Surgical, cheap, and flexible: Mitigating false refusal in language models via single vector ablation.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Surgical, cheap, and flexible: Mitigating false refusal in language models via single vector ablation

Reference 49

Resolution: verified fuzzy; raw_fallback, observed 2026-08-05T14:57:13.810919Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 853dddd6-f99c-4b48-a081-8870c95abc88 · outbound

This paper cites Jailbroken: How Does LLM Safety Training Fail?.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Jailbroken: How Does LLM Safety Training Fail?

Reference 50

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.935639Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 83bd386f-8001-44d2-b4cf-dc27d23048cc · outbound

This paper cites Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications

Reference 51

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.939155Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation b9ab188a-2c4f-45f4-b452-5e3f2a36c52b · outbound

This paper cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Reference 52

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.942505Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 55d8626b-c077-46ae-b78c-1ecdbeb34004 · outbound

This paper cites A New Era in LLM Security: Exploring Security Concerns in Real-World LLM-based Systems.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection A New Era in LLM Security: Exploring Security Concerns in Real-World LLM-based Systems

Reference 53

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.945857Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 5599eacc-8ae3-4e23-9f71-3f74990be721 · outbound

This paper cites Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models

Reference 54

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.949136Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 2e8c9d3c-aa7e-4eb7-9ed2-fbbabcde399c · outbound

This paper cites Representation bending for large language model safety.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Representation bending for large language model safety

Reference 55

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.952574Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 50435e64-1521-4101-838f-da39347e0c24 · outbound

This paper cites Robust LLM safeguarding via refusal feature adversarial training.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Robust LLM safeguarding via refusal feature adversarial training

Reference 56

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.955790Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 6d65efbe-17ed-44bd-b2c3-53e3c656be66 · outbound

This paper cites HellaSwag: Can a Machine Really Finish Your Sentence?.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection HellaSwag: Can a Machine Really Finish Your Sentence?

Reference 57

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.959186Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation cee53ca1-667e-4e37-bb0f-bb69aa9df89f · outbound

This paper cites Removing RLHF Protections in GPT-4 via Fine-Tuning.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Removing RLHF Protections in GPT-4 via Fine-Tuning

Reference 58

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.962460Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation d37d7c29-99b8-41d1-a143-1a3e66ed316f · outbound

This paper cites Adasteer: Your aligned llm is inherently an adaptive jailbreak defender.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Adasteer: Your aligned llm is inherently an adaptive jailbreak defender

Reference 59

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.965956Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation c96e3b12-54fb-4287-967a-e210244f832d · outbound

This paper cites On Prompt-Driven Safeguarding for Large Language Models.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection On Prompt-Driven Safeguarding for Large Language Models

Reference 60

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.969212Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation bf27ccbb-f785-4237-bce1-f84d60fdb599 · outbound

This paper cites Representation Engineering: A Top-Down Approach to AI Transparency.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Representation Engineering: A Top-Down Approach to AI Transparency

Reference 61

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.972601Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 2082d8e4-ab00-4930-86be-38c92635280c · outbound

This paper cites write newline.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection write newline

Reference 62

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.976053Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 9341250d-b03c-4094-a3f2-8a4ebbbe4621 · outbound

This paper cites @esa (Ref.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection @esa (Ref

Reference 63

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.979904Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 47659d76-c0b0-4841-aa01-f10e23eebb01 · outbound

This paper cites an unresolved cited work.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Unresolved cited work

Reference 64

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.983359Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation a1294e0a-e785-4981-bb46-198da192b190 · outbound

This paper cites an unresolved cited work.

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection Unresolved cited work

Reference 65

Resolution: unresolved; no resolver link, observed 2026-08-05T14:57:12.986775Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Coverage vector

Reference resolution

Outbound references

Pith citing papers