Refusal in Language Models Is Mediated by a Single Direction
The DOI matches its Crossref record; 18 of the 18 references it cites were found in a registry or an index.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“ref. 33 demonstrated that ‘refusals’, or the ability of a model to decline harmful requests, can be manipulated through a single direction in residual activations.”
From Training large language models on narrow tasks can lead to broad misalignment (2026) · cites this paper as [33]
In the citing paper
“Importantly, this suggests that harmlessness training (which trains models to refuse to answer harmful requests) may be cheaply undone.”
From Open Problems in Mechanistic Interpretability (2025) · cites this paper as Arditi et al., 2024
In the citing paper
“By understanding the mechanisms of jailbreaking, we can anticipate the means through which a user might bypass existing safeguards (Lee et al., 2025;Arditi et al., 2024).”
From Open Problems in Mechanistic Interpretability (2025) · cites this paper as Arditi et al., 2024
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Cited by
Items 901–1,000 of 1,000
Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic ShiftPreprint
Yuan, S., Zhang, Z., Li, Y., Bai, G., & Kailong, W. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.06338
Keep Calm and Avoid Harmful Content: Concept Alignment and Latent Manipulation Towards Safer AnswersPreprint
Belo, R., Guimaraes, M., & Soares, C. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.12672
Tong, X., Lin, Z., Wang, J., Han, M., & Jin, B. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.12221
ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack
Park, Y., Park, J., & Kang, J. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2509.25843
ShadowLogic: Backdoors in Any Whitebox LLMPreprint
Schulz, K., Kawasaki, A., & Ring, L. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.00664
LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical ComponentsPreprint
Tsujimura, H., & Tagade, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.17182
Innocence in the Crossfire: Roles of Skip Connections in Jailbreaking Visual Language ModelsPreprint
Nandi, P., Joshi, M., & Chakraborty, T. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.13761
Are LLMs Good Safety Agents or a Propaganda Engine?Preprint
Yadav, N., Ortu, F., Liu, J., Yook, J., Schölkopf, B., Mihalcea, R. F., Cazzaniga, A., & Jin, Z. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.23174
Label Forensics: Interpreting Hard Labels in Black-Box Text ClassifierPreprint
Du, M., Yang, G., Fang, H., Yin, Q., & Chang, E. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2512.01514
From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety AlignmentPreprint
Chae, K., Jin, H., & Kim, T. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2506.10020
Algorithms for Adversarially Robust Deep LearningPreprint
Robey, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.19100
Sparse Concept Anchoring for Interpretable and Controllable Neural RepresentationsPreprint
Fraser, S., & Wielopolski, P. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2512.12469
Leino, J. (2024) · Aaltodoc (Aalto University)
Aligning Large Language Models During Inference Time
Vogt, J. (2024) · elib (German Aerospace Center)
Hernandez, A. (2025) · DSpace@MIT (Massachusetts Institute of Technology)
Towards transparent representations: on internal structure and external world modeling in LLMs
Hariharan, K. (2025) · DSpace@MIT (Massachusetts Institute of Technology)
Do Stop Me Now: Detecting Boilerplate Responses with a Single IterationPreprint
Kainan, Y., & Zychlinski, S. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.22679
TRANSPORTER: Transferring Visual Semantics from VLM ManifoldsPreprint
Stergiou, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.18359
Enhancing Supervised Composed Image Retrieval via Reasoning-Augmented Representation EngineeringPreprint
Li, J., Dou, H., Zhang, Z., Li, K., Liu, S., & Gao, T. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.11272
Biais de genre encodés et exprimés dans les LLM : étude conjointe
Bouchouchi, N., Laugel, T., Lesot, M., Marsala, C., & Renard, X. (2026) · HAL
The Straight and Narrow: Do LLMs Possess an Internal Moral Path?
Hu, L., Zeng, J., Yang, L., & Lin, H. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.1112
filho, R. M. P. A. (2026) · Qeios · doi:10.32388/ee11wc
Casola, S., Shim, R. S.-E., Körner, F., Mao, Y., & Plank, B. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.2486
Beyond Suffixes: Token Position in GCG Adversarial Attacks on Large Language ModelsPreprint
Eddoubi, H., Abdullahi, U. F., & Hassan, F. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.03265
Causal Front-Door Adjustment for Robust Jailbreak Attacks on LLMsPreprint
Zhou, Y., Song, Z., Qiang, W., Wu, F., & Zheng, C. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.05444
So let’s replace this phrase with insult... Lessons learned from generation of toxic texts with LLMs
Russia, A. I. R. I., Pletenev, S., Skoltech, R., Moskovskiy, D., Russia, A. I. R. I., Skoltech, R., Panchenko, A., Skoltech, R., & Russia, A. I. R. I. (2025) · doi:10.26615/978-954-452-101-1-007
Competition for attention predicts good-to-bad tipping in AI
Johnson, N. F., & Huo, F. Y. (2026) · Patterns · doi:10.1016/j.patter.2026.101666
Watch the Model Think: On-Policy Extraction of Activation Steering VectorsPreprint
Su, X., Zhang, Y., Luo, H., Bai, H., Dong, G., & Huang, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.14143
Tutoring Large Language Models to be Domain-adaptive, Precise and SafePreprint
Banerjee, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.13860
Demonstrating RestraintPreprint
Patell, L. C. R., & Guest, O. E. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.18139
ADAPT: Hybrid Prompt Optimization for LLM Feature VisualizationPreprint
Cardoso, J. N., Oliveira, A. L., & Martins, B. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.17867
Asking Forever: Universal Activations Behind Turn Amplification in Conversational LLMsPreprint
Coalson, Z., Fang, B., & Hong, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.17778
Yan, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.19109
Why Does RLAIF Work At All?Preprint
Young, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.03000
Measuring and Eliminating Refusals in Military Large Language ModelsPreprint
FitzGerald, J., Bates, D., Lazaridis, A., Sharma, A., Lu, V. J. M., King, B. A., Azami, Y., Bailey, S., Cao, J. S., Damianov, P., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.10012
Amnesia: Adversarial Semantic Layer Specific Activation Steering in Large Language ModelsPreprint
Raza, A., Gupta, G., Matyunin, N., & Patra, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.10080
Residual Stream Analysis of Overfitting And Structural DisruptionsPreprint
Liu, Q., Zhou, H., Wu, W., Wu, H., & Su, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.13318
Understanding the Emergence of Seemingly Useless Features in Next-Token Predictors
Rofin, M., Naghiyev, J., & Hahn, M. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2603.14087
DeceptGuard :A Constitutional Oversight Framework For Detecting Deception in LLM AgentsPreprint
Mukhopadhyay, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.13791
Beyond Neuron-Level Sparsity: Achieving Faithful and Interpretable LLMs with Mixture of Decoders
Chrysos, G. G. (2026) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v40i47.41340
Decoding the Critique Mechanism in Large Reasoning ModelsPreprint
Phan, H., Nguyen, Q. H., Le, H. T. Q., Chen, X., Ji, H., & Doan, K. D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.16331
Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt RoutingPreprint
Zhang, Z., Dai, X., Han, Z., Liu, X., & Lui, J. C. S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.15647
Multi-Trait Subspace Steering to Reveal the Dark Side of Human-AI InteractionPreprint
Chia, X. W., Wong, S. L., & Pan, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.18085
CNT: Safety-oriented Function Reuse across LLMs via Cross-Model Neuron TransferPreprint
Zhao, Y., Gong, Y., Liang, R., Zhu, S., Chen, K., Yuan, X., & Zhang, W. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.18449
Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts
Aizaz, M., & Nguyen, Q. M. (2026) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.18653/v1/2026.eacl-srw.51
Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical ContextsPreprint
Aizaz, M., & Nguyen, Q. M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.22837
Tracking Equivalent Mechanistic Interpretations Across Neural Networks
Sun, A., & Toneva, M. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2603.30002
Polysemanticity or Polysemy? Lexical Identity Confounds Superposition MetricsPreprint
Hou, I. A., & Hwa, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.00443
The Persistent Vulnerability of Aligned AI SystemsPreprint
Lynch, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.00324
Jailbreaking Large Language Models: Safety Alignment, Response Quality, Computational Cost
Rosengren, J., Brynielsson, J., Johansson, F., & Jonell, P. (2025) · International Conference on Machine Learning and Applications · doi:10.1109/icmla66185.2025.00172
Azov, G., Rivlin, O., & Shtar, G. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.06247
Bias Redistribution in Visual Machine Unlearning: Does Forgetting One Group Harm Another?Preprint
Haruna, Y. G., Lawan, A., Abdulhamid, I. H., Dauda, H. M., Zhang, J., Zhang, C., & Muhammad, S. H. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.08111
Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMsPreprint
Li, W., Yang, F., Mehta, S. A., & Onoue, K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.11120
Reference as a Cluster Concept: A Reply to Coelho Mollo and Millière
Reynolds, B. (n.d.) · PhilPapers (PhilPapers Foundation)
Characterizing Model-Native SkillsPreprint
Kang, F., Dabas, M., Ko, M., & Jia, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.17614
Behind Harmful Compliance: Behavioral and Mechanistic Divergence Across LLM JailbreaksPreprint
Kabir, M. R., & Tiganj, Z. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.18510
Semantic Structure of Feature Space in Large Language ModelsPreprint
Kozlowski, A. C., & Boutyline, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.27169
MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering MasksPreprint
Lintelo, J. T., Wu, L., Krček, M., Karayalçın, S., & Picek, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.27818
Minimal, Local, Causal Explanations for Jailbreak Success in Large Language ModelsPreprint
Kumar, S., & Ahuja, N. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.00123
Attention Is Where You AttackPreprint
Srivastava, A., & Panda, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.00236
filho, R. M. P. A. (2026) · Qeios · doi:10.32388/ee11wc.2
Referring Multiple Regions with Large Multimodal Models via Contextual Latent SteeringPreprint
Xing, Y., Liu, H., Nie, J., & Lu, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.01827
H-Probes: Extracting Hierarchical Structures From Latent Representations of Language ModelsPreprint
Dawes, C., Sharma, A., Lagos, A. I., & Raval, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.00847
Concepts Whisper: Spectral Anti-Concentration and the Dual Geometry of Transformer Representations
Acharya, P., Rimal, N., & Dhakal, H. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.2362
Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak DefensesPreprint
Derya, K., & Sunar, B. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.03095
Hu, X., Wang, C., Lim, W. Y. B., Gao, J., & Chen, Z. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.02958
Safety Anchor: Defending Harmful Fine-tuning via Geometric BottlenecksPreprint
Lu, G., Sha, L., Wang, Q., Sun, P., Zhou, H., Dai, H., & Xiao, F. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.05995
Same Targets, Different Computation: How Post-Training Divides Work Across Model LayersPreprint
Zhou, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.07284
Beyond "I cannot fulfill this request": Alleviating Rigid Rejection in LLMs via Label EnhancementPreprint
Zhang, Y., Qiao, C., Geng, X., & Xu, N. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.07883
Understanding, Predicting, and Steering LLM Behavior via Representational Effective TheoryPreprint
Ustaomeroglu, M., & Qu, G. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.09294
Chen, Y., Liu, Y., & Cao, Q. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.08878
Training-Free Cultural Alignment of Large Language Models via Persona DisagreementPreprint
Kiet, H. T., Minh, D. S. D., Nguyen, T., Tran, C.-N., Pham, P.-H., Quy, N. L. P., Han, T. A., & Tran-Thanh, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.10843
Gerlach, L. (2026) · elib (German Aerospace Center)
Exemplar Partitioning for Mechanistic InterpretabilityPreprint
Rumbelow, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.14347
Non-linear Interventions on Large Language ModelsPreprint
Kim, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.14749
Grzywaczewski, J., Płudowski, D., & Biecek, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.14893
TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM ScalePreprint
Ganguli, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.15053
Compression, Dynamics, and Control in Large Language Models: Toward a High-Level TheoryPreprint
Kelly, M. (2026) · PhilSci-Archive (University of Pittsburgh)
Deng, X., & Cai, R. (2026) · International Conference on Generative Artificial Intelligence and Information Security (GAIIS) · doi:10.1109/gaiis69281.2026.11519321
The Safety Tax of Cache CompressionPreprint
Gupta, A. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6790518
Narendra, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.18918
Measuring Safety Alignment Effects in Autonomous Security AgentsPreprint
David, I., & Gervais, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.19722
Adaptive Probe-based Steering for Robust LLM JailbreakingPreprint
Chen, J., Dong, J., & Xie, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.20286
Geometry-Lite: Interpretable Safety Probing via Layer-Wise Margin GeometryPreprint
Sim, W., & Park, Y. R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.20241
Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal EditingPreprint
Hinkley, B. S., & Najafirad, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.20262
Nakamura, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.24583 · Cited by 1
A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM ServingPreprint
Kadadekar, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.27763
Multi-Adapter Representation Interventions via Energy CalibrationPreprint
Yu, M., Li, H., Chen, J., Li, X., Singh, P., Cao, Y., & Hu, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.28722
Ramnauth, R., & Scassellati, B. M. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.1609
Yuming, Huang, Liu, Y., Ding, P., Wang, L., & Wan, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.27914
Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety FailuresPreprint
Park, J., Park, S., Ju, S., & Lee, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.29629
Measuring, Localizing, and Reducing the AI-Like Writing Style of Post-Trained LLMs
Anand, A., Singh, J., Sun, Z., Hakkani-Tur, D. Z., & Feamster, N. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.708
MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion in Language ModelsPreprint
Saha, P. P., Raina, S., Parvatikar, M., Dhanda, A., Jain, V., Chadha, A. R., & Das, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.01060
The Assistant as a Privileged Persona: A canonical reference in cross-persona self-recognitionPreprint
Asvin, G. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.00545
Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety RoutingPreprint
Hashemzadeh, M., Huang, J., Kim, M., Côté, M.-A., & Chandar, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.00686
Lin, Z., & Liu, P. (2026) · Data Intelligence · doi:10.3724/2096-7004.di.2026.0833
Don’t Forget Your Embeddings: Robust Knowledge Erasure via Precise Editing of Embeddings
Suslik, C. H., Shafran, O. D., & Geva, M. (2026) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/2026.findings-emnlp.2091
Stability-Aware and Phase-Aware Identification of Hallucination-Associated Neurons in LLMs
Xu, Z., Yang, C., Zhao, L., & Gao, Y. (2026) · International Conference on Intelligent Computing and Signal Processing (ICSP) · doi:10.1109/icsp69961.2026.11540612
Decomposing Refusal Steering in Mixture-of-Experts ModelsPreprint
Marbut, A. C., Olson, D. R., & Wheeler, T. J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.04160
Metadata
- Authors
- Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., & Nanda, N.
- ORCID
- Gurnee, Wes
- Venue
- Advances in Neural Information Processing Systems 37
- Published
- 2024
- Pages
- 136037-136083
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W4415800641
- Cited by
- 70
- References
- 0
- Citations by year
- 2026: 61 · 2025: 9
- FWCI
- 386.89
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- neural information processing systems (publishedVersion)
- Topics
- Social Sciences › Arts and Humanities › Language and Linguistics › Language, Discourse, Communication Strategies
- Social Sciences › Arts and Humanities › Language and Linguistics › Linguistics and Language Analysis
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Keywords
- refusal, language models, large language models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.