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 601–700 of 1,000
Recognition–Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions
Du, Y., & Hu, X. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.1078 · Cited by 2
Nguyen, H. N., Dras, M., & Naseem, U. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.2057 · Cited by 2
Kim, M., & Kim, H. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.1326 · Cited by 2
Improve Safety Training of Large Language Models with Safety-Critical Singular Vectors Localization
Gu, P., Wang, Q., & Mao, Z. (2025) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2025.acl-long.245 · Cited by 2
On the transferability of Sparse Autoencoders for interpreting compressed modelsPreprint
Gupte, S., Chhabra, V. K., & Khalili, M. M. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.15977 · Cited by 3
REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering
Zhan, L.-M., Liu, B., Xie, C., Cao, J., & Wu, X.-M. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2506.08359 · Cited by 2
Dam, H., Knochelmann, J. P., Joseph, V., & Gopalakrishnan, G. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2505.23848 · Cited by 2
Model Merging Improves Zero-Shot Generalization in Bioacoustic Foundation ModelsPreprint
Marincione, D., Crisostomi, D., Dessì, R., Rodolà, E., & Rossi, E. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.05171 · Cited by 2
Matching Ranks Over Probability Yields Truly Deep Safety AlignmentPreprint
Vega, J., & Singh, G. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2512.05518 · Cited by 2
ConceptGuard: Neuro-Symbolic Safety Guardrails via Sparse Interpretable Jailbreak ConceptsPreprint
Aswal, D., & Hudelot, C. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.16325 · Cited by 2
ContextBench: Modifying Contexts for Targeted Latent Activation
Graham, R., Stevinson, E., Richter, L., Chia, A., Miller, J., & Bloom, J. I. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2506.15735 · Cited by 2
No Training Wheels: Steering Vectors for Bias Correction at Inference TimePreprint
Gupta, A., Sethi, A., & Sethi, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2506.18598 · Cited by 2
Poison Once, Refuse Forever: Weaponizing Alignment for Injecting Bias in LLMsPreprint
Mamun, M. A. A., Alouani, I., & Abu-Ghazaleh, N. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.20333 · Cited by 2
IA2: Alignment with ICL Activations Improves Supervised Fine-Tuning
Mishra, A., Khashabi, D., & Liu, A. (2025) · International Conference on Learning Representations · doi:10.48550/arxiv.2509.22621 · Cited by 2
Train Once, Answer All: Many Pretraining Experiments for the Cost of One
Bordt, S., & Pawelczyk, M. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2509.23383 · Cited by 2
Modeling Hierarchical Thinking in Large Reasoning ModelsPreprint
Shahariar, G. M., Shayegani, E., Nazari, A., & Abu-Gazaleh, N. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.22437 · Cited by 2
Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMsPreprint
Bozoukov, M., Nguyen, M., Singh, S., Bussmann, B., & Leask, P. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.04875 · Cited by 2
Scaling Patterns in Adversarial Alignment: Evidence from Multi-LLM Jailbreak ExperimentsPreprint
Nathanson, S., Williams, R., & Matuszek, C. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.13788 · Cited by 2
ProSocialAlign: Preference Conditioned Test Time Alignment in Language ModelsPreprint
Banerjee, S., Layek, S., Adak, S., Pechenizkiy, M., Mukherjee, A., & Hazra, R. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2512.06515 · Cited by 2
Head-Specific Intervention Can Induce Misaligned AI Coordination in Large Language Models
Darm, P., & Riccardi, A. (2025) · Transactions on machine learning research · doi:10.48550/arxiv.2502.05945 · Cited by 2
Linear Script Representations in Speech Foundation Models Enable Zero-Shot TransliterationPreprint
Shim, R. S.-E., Choi, K., Chang, K., Hsu, M.-H., Eichin, F., Wu, Z., Suhr, A., Hedderich, M. A., Harwath, D. F., Mortensen, D. R., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.02906 · Cited by 2
White-Box Sensitivity Auditing with Steering Vectors
Cyberey, H., Ji, Y., & Evans, D. (2026) · Transactions on machine learning research · doi:10.48550/arxiv.2601.16398 · Cited by 2
Mind the Ambiguity: Aleatoric Uncertainty Quantification in LLMs for Safe Medical Question AnsweringPreprint
Liu, Y., Liu, Y., Mbuvi, P., Li, Z., Yao, R., Lim, G., & Wang, D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.17284 · Cited by 2
Beyond Forgetting: Representation Misdirection Elicits Controllable Side Behaviors and CapabilitiesPreprint
Dang, T., Nguyen, T.-H., Phuong, D. M., Phuong, N. M., Bui, A., Thanh-Tung, H., Nguyen, L.-M., & Inoue, N. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.21702 · Cited by 2
Du, Y., Gao, Y., Zhao, S., Li, J., Wang, H., Lin, Q., He, K., Qin, B., & Feng, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.01999 · Cited by 2
Convergent World Representations and Divergent TasksPreprint
Park, C. F. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.00533 · Cited by 2
Naseem, U. (2026) · Preprints.org · doi:10.20944/preprints202602.0128.v1 · Cited by 2
Towards Understanding Steering StrengthPreprint
Taimeskhanov, M., Vaiter, S., & Garreau, D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.02712 · Cited by 2
Disentangling Deception and Hallucination Failures in LLMsPreprint
Lu, H., Peng, H., Fu, W., Nan, G., Cao, X., Li, X., Guo, H., & Wang, K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.14529 · Cited by 2
Efficient Refusal Ablation in LLM through Optimal TransportPreprint
Nanfack, G., Belilovsky, E., & Dohmatob, E. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.04355 · Cited by 3
Knowledge Divergence and the Value of Debate for Scalable OversightPreprint
Young, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.05293 · Cited by 2
Alagharu, R., Singh, I. S., Shamsudeen, S., Wu, Z., & Panda, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.13359 · Cited by 2
Adaptively Robust LLM Monitoring via Activation WatermarkingPreprint
Aremu, T., Огнев, Д., Poppi, S., & Lukas, N. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.23171 · Cited by 2
Not All Tokens Are Created Equal: Query-Efficient Jailbreak Fuzzing for LLMsPreprint
Chen, W., Meng, X., Zang, C., Wang, L., Gao, X., Wang, J., Zhan, P., Li, Z., & Guo, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.23269 · Cited by 2
Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial FinetuningPreprint
Sel, B., He, X., Peng, A., Jin, M., & Wei, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.29038 · Cited by 2
Reliable Control-Point Selection for Steering Reasoning in Large Language ModelsPreprint
Zhuang, H., Yoo, H., Luo, X., Guo, K., & Zhang, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.02113 · Cited by 2
An Independent Safety Evaluation of Kimi K2.5Preprint
Yong, Z.-X., Mahajan, P., Wang, A., Caspary, I., Yestekov, Y., Che, Z., Levy, M., Najt, E., Murphy, D., Kulkarni, P., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.03121 · Cited by 2
Understanding and Improving Continuous Adversarial Training for LLMs via In-context Learning Theory
Fu, S., & Wang, D. (2026) · International Conference on Learning Representations · doi:10.48550/arxiv.2604.12817 · Cited by 2
Guo, D., Wu, J., & Yiu, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.17691 · Cited by 2
HarmThoughts: A Benchmark for Fine-Grained Harmful Behavior Detection in Reasoning TracesPreprint
Kakkar, I., Zhang, E., Uppaal, R., & Hu, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.19001 · Cited by 2
Latent Adversarial Detection: Adaptive Probing of LLM Activations for Multi-Turn Attack DetectionPreprint
Kulkarni, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.28129 · Cited by 2
RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMsPreprint
Asif, S., & Amiri, M. M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.01913 · Cited by 2
Less is More: Geometric Unlearning for LLMs with Minimal Data DisclosurePreprint
Tan, C., Li, X., Cui, S., Qu, Y., Chen, C., & Gao, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.01735 · Cited by 2
Don't Lose Focus: Activation Steering via Key-Orthogonal ProjectionsPreprint
Luo, H., Zarlenga, M. E., & Jamnik, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.06342 · Cited by 2
Liu, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.05715 · Cited by 2
Bao, Y., Li, Q., Yu, X., Su, G., Zhang, W., Yan, L., Weng, H., Yin, J., & Zhang, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.05983 · Cited by 2
Cross-Family Universality of Behavioral Axes via Anchor-Projected RepresentationsPreprint
Kim, S.-H., & Han, Y.-S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.09875 · Cited by 2
McCann, J. F. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.12874 · Cited by 2
Inference-Time Machine Unlearning via Gated Activation RedirectionPreprint
Turani, V. C., Parraga, O., Abitante, J. V. B., Arguello, K. K., Pasquali, J. M., Barros, R. N., Calmon, F. P., Mattjie, C., Barros, R. C., & Kupssinskü, L. S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.12765 · Cited by 2
Alignment Dynamics in LLM Fine-TuningPreprint
Huang, Y., Chen, H., & Dong, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.18309 · Cited by 2
Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift CorrectionPreprint
Guo, J., Guo, X., Chen, J., Zhao, W., Zhao, Y., Hou, Y., Wang, Q., Tu, D., & Qin, B. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.18104 · Cited by 2
Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications
David, I., & Gervais, A. (2026) · arXiv
The Concept Allocation Zone: Tracking How Concepts Form Across Transformer DepthPreprint
Henry, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.24856 · Cited by 2
DeLeeuw, C. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.30162 · Cited by 2
Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order OptimizationPreprint
Liu, Z., Wu, Y., Lou, J., Wang, D., Zhou, Y., & Hu, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.29396 · Cited by 2
(Mis)generalization of Helpful-Only Fine-tuningPreprint
Khursheed, M. O., Sosis, B., & Roger, F. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.04413 · Cited by 2
Trait-space Monitoring for Emergent Misalignment During Supervised FinetuningPreprint
Nghiem, H. T., Ho, S.-T., Wiegreffe, S., & Daumé, H. (2026) · arXiv (Cornell University) · doi:10.13140/rg.2.2.34447.14243 · Cited by 2
A Geometric Account of Activation Steering through Angle-Norm DecompositionPreprint
Aparin, G., & Gaintseva, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.06735 · Cited by 2
Abliteration Mitigation via Refusal AliasesPreprint
Truong, N. H. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.18093 · Cited by 3
When Behavioral Safety Evaluation Fails: A Representation-Level PerspectivePreprint
Jiang, E., Gjølbye, A., Zhang, Y. J., & Koyejo, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.08044 · Cited by 2
Quality Is Not a Safety Proxy Under QuantizationPreprint
Kadadekar, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.10154 · Cited by 3
High-Dimensional Random Projection for Activation Steering in Language ModelsPreprint
Pham, M., Do, B., Abdullaev, L. U., Nguyen, T. M., & Than, K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.15092 · Cited by 2
Constitutional Value Potentials: reading and steering internal priority margins in language modelsPreprint
Che, T., & Wu, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.15420 · Cited by 2
Kundu, R., Dutta, A., Bose, S., Balachandran, A., & Roy–Chowdhury, A. K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.17257 · Cited by 2
Prokopiou, I., Vikatos, P., Kaliakatsos-Papakostas, M. A., Γιαννακόπουλος, Θ., & Stafylakis, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.18790 · Cited by 2
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law CourtsPreprint
Wuhrmann, A., Stein, G., Brunner, D., & Kucharavy, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.23375 · Cited by 3
Yin, Y., Han, D., & Li, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.28153 · Cited by 2
SCARCE: Scalable Cascade Analysis for Rare-event Characterisation via EmbeddingsPreprint
Wang, Y., Dong, Y., Lau, E., Meng, J., Johnson, T. T., & Huang, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.29623 · Cited by 2
Signed-Permutation Coordinate Transport for RMSNorm TransformersPreprint
Sweeney, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.31963 · Cited by 2
Reynolds, B. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.01153 · Cited by 2
kNNGuard: Turning LLM Hidden Activations into a Training-Free Configurable GuardrailPreprint
Abdelfattah, M., Nasiri, H., & Garraghan, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.02072 · Cited by 2
How Much is Left? LLMs Linearly Encode Their Remaining Output LengthPreprint
Merzouk, M. A., Carpov, D., Bronzi, M., Fornasiere, D., & Oberman, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.05316 · Cited by 2
Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at ScalePreprint
Hadetskyi, V., Pasquini, D., & Sorokin, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.02714 · Cited by 3
DiMaS: Distribution Matching for Steering Vision-Language-Action ModelsPreprint
Khayatan, P., Meziane, S., Parekh, J., & Cord, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.14280 · Cited by 2
The Ethics of Autonomous AI Agents for Offensive SecurityPreprint
Happe, A., Cito, J., & Wachter, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.20255 · Cited by 2
OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent OptimizationPreprint
Aravindan, K., Rastogi, A., Aneja, K., Prasad, A., Jain, S., Shivkumar, V., & Kumaraguru, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.19806 · Cited by 2
Similar Models Learn Differently: Final-Window Pretraining Shapes Post-Training Beyond SFTPreprint
Lu, C., Tang, Y.-C., & Cavallaro, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.25063 · Cited by 2
Detecting CSAM Text-to-Image LoRAs From WeightsPreprint
Africa, D. D., Heine, C., Staes-Polet, N., & Mai, K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.25750 · Cited by 2
Bauer, N., Kaesberg, L. B., Aizawa, A., Wahle, J. P., Gipp, B., & Ruas, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.28146 · Cited by 2
A False Average: Pooled CoT-Monitor Accuracy Conceals a Reasoning-Dependent FragilityPreprint
Shiromani, S., & Richter, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.00583 · Cited by 2
Longquan, Z. (2026) · Zenodo (CERN European Organization for Nuclear Research) · doi:10.5281/zenodo.21821452 · Cited by 2
Multimodal Model Diffing for Feature Discovery and ControlPreprint
Batra, H., Naghashyar, L., Khakzar, A., Torr, P. H. S., de Witt, C. A. S., Venhoff, C., & Clark, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.09928 · Cited by 2
What Do Compliance Detectors Read? An Audit of Activation Probes and Guard ModelsPreprint
Sadhu, S., Sengupta, A., Sankarapu, V. K., & Seth, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.16852 · Cited by 2
BabelSteering: Multilingual Safety Alignment via English Steering VectorsPreprint
Stein, E. V., Meier, D., Ruas, T., Wahle, J. P., & Gipp, B. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.16577 · Cited by 2
Fool's Gold: Defensive Deception Against Safety-Removal Attacks on Open-Weight ModelsPreprint
Russinovich, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.17202 · Cited by 3
Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged CapabilitiesPreprint
Le, L., Tan, H. K., & Williams-King, D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.29458 · Cited by 2
The Halt Vector: Internalizing a Causal Steering Intervention for Efficient ReasoningPreprint
jayabahu, D., & Adeleke, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.28859 · Cited by 2
A Causal Model for Locating and Unlocking Sandbagging in Model OrganismsPreprint
Tan, H. K., Le, L., & Williams-King, D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.29461 · Cited by 2
Math for AI safety: an invitation for mathematiciansPreprint
Levine, L. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.15289 · Cited by 2
Reblitz-Richardson, O. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.14759 · Cited by 2
Unread or Unenforced? Separating Representation from Enforcement Failure in Content GuardsPreprint
Zhang, H., Feng, Y., Zheng, S., Liu, H., Xu, H., Luo, X., Wang, Z., Zandsalimy, M., & Sushmita, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.26178 · Cited by 2
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model AnswersPreprint
Ahmed, M. S., Poech, L. G., & Röttger, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.31342 · Cited by 2
A mechanistic study of language model introspectionPreprint
Zou, J., Sun, X., Kong, L., & Wang, T. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.35108 · Cited by 2
LLMs Learn to Evade Latent Monitors from Prior Feedback AlonePreprint
Keenan, H. L., Leckie, C. A., & Erfani, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.36490 · Cited by 3
The Sharp Directions Are Against You: Curvature Analysis of Activation SteeringPreprint
Azizi, S., Fayyazi, A., Razmara, P., Potraghloo, E. B., Kundu, S., & Pedram, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.16084 · Cited by 2
Verbalizable Representations Form a Global Workspace in Language Models
Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., et al. (2026) · arXiv · doi:10.48550/arxiv.2607.15495 · Cited by 1
Self-citationZhang, S., Zhai, Y., Guo, K., Hu, H., Guo, S., Fang, Z., Zhao, L., Shen, C., Wang, C., & Wang, Q. (2025) · arXiv · doi:10.48550/arxiv.2502.07557 · Cited by 1
Robustly Improving LLM Fairness in Realistic Settings via Interpretability
Karvonen, A., & Marks, S. (2025) · arXiv · doi:10.48550/arxiv.2506.10922 · Cited by 1
Can sparse autoencoders be used to decompose and interpret steering vectors?
Mayne, H., Yang, Y., & Mahdi, A. (2024) · arXiv · doi:10.48550/arxiv.2411.08790 · Cited by 1
Controlling Language and Style of Multi-lingual Generative Language Models with Control Vectors
Leino, J., & Karlgren, J. (2025) · Northern European Journal of Language Technology · doi:10.3384/nejlt.2000-1533.2025.5888 · Cited by 1
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.