mlsanityData to 7 Oct 2026

Refusal Behavior in Large Language Models: A Nonlinear Perspective

arXiv 2501.08145 ↗First posted 14 Jan 2025cs.CL cs.AIcited by 13

Not on the map. Why?

Abstract (arXiv)

Refusal behavior in large language models (LLMs) enables them to decline responding to harmful, unethical, or inappropriate prompts, ensuring alignment with ethical standards. This paper investigates refusal behavior across six LLMs from three architectural families. We challenge the assumption of refusal as a linear phenomenon by employing dimensionality reduction techniques, including PCA, t-SNE, and UMAP. Our results reveal that refusal mechanisms exhibit nonlinear, multidimensional characteristics that vary by model architecture and layer. These findings highlight the need for nonlinear interpretability to improve alignment research and inform safer AI deployment strategies.

Key facts. Cited by 13 papers in the map (data through 7 Oct 2026). Builds on 14 of its 31 references that are in the map. First posted on arXiv 14 Jan 2025.

Builds on

14 of 31 references.

References · 14

  1. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 2,914
  2. 2018UMAP: Uniform Manifold Approximation and Projection for Dimension Reductioncited by 2,841
  3. 2024Qwen2 Technical Reportcited by 2,076
  4. 2022BLOOM: A 176B-Parameter Open-Access Multilingual Language Modelcited by 1,968
  5. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  6. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  7. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  8. 2022Toy Models of Superpositioncited by 928
  9. 2023Steering Language Models With Activation Engineeringcited by 824
  10. 2024Alignment faking in large language modelscited by 316
  11. 2016Linear Algebraic Structure of Word Senses, with Applications to Polysemycited by 165
  12. 2015Sparse Overcomplete Word Vector Representationscited by 43
2 more
  1. 2018How deep is deep enough? -- Quantifying class separability in the hidden layers of deep neural networkscited by 28
  2. 2024What Features in Prompts Jailbreak LLMs? Investigating the Mechanisms Behind Attackscited by 24

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 8; 2026: 5.

13 papers.

Sort:

By area:

  1. 2025Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safetycited by 10
  2. 2025The Geometry of Harmfulness in LLMs through Subconcept Probingcited by 7
  3. 2025From Directions to Cones: Exploring Multidimensional Representations of Propositional Facts in LLMscited by 3
  4. 2025Poison Once, Refuse Forever: Weaponizing Alignment for Injecting Bias in LLMscited by 1
  5. 2025AlignTree: Efficient Defense Against LLM Jailbreak Attackscited by 1
  6. 2025Graph-Regularized Sparse Autoencoders for LLM Safety Steeringcited by 1
  7. 2026BabelSteering: Multilingual Safety Alignment via English Steering Vectorscited by 1
  8. 2025From Rogue to Safe AI: The Role of Explicit Refusals in Aligning LLMs with International Humanitarian Lawcited by 0
  9. 2025State-Dependent Refusal and Learned Incapacity in RLHF-Aligned Language Modelscited by 0
  10. 2026Neuroscience-Inspired Analyses of Visual Interestingness in Multimodal Transformerscited by 0
  11. 2026Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activationscited by 0
  12. 2026Refusal Lives Downstream of Persona in Chat Modelscited by 0
1 more
  1. 2026The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuningcited by 0