mlsanityData to 7 Oct 2026

Navigating the OverKill in Large Language Models

arXiv 2401.17633 ↗First posted 31 Jan 2024cs.CL cs.AIcited by 40

Not on the map. Why?

Cited by 27 in LLM safety and security, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to an over-attention of harmful words like 'kill' and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such over-attention by amplifying the difference in the model's output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the over-attention from the model via contrastive decoding. Empirical results indicate that our method has achieved an average reduction of the refusal rate by 20% while having almost no impact on safety.

Key facts. Cited by 40 papers in the map (data through 7 Oct 2026). Builds on 19 of its 31 references that are in the map. First posted on arXiv 31 Jan 2024.

Builds on

19 of 31 references.

References · 19

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 13,958
  3. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  4. 2022Large Language Models are Zero-Shot Reasonerscited by 4,323
  5. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
  6. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,348
  7. 2021SimCSE: Simple Contrastive Learning of Sentence Embeddingscited by 2,168
  8. 2020Shortcut Learning in Deep Neural Networkscited by 1,928
  9. 2018CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledgecited by 1,793
  10. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 709
  11. 2022Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learnedcited by 702
  12. 2022Contrastive Decoding: Open-ended Text Generation as Optimizationcited by 598
7 more
  1. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 445
  2. 2023Shadow Alignment: The Ease of Subverting Safely-Aligned Language Modelscited by 284
  3. 2023The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learningcited by 251
  4. 2023Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learningcited by 235
  5. 2022Human-Machine Collaboration Approaches to Build a Dialogue Dataset for Hate Speech Counteringcited by 40
  6. 2024Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedbackcited by 25
  7. 2022MINER: Improving Out-of-Vocabulary Named Entity Recognition from an Information Theoretic Perspectivecited by 17

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 8; 2025: 18; 2026: 14.

40 papers.

Sort:

By area:

  1. 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independencecited by 104
  2. 2025LLMs Encode Harmfulness and Refusal Separatelycited by 75
  3. 2025From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Trainingcited by 49
  4. 2024Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablationcited by 48
  5. 2024R2-Guard: Robust Reasoning Enabled LLM Guardrail via Knowledge-Enhanced Logical Reasoningcited by 41
  6. 2024Automatic Pseudo-Harmful Prompt Generation for Evaluating False Refusals in Large Language Modelscited by 35
  7. 2025Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuningcited by 22
  8. 2025X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Multi-Turn Jailbreaks without Compromising Usabilitycited by 20
  9. 2025FORTRESS: Frontier Risk Evaluation for National Security and Public Safetycited by 20
  10. 2024Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patchingcited by 16
  11. 2024You Know What I'm Saying: Jailbreak Attack via Implicit Referencecited by 13
  12. 2026Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Controlcited by 12
28 more
  1. 2025Think Twice, Generate Once: Safeguarding by Progressive Self-Reflectioncited by 11
  2. 2025On-the-fly Preference Alignment via Principle-Guided Decodingcited by 10
  3. 2024JailbreakLens: Visual Analysis of Jailbreak Attacks Against Large Language Modelscited by 7
  4. 2025Latent-space adversarial training with post-aware calibration for defending large language models against jailbreak attackscited by 7
  5. 2025EVOREFUSE: Evolutionary Prompt Optimization for Evaluation and Mitigation of LLM Over-Refusal to Pseudo-Malicious Instructionscited by 6
  6. 2024Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Controlcited by 5
  7. 2025Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Checkcited by 5
  8. 2025aiXamine: Simplified LLM Safety and Securitycited by 4
  9. 2026Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinementcited by 3
  10. 2025How Semantically Stable Are LLM Refusals? Measuring Confusion in Local Safety Boundariescited by 2
  11. 2026Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learningcited by 2
  12. 2026Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rulescited by 2
  13. 2024What's New in My Data? Novelty Exploration via Contrastive Generationcited by 1
  14. 2025MoGU V2: Toward a Higher Pareto Frontier Between Model Usability and Securitycited by 1
  15. 2025Reimagining Safety Alignment with An Imagecited by 1
  16. 2026Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levelscited by 1
  17. 2025MEUV: Achieving Fine-Grained Capability Activation in Large Language Models via Mutually Exclusive Unlock Vectorscited by 0
  18. 2025UpSafe^circC: Upcycling for Controllable Safety in Large Language Modelscited by 0
  19. 2025Understanding and Mitigating Over-refusal for Large Language Models via Representation Interventioncited by 0
  20. 2026Can One-sided Arguments Lead to Response Change in Large Language Models?cited by 0
  21. 2026Residual Stream Analysis of Overfitting And Structural Disruptionscited by 0
  22. 2026A Systematic Study of Training-Free Methods for Trustworthy Large Language Modelscited by 0
  23. 2026Mitigating Many-shot Jailbreak Attacks with One Single Demonstrationcited by 0
  24. 2026Safety Cost of Steering Vectors Is Separable and Reduciblecited by 0
  25. 2026aiXamine: Unified Black-Box Evaluation of Cross-Dimensional Trade-offs in LLM Safety, Security, and Privacycited by 0
  26. 2026Door-in-the-Face Requests and Refusal Behaviour in Large Language Modelscited by 0
  27. 2026Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibrationcited by 0
  28. 2026EOPSA: Efficient On-Policy Self-Distilled Safety Alignmentcited by 0