mlsanityData to 7 Oct 2026

Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron

arXiv 2602.02027 ↗First posted 2 Feb 2026cs.AI cs.LGcited by 0

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 47 of its 55 references that are in the map. First posted on arXiv 2 Feb 2026.

Builds on

47 of 55 references.

Closest ancestors · 7

Same direction, newest first

  1. 2025Breaking the Safety-Capability Tradeoff: Reinforcement Learning with Verifiable Rewards Maintains Safety Guardrails in LLMs26 Nov
  2. 2025CARE: Decoding Time Safety Alignment via Rollback and Introspection Intervention1 Sep
  3. 2025FalseReject: A Resource for Improving Contextual Safety and Mitigating Over-Refusals in LLMs via Structured Reasoning12 May
  4. 2025SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals3 Feb
  5. 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deep10 Jun
  6. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning5 Mar
  7. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!5 Oct

Other references · 40

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2024The Llama 3 Herd of Modelscited by 14,192
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  5. 2025Qwen3 Technical Reportcited by 8,195
  6. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  7. 2023Qwen Technical Reportcited by 3,273
  8. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  9. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 2,914
  10. 2023Jailbreaking Black Box Large Language Models in Twenty Queriescited by 1,326
  11. 2022Red Teaming Language Models with Language Modelscited by 983
  12. 2024Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluatorscited by 865
28 more
  1. 2023AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Modelscited by 738
  2. 2022Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learnedcited by 702
  3. 2024TinyLlama: An Open-Source Small Language Modelcited by 604
  4. 2024How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMscited by 533
  5. 2023GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Promptscited by 520
  6. 2019Neural Text Generation with Unlikelihood Trainingcited by 489
  7. 2023XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Modelscited by 477
  8. 2023Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrationscited by 401
  9. 2023The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learningcited by 251
  10. 2024SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decodingcited by 233
  11. 2023LLM Self Defense: By Self Examination, LLMs Know They Are Being Trickedcited by 226
  12. 2023Mitigating the Alignment Tax of RLHFcited by 181
  13. 2025Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teamingcited by 172
  14. 2025Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonablecited by 120
  15. 2024Revisiting Catastrophic Forgetting in Large Language Model Tuningcited by 93
  16. 2024Reinforcement Learning for LLM Post-Training: A Surveycited by 91
  17. 2024PAD: Personalized Alignment of LLMs at Decoding-Timecited by 58
  18. 2025PromptShield: Deployable Detection for Prompt Injection Attackscited by 43
  19. 2024Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Modelscited by 21
  20. 2024Dynamic Rewarding with Prompt Optimization Enables Tuning-free Self-Alignment of Language Modelscited by 12
  21. 2025Persona-judge: Personalized Alignment of Large Language Models via Token-level Self-judgmentcited by 11
  22. 2025Alignment-Augmented Speculative Decoding with Alignment Sampling and Conditional Verificationcited by 10
  23. 2020A bio-inspired bistable recurrent cell allows for long-lasting memorycited by 8
  24. 2025Improved Supervised Fine-Tuning for Large Language Models to Mitigate Catastrophic Forgettingcited by 8
  25. 2025Speculative Safety-Aware Decodingcited by 7
  26. 2025PITA: Preference-Guided Inference-Time Alignment for LLM Post-Trainingcited by 2
  27. 2025RLHF: A comprehensive Survey for Cultural, Multimodal and Low Latency Alignment Methodscited by 2
  28. 2025Bidirectional Intention Inference Enhances LLMs' Defense Against Multi-Turn Jailbreak Attackscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance
  2. 2026Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order Optimization
  3. 2026NeST: Neuron Selective Tuning for LLM Safety
  4. 2026DNAlign: Dynamic Null-Space Safe Alignment for LLMs
  5. 2024Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons
  6. 2025NeuronTune: Fine-Grained Neuron Modulation for Balanced Safety-Utility Alignment in LLMs
  7. 2026SafeNeuron: Neuron-Level Safety Alignment for Large Language Models
  8. 2026Safeguarding LLM Fine-tuning via Push-Pull Distributional Alignment