mlsanityData to 14 Oct 2026

FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering

arXiv 2504.14492 ↗First posted 20 Apr 2025cs.CLcited by 40

On the map: Mechanistic interpretability → Probing deception, bias and beliefs

Abstract (arXiv)

Large language models (LLMs) are prone to capturing biases from training corpus, leading to potential negative social impacts. Existing prompt-based debiasing methods exhibit instability due to their sensitivity to prompt changes, while fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting. In this paper, we propose FairSteer, a novel inference-time debiasing framework without requiring customized prompt design or model retraining. Motivated by the linear representation hypothesis, our preliminary investigation demonstrates that fairness-related features can be encoded into separable directions in the hidden activation space. FairSteer operates in three steps: biased activation detection, debiasing steering vector (DSV) computation, and dynamic activation steering. Specifically, it first trains a lightweight linear classifier to detect bias signatures in activations, and then computes DSVs as intervention directions derived from small contrastive prompt pairs. Subsequently, it performs debiasing by adjusting activations with DSVs in the inference stage. Comprehensive evaluation with six LLMs demonstrates the superiority of FairSteer across question-answering, counterfactual input evaluation and open-ended text generation tasks. Code will be released.

Key facts. Cited by 40 papers in the map (data through 14 Oct 2026). Placed in Probing deception, bias and beliefs, one of 13 directions in Mechanistic interpretability. Builds on 62 of its 78 references that are in the map. First posted on arXiv 20 Apr 2025.

Builds on

62 of 78 references.

References · 62

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  2. 2020Language Models are Few-Shot Learnerscited by 32,419
  3. 2023GPT-4 Technical Reportcited by 18,396
  4. 2022Training language models to follow instructions with human feedbackcited by 15,971
  5. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 14,132
  6. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  7. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,886
  8. 2022Large Language Models are Zero-Shot Reasonerscited by 4,374
  9. 2016Pointer Sentinel Mixture Modelscited by 3,096
  10. 2020Learning to summarize from human feedbackcited by 2,588
  11. 2018Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answeringcited by 2,164
  12. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
38 more
  1. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
  2. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  3. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  4. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  5. 2022Toy Models of Superpositioncited by 971
  6. 2020CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Modelscited by 710
  7. 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 701
  8. 2023Bias and Fairness in Large Language Models: A Surveycited by 639
  9. 2019The Woman Worked as a Babysitter: On Biases in Language Generationcited by 541
  10. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 455
  11. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 441
  12. 2021BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generationcited by 412
  13. 2021Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLPcited by 337
  14. 2022Prompting GPT-3 To Be Reliablecited by 260
  15. 2020Measuring and Reducing Gendered Correlations in Pre-trained Modelscited by 224
  16. 2018Gender Bias in Neural Natural Language Processingcited by 211
  17. 2019Gender Bias in Contextualized Word Embeddingscited by 203
  18. 2021Societal Biases in Language Generation: Progress and Challengescited by 202
  19. 2019Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphologycited by 200
  20. 2020Social Biases in NLP Models as Barriers for Persons with Disabilitiescited by 182
  21. 2020Towards Debiasing Sentence Representationscited by 171
  22. 2023Linear Representations of Sentiment in Large Language Modelscited by 163
  23. 2020NeuroLogic Decoding: (Un)supervised Neural Text Generation with Predicate Logic Constraintscited by 149
  24. 2020UnQovering Stereotyping Biases via Underspecified Questionscited by 126
  25. 2023Increasing Diversity While Maintaining Accuracy: Text Data Generation with Large Language Models and Human Interventionscited by 125
  26. 2019It's All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitutioncited by 103
  27. 2021FairFil: Contrastive Neural Debiasing Method for Pretrained Text Encoderscited by 75
  28. 2021Mitigating Political Bias in Language Models Through Reinforced Calibrationcited by 73
  29. 2023Fast Model Debias with Machine Unlearningcited by 72
  30. 2024Uncovering Safety Risks of Large Language Models through Concept Activation Vectorcited by 67
  31. 2024Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypescited by 59
  32. 2023Bias Against 93 Stigmatized Groups in Masked Language Models and Downstream Sentiment Classification Taskscited by 51
  33. 2021Diverse Adversaries for Mitigating Bias in Trainingcited by 48
  34. 2022Critic-Guided Decoding for Controlled Text Generationcited by 45
  35. 2024A Language Model's Guide Through Latent Spacecited by 43
  36. 2023Using In-Context Learning to Improve Dialogue Safetycited by 42
  37. 2022Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Expertscited by 36
  38. 2024FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMscited by 36

50 most cited of 62 All 62 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 16; 2026: 24.

40 papers.

Sort:

By area:

  1. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  2. 2025Angular Steering: Behavior Control via Rotation in Activation Spacecited by 53
  3. 2025SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Modelscited by 29
  4. 2025Silenced Biases: The Dark Side LLMs Learned to Refusecited by 10
  5. 2025Guiding Giants: Lightweight Controllers for Weighted Activation Steering in LLMscited by 8
  6. 2025ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behaviorcited by 7
  7. 2025BiasFilter: An Inference-Time Debiasing Framework for Large Language Modelscited by 6
  8. 2026Selective Steering: Norm-Preserving Control Through Discriminative Layer Selectioncited by 6
  9. 2025C2PO: Diagnosing and Disentangling Bias Shortcuts in LLMscited by 5
  10. 2026The Hidden Signal of Verifier Strictness: Controlling and Improving Step-Wise Verification via Selective Latent Steeringcited by 5
  11. 2025BiasGuard: A Reasoning-enhanced Bias Detection Tool For Large Language Modelscited by 4
  12. 2025DSO: Direct Steering Optimization for Bias Mitigationcited by 3
28 more
  1. 2026Self-Debias: Self-correcting for Debiasing Large Language Modelscited by 3
  2. 2026Where Steering Signals Come From: Activation Source Selection in Activation Steeringcited by 3
  3. 2025BiasFreeBench: a Benchmark for Mitigating Bias in Large Language Model Responsescited by 2
  4. 2025Robust Preference Alignment via Directional Neighborhood Consensuscited by 2
  5. 2025Can SAEs reveal and mitigate racial biases of LLMs in healthcare?cited by 2
  6. 2026Knowing without Acting: The Disentangled Geometry of Safety Mechanisms in Large Language Modelscited by 2
  7. 2026Locating and Controlling Implicit Personalization in Large Language Modelscited by 2
  8. 2025Detection, Classification, and Mitigation of Gender Bias in Large Language Modelscited by 1
  9. 2025BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Injectioncited by 1
  10. 2026Tracing the Latent Threads: A Mechanistic Study of How LLMs Represent and Operationalize Race and Ethnicity Cuescited by 1
  11. 2026Faithful-Patchscopes: Understanding and Mitigating Model Bias in Hidden Representations Explanation of Large Language Modelscited by 1
  12. 2026Routing Sensitivity Without Controllability: A Diagnostic Study of Fairness in MoE Language Modelscited by 1
  13. 2026STEREODISCO: Discovering Stereotypicality in LLMscited by 1
  14. 2026GAPS: Dimension-Level Gates for Conditional Activation Steeringcited by 1
  15. 2025Debiasing LLMs by Masking Unfairness-Driving Attention Headscited by 0
  16. 2026Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancementcited by 0
  17. 2026Pref-CTRL: Preference Driven LLM Alignment using Representation Editingcited by 0
  18. 2026Steered Generation via Gradient-Based Optimization on Sparse Query Featurescited by 0
  19. 2026Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Modelscited by 0
  20. 2026The Fairness Collapse Phenomenon: Bias Amplification in Language Models Trained on Synthetic Datacited by 0
  21. 2026Your AI, On a Dial: Controlling Investment Bias in LLMs with a Single Neuroncited by 0
  22. 2026StateSwap: Probing Support-Elimination Hidden States in Multiple-Choice Questionscited by 0
  23. 2026Empathy Is Steerable but Multi-Axial: Mechanism Geometry and Persona Effects in LLMscited by 0
  24. 2026A Survey on the Linear Representation Hypothesiscited by 0
  25. 2026IMPLICIT-Bench: Measuring Implicit Bias in Text-to-Image Models under Neutral Promptscited by 0
  26. 2026Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventionscited by 0
  27. 2026Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steeringcited by 0
  28. 2026Unfair and Unsafe: Everyday Bias Triggers Emergent Misalignmentcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Latent space bias directions in LLMs capture confidence, not fairness
  2. 2026CatRAG: Functor-Guided Structural Debiasing with Retrieval Augmentation for Fair LLMs
  3. 2026DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation
  4. 2024Teacher-Student Training for Debiasing: General Permutation Debiasing for Large Language Models
  5. 2025Activation Steering for Bias Mitigation: An Interpretable Approach to Safer LLMs
  6. 2024Towards Resource Efficient and Interpretable Bias Mitigation in Large Language Models
  7. 2025Mitigating Biases in Language Models via Bias Unlearning
  8. 2025Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs