mlsanityData to 14 Oct 2026

Dissecting Bias in LLMs: A Mechanistic Interpretability Perspective

arXiv 2506.05166 ↗First posted 5 Jun 2025cs.CL cs.AIcited by 15

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

Abstract (arXiv)

Large Language Models (LLMs) are known to exhibit social, demographic, and gender biases, often as a consequence of the data on which they are trained. In this work, we adopt a mechanistic interpretability approach to analyze how such biases are structurally represented within models such as GPT-2 and Llama2. Focusing on demographic and gender biases, we explore different metrics to identify the internal edges responsible for biased behavior. We then assess the stability, localization, and generalizability of these components across dataset and linguistic variations. Through systematic ablations, we demonstrate that bias-related computations are highly localized, often concentrated in a small subset of layers. Moreover, the identified components change across fine-tuning settings, including those unrelated to bias. Finally, we show that removing these components not only reduces biased outputs but also affects other NLP tasks, such as named entity recognition and linguistic acceptability judgment because of the sharing of important components with these tasks.

Key facts. Cited by 15 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 25 of its 38 references that are in the map. First posted on arXiv 5 Jun 2025.

Builds on

25 of 38 references.

Closest ancestors · 1

Same direction, newest first

  1. 2025Linear Representations of Political Perspective Emerge in Large Language Models3 Mar

Other references · 24

  1. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  2. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
  3. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  4. 2018Neural Network Acceptability Judgmentscited by 958
  5. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  6. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  7. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  8. 2021Causal Abstractions of Neural Networkscited by 422
  9. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 359
  10. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 209
  11. 2019Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphologycited by 200
  12. 2023Localizing Model Behavior with Path Patchingcited by 195
12 more
  1. 2023Gender bias and stereotypes in Large Language Modelscited by 187
  2. 2023Interpretability at Scale: Identifying Causal Mechanisms in Alpacacited by 103
  3. 2024Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypescited by 59
  4. 2023Nationality Bias in Text Generationcited by 50
  5. 2023Investigating Subtler Biases in LLMs: Ageism, Beauty, Institutional, and Nationality Bias in Generative Modelscited by 32
  6. 2024Backward Lens: Projecting Language Model Gradients into the Vocabulary Spacecited by 32
  7. 2023Gender Biases in Automatic Evaluation Metrics for Image Captioningcited by 24
  8. 2023Identifying and Adapting Transformer-Components Responsible for Gender Bias in an English Language Modelcited by 22
  9. 2024ChatGPT Based Data Augmentation for Improved Parameter-Efficient Debiasing of LLMscited by 18
  10. 2025BiasEdit: Debiasing Stereotyped Language Models via Model Editingcited by 18
  11. 2016Quantifying and Reducing Stereotypes in Word Embeddingscited by 15
  12. 2023The Impact of Debiasing on the Performance of Language Models in Downstream Tasks is Underestimatedcited by 7

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 5; 2026: 10.

15 papers.

Sort:

By area:

  1. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  2. 2025Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretabilitycited by 2
  3. 2025Explainability of Large Language Models: Opportunities and Challenges toward Generating Trustworthy Explanationscited by 2
  4. 2026The Homogenization Problem in LLMs: Towards Meaningful Diversity in AI Safetycited by 2
  5. 2026Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Modelscited by 2
  6. 2025BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Injectioncited by 1
  7. 2025The Psychology of Falsehood: A Human-Centric Survey of Misinformation Detectioncited by 1
  8. 2025Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?cited by 1
  9. 2026Towards general embodied intelligence: integrating large language models, knowledge bases, and reasoning capabilities to build the next generation of AI agentscited by 1
  10. 2026White-Box Sensitivity Auditing with Steering Vectorscited by 0
  11. 2026UGID: Unified Graph Isomorphism for Debiasing Large Language Modelscited by 0
  12. 2026GKnow: Measuring the Entanglement of Gender Bias and Factual Gendercited by 0
3 more
  1. 2026A Mechanistic Understanding of Pronoun Fidelity in LLMscited by 0
  2. 2026S^3martCirc: Self-supervised Smart Circuit Discoverycited by 0
  3. 2026Thin Evidence, Thick Priors: How Language Models Substitute Identity for Missing Financial Factscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Interpreting Bias in Large Language Models: A Feature-Based Approach
  2. 2024Towards Resource Efficient and Interpretable Bias Mitigation in Large Language Models
  3. 2025Unsupervised Concept Vector Extraction for Bias Control in LLMs
  4. 2026Position: LLMs Must Use Functor-Based and RAG-Driven Bias Mitigation for Fairness
  5. 2024Decoding Biases: Automated Methods and LLM Judges for Gender Bias Detection in Language Models
  6. 2025Activation Steering for Bias Mitigation: An Interpretable Approach to Safer LLMs
  7. 2026Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study
  8. 2024CBEval: A framework for evaluating and interpreting cognitive biases in LLMs