mlsanityData to 14 Oct 2026

Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?

arXiv 2512.20796 ↗First posted 23 Dec 2025cs.CLcited by 1

Not on the map. Why?

Abstract (arXiv)

We investigate how independent demographic bias mechanisms are from general demographic recognition in language models. Using a multi-task evaluation setup where demographics are associated with names, professions, and education levels, we measure whether models can be debiased while preserving demographic detection capabilities. We compare attribution-based and correlation-based methods for locating bias features. We find that targeted sparse autoencoder feature ablations in Gemma-2-9B reduce bias without degrading recognition performance: attribution-based ablations mitigate race and gender profession stereotypes while preserving name recognition accuracy, whereas correlation-based ablations are more effective for education bias. Qualitative analysis further reveals that removing attribution features in education tasks induces ``prior collapse'', thus increasing overall bias. This highlights the need for dimension-specific interventions. Overall, our results show that demographic bias arises from task-specific mechanisms rather than absolute demographic markers, and that mechanistic inference-time interventions can enable surgical debiasing without compromising core model capabilities.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 25 of its 32 references that are in the map. First posted on arXiv 23 Dec 2025.

Builds on

25 of 32 references.

References · 25

  1. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  2. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  3. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  5. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  6. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  7. 2020Language (Technology) is Power: A Critical Survey of "Bias" in NLPcited by 956
  8. 2023Steering Language Models With Activation Engineeringcited by 922
  9. 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 701
  10. 2018Gender Bias in Coreference Resolution: Evaluation and Debiasing Methodscited by 659
  11. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 532
  12. 2018Gender Bias in Coreference Resolutioncited by 442
13 more
  1. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  2. 2021BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generationcited by 412
  3. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  4. 2021Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLPcited by 337
  5. 2013Direct and Indirect Effectscited by 280
  6. 2019Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Themcited by 268
  7. 2024Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanismscited by 149
  8. 2021Causal Intervention for Leveraging Popularity Bias in Recommendationcited by 116
  9. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  10. 2024NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internalscited by 45
  11. 2025Social Bias Benchmark for Generation: A Comparison of Generation and QA-Based Evaluationscited by 16
  12. 2025Aligned but Blind: Alignment Increases Implicit Bias by Reducing Awareness of Racecited by 15
  13. 2025Dissecting Bias in LLMs: A Mechanistic Interpretability Perspectivecited by 15

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Topics as Proxies for Sociodemographics: How Conversational Context Affects LLM Answerscited by 2

Similar, not cited

Close in wording, no citation either way

  1. 2025Mitigating Biases in Language Models via Bias Unlearning
  2. 2025DemoBias: An Empirical Study to Trace Demographic Biases in Vision Foundation Models
  3. 2026ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues
  4. 2024Towards Resource Efficient and Interpretable Bias Mitigation in Large Language Models
  5. 2025Investigating Intersectional Bias in Large Language Models using Confidence Disparities in Coreference Resolution
  6. 2026Decoding-Time Debiasing via Process Reward Models: From Controlled Fill-in to Open-Ended Generation
  7. 2025Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to Models
  8. 2026Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder