A Representation-Level Assessment of Bias Mitigation in Foundation Models
Not on the map. Why?
Abstract (arXiv)
We investigate how successful bias mitigation reshapes the embedding space of encoder-only and decoder-only foundation models, offering an internal audit of model behaviour through representational analysis. Using BERT and Llama2 as representative architectures, we assess the shifts in associations between gender and occupation terms by comparing baseline and bias-mitigated variants of the models. Our findings show that bias mitigation reduces gender-occupation disparities in the embedding space, leading to more neutral and balanced internal representations. These representational shifts are consistent across both model types, suggesting that fairness improvements can manifest as interpretable and geometric transformations. These results position embedding analysis as a valuable tool for understanding and validating the effectiveness of debiasing methods in foundation models. To further promote the assessment of decoder-only models, we introduce WinoDec, a dataset consisting of 4,000 sequences with gender and occupation terms, and release it to the general public. (https://github.com/winodec/wino-dec)
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 30 of its 39 references that are in the map. First posted on arXiv 17 Mar 2026.
Builds on
30 of 39 references.
References · 30
- 2017Attention Is All You Needcited by 60,361
- 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
- 2022Training language models to follow instructions with human feedbackcited by 15,971
- 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
- 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 13,028
- 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,375
- 2019A Survey on Bias and Fairness in Machine Learningcited by 1,917
- 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
- 2020StereoSet: Measuring stereotypical bias in pretrained language modelscited by 924
- 2016Semantics derived automatically from language corpora contain human-like biasescited by 901
- 2020CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Modelscited by 710
- 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 701
18 more
- 2018Gender Bias in Coreference Resolution: Evaluation and Debiasing Methodscited by 659
- 2023Bias and Fairness in Large Language Models: A Surveycited by 639
- 2018Gender Bias in Coreference Resolutioncited by 442
- 2019On Measuring Social Biases in Sentence Encoderscited by 398
- 2020Measuring and Reducing Gendered Correlations in Pre-trained Modelscited by 224
- 2018Gender Bias in Neural Natural Language Processingcited by 211
- 2019Gender Bias in Contextualized Word Embeddingscited by 203
- 2018Dissecting Contextual Word Embeddings: Architecture and Representationcited by 146
- 2019It's All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitutioncited by 103
- 2024Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrievalcited by 51
- 2020Double-Hard Debias: Tailoring Word Embeddings for Gender Bias Mitigationcited by 36
- 2024Creativity Has Left the Chat: The Price of Debiasing Language Modelscited by 29
- 2022Text Style Transfer for Bias Mitigation using Masked Language Modelingcited by 28
- 2022Gender Biases and Where to Find Them: Exploring Gender Bias in Pre-Trained Transformer-based Language Models Using Movement Pruningcited by 21
- 2023The Impact of Debiasing on the Performance of Language Models in Downstream Tasks is Underestimatedcited by 7
- 2025Concept-Level Explainability for Auditing & Steering LLM Responsescited by 6
- 2024Stereotype Detection in LLMs: A Multiclass, Explainable, and Benchmark-Driven Approachcited by 3
- 2024Bias in Text Embedding Modelscited by 1
Built on by
Not cited yet.