Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?
- URL: http://arxiv.org/abs/2512.20796v1
- Date: Tue, 23 Dec 2025 21:44:20 GMT
- Title: Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?
- Authors: Zhengyang Shan, Aaron Mueller,
- Abstract summary: We investigate how independent demographic bias mechanisms are from general demographic recognition in language models.<n>We find that attribution-based ablations mitigate race and gender profession stereotypes while preserving name recognition accuracy.<n>We find that correlation-based ablations are more effective for education bias.
- Score: 17.978167351646288
- License: http://creativecommons.org/licenses/by/4.0/
- Abstract: 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.
Related papers
- How far can bias go? -- Tracing bias from pretraining data to alignment [54.51310112013655]
This study examines the correlation between gender-occupation bias in pre-training data and their manifestation in LLMs.<n>Our findings reveal that biases present in pre-training data are amplified in model outputs.
arXiv Detail & Related papers (2024-11-28T16:20:25Z) - MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge [6.208151505901749]
Models trained on real-world data often mirror and exacerbate existing social biases.<n>We introduce a novel adversarial training strategy that operates independently of prior bias-type knowledge.<n>Our method effectively reduces social biases without the need for demographic annotations.
arXiv Detail & Related papers (2024-08-10T09:11:01Z) - Dissecting Causal Biases [0.0]
This paper focuses on a class of bias originating in the way training data is generated and/or collected.
Four sources of bias are considered, namely, confounding, selection, measurement, and interaction.
arXiv Detail & Related papers (2023-10-20T09:12:10Z) - Gender Biases in Automatic Evaluation Metrics for Image Captioning [87.15170977240643]
We conduct a systematic study of gender biases in model-based evaluation metrics for image captioning tasks.
We demonstrate the negative consequences of using these biased metrics, including the inability to differentiate between biased and unbiased generations.
We present a simple and effective way to mitigate the metric bias without hurting the correlations with human judgments.
arXiv Detail & Related papers (2023-05-24T04:27:40Z) - Metrics for Dataset Demographic Bias: A Case Study on Facial Expression Recognition [4.336779198334903]
One of the most prominent types of demographic bias are statistical imbalances in the representation of demographic groups in the datasets.
We develop a taxonomy for the classification of these metrics, providing a practical guide for the selection of appropriate metrics.
The paper provides valuable insights for researchers in AI and related fields to mitigate dataset bias and improve the fairness and accuracy of AI models.
arXiv Detail & Related papers (2023-03-28T11:04:18Z) - General Greedy De-bias Learning [163.65789778416172]
We propose a General Greedy De-bias learning framework (GGD), which greedily trains the biased models and the base model like gradient descent in functional space.
GGD can learn a more robust base model under the settings of both task-specific biased models with prior knowledge and self-ensemble biased model without prior knowledge.
arXiv Detail & Related papers (2021-12-20T14:47:32Z) - Statistical discrimination in learning agents [64.78141757063142]
Statistical discrimination emerges in agent policies as a function of both the bias in the training population and of agent architecture.
We show that less discrimination emerges with agents that use recurrent neural networks, and when their training environment has less bias.
arXiv Detail & Related papers (2021-10-21T18:28:57Z) - Balancing out Bias: Achieving Fairness Through Training Reweighting [58.201275105195485]
Bias in natural language processing arises from models learning characteristics of the author such as gender and race.
Existing methods for mitigating and measuring bias do not directly account for correlations between author demographics and linguistic variables.
This paper introduces a very simple but highly effective method for countering bias using instance reweighting.
arXiv Detail & Related papers (2021-09-16T23:40:28Z) - Towards Controllable Biases in Language Generation [87.89632038677912]
We develop a method to induce societal biases in generated text when input prompts contain mentions of specific demographic groups.
We analyze two scenarios: 1) inducing negative biases for one demographic and positive biases for another demographic, and 2) equalizing biases between demographics.
arXiv Detail & Related papers (2020-05-01T08:25:11Z)
This list is automatically generated from the titles and abstracts of the papers in this site.
This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.