Tracing Bias Propagation in Large Foundation Models: A Path-Level Framework for Fairness Auditing and Intervention
Keywords:
foundation models, bias auditing, fairness, path-level intervention, mechanistic interpretability, socio-technical systemsAbstract
The rapid deployment of large foundation models across critical societal domains has intensified concerns about encoded biases that produce discriminatory outcomes. Existing fairness auditing methods predominantly rely on input-output behavioral testing, a surface-level approach that fails to reveal how bias propagates through the internal computational graph of transformer architectures. This paper proposes a path-level framework for tracing bias propagation in large foundation models, enabling granular auditing and targeted intervention. Within the framework, model computation is modeled as a directed acyclic graph of activation pathways spanning attention heads, feedforward layers, and residual streams. Bias metrics are defined at the path level, capturing how stereotypical representations are formed, amplified, or suppressed along specific routes. The framework supports a dual function: first, it allows systematic auditing by localizing discriminatory pathways and quantifying their contribution to final outputs; second, it facilitates fine-grained intervention strategies such as path pruning, value editing, and contrastive regularization. The paper examines structural trade-offs between auditing fidelity and computational overhead, discusses infrastructure requirements for runtime path monitoring in production systems, and addresses governance implications for model transparency and regulatory compliance. A system-level perspective reveals that path-level fairness mechanisms must be integrated into the full model lifecycle, from pretraining to deployment and continuous monitoring. The analysis extends to sustainability dimensions, acknowledging the energy costs of high-resolution bias tracing. By framing fairness as an emergent property of path-level computational dynamics, the paper contributes a scalable conceptual architecture that bridges mechanistic interpretability and sociotechnical fairness, opening avenues for more accountable foundation model ecosystems.
References
1. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). ACM.
2. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., ... & Liang, P. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.
3. Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in Neural Information Processing Systems, 29, 4349–4357.
4. Zhao, J., Wang, T., Yatskar, M., Ordonez, V., & Chang, K. W. (2019). Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 15–20). Association for Computational Linguistics.
5. Nadeem, M., Bethke, A., & Reddy, S. (2020). StereoSet: Measuring stereotypical bias in pretrained language models. arXiv preprint arXiv:2004.09456.
6. Gehman, S., Gururangan, S., Sap, M., Choi, Y., & Smith, N. A. (2020). RealToxicityPrompts: Evaluating neural toxic degeneration in language models. In Findings of the Association for Computational Linguistics: EMNLP 2020 (pp. 3356–3369). Association for Computational Linguistics.
7. Dathathri, S., Madotto, A., Lan, J., Hung, J., Frank, E., Molino, P., ... & Goyal, Y. (2020). Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representations.
8. Meng, K., Bau, D., Andonian, A., & Belinkov, Y. (2022). Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems, 35, 17359–17372.
9. Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., ... & Olah, C. (2021). A mathematical framework for transformer circuits. Transformer Circuits Thread, Anthropic.
10. Geva, M., Caciularu, A., Wang, K. R., & Goldberg, Y. (2022). Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (pp. 30–45). Association for Computational Linguistics.
11. Vig, J., & Belinkov, Y. (2020). Analyzing the structure of attention in a transformer language model. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (pp. 63–70). Association for Computational Linguistics.
12. Liang, P. P., Wu, C., Morency, L. P., & Salakhutdinov, R. (2021). Towards understanding and mitigating social biases in language models. In International Conference on Machine Learning (pp. 6565–6576). PMLR.
13. Schick, T., Udupa, S., & Schütze, H. (2021). Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP. Transactions of the Association for Computational Linguistics, 9, 1408–1424.
14. Barocas, S., Hardt, M., & Narayanan, A. (2019). Fairness and machine learning. fairmlbook.org.
15. Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019). Fairness and abstraction in sociotechnical systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 59–68). ACM.
16. C. Shi, S. Li, W. Lu, W. Wu, C. Wang, Z. Cheng, F. Shen, and T. Chua (2026)TraceRouter: robust safety for large foundation models via path-level intervention.arXiv preprint arXiv:2601.21900.
17. Belrose, N., Scherlis, A., & Steinhardt, J. (2023). Eliciting latent knowledge from language models with classifier-free guidance. arXiv preprint arXiv:2302.06675.
18. Dinan, E., Fan, A., Williams, A., Urbanek, J., Kiela, D., & Weston, J. (2020). Queens are powerful too: Mitigating gender bias in dialogue generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (pp. 8173–8188). Association for Computational Linguistics.
19. Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., ... & Olah, C. (2022). In-context learning and induction heads. Transformer Circuits Thread, Anthropic.
20. Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., ... & Barnes, P. (2020). Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 33–44). ACM.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Journal of Data Intelligence and AI Systems
This work is licensed under a Creative Commons Attribution 4.0 International License.
This article is published under the Creative Commons Attribution 4.0 International License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
