{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,22]],"date-time":"2026-09-22T06:06:32Z","timestamp":1790057192593,"version":"4.0.1"},"reference-count":24,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,9,22]],"date-time":"2026-09-22T00:00:00Z","timestamp":1790035200000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Artif. Intell."],"abstract":"<jats:sec>\n                    <jats:title>Introduction<\/jats:title>\n                    <jats:p>Large Language Models (LLMs) are becoming widely adopted for reasoning and decision-support tasks, yet they can inherit and reproduce gender-related biases present in their training data. Most existing bias-mitigation strategies depend on fine-tuning, reinforcement learning, prompt engineering, or interventions during pre-training, requiring either parameter updates or extensive manual prompt design. These requirements limit their applicability when the underlying model is available only as a closed-source or black-box system.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>\n                      This work explores whether bias-mitigation knowledge can instead be separated from the model and reused without altering its parameters. We propose a feedback-driven external alignment memory framework for\n                      <jats:italic>post-hoc<\/jats:italic>\n                      gender bias mitigation that transforms identified bias into concise corrective rules through an iterative feedback process. These rules are stored independently of the model parameters and are retrieved during inference to influence future responses. Retrieval is performed using cosine-similarity matching between an incoming query and previously stored examples, allowing corrective knowledge to be reused while maintaining complete model independence.\n                    <\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>Evaluation on selected subsets of BBQ, BiasNLI, CoBias, CrowS-Pairs, and WinoBias shows measurable reductions in gender-related bias. In the 120-record evaluation, Gender Assumption (GA) decreases from 15.83 to 7.08%, while Stereotypical Gender Assumption (SGA) decreases from 24.16 to 7.08%. Gender Neutral responses increase from 75.00 to 90.415%, and response quality improves from 4.15 to 4.211. A larger 500-record evaluation further indicates that corrective rules learned earlier continue to provide benefits when applied beyond the original feedback corpus, reducing GA from 13.8 to 12.0% and SGA from 14.4 to 11.0%. During the same evaluation, Gender Neutral responses increase from 78.6 to 83.2%, while response quality improves from 4.00 to 4.05.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Discussion<\/jats:title>\n                    <jats:p>These results indicate that corrective alignment knowledge can be maintained as a reusable external memory and incorporated during inference to improve fairness in black-box large language models without retraining or modifying model parameters.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/frai.2026.1925701","type":"journal-article","created":{"date-parts":[[2026,9,22]],"date-time":"2026-09-22T05:40:34Z","timestamp":1790055634000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Feedback-driven rule induction and retrieval-augmented bias mitigation for large language models"],"prefix":"10.3389","volume":"9","author":[{"given":"Abhishek","family":"Kumar","sequence":"first","affiliation":[{"name":"School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology","place":["Chennai, India"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Aishwaryaa Shree","family":"Muralitharan","sequence":"additional","affiliation":[{"name":"School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology","place":["Chennai, India"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Vinesh Kannaa","family":"Balaji","sequence":"additional","affiliation":[{"name":"School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology","place":["Chennai, India"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Bhargavi","family":"Renta Chintala","sequence":"additional","affiliation":[{"name":"School of Computer Science and Engineering (SCOPE), Vellore Institute of Technology","place":["Chennai, India"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,9,22]]},"reference":[{"key":"B1","article-title":"\u201cThe problem with bias: from allocative to representational harms in machine learning,\u201d","author":"Barocas","year":"2017","journal-title":"Proceedings of SIGCIS"},{"key":"B2","first-page":"3615","article-title":"\u201cRetrieval-augmented contrastive learning for dynamic graph anomaly detection,\u201d","volume-title":"Proceedings of the 49th international ACM SIGIR conference on research and development in information retrieval, SIGIR '26","author":"Berahmand","year":""},{"key":"B3","first-page":"3620","article-title":"\u201cRetrieval-augmented contrastive learning for knowledge tracing,\u201d","volume-title":"Proceedings of the 49th international ACM SIGIR conference on research and development in information retrieval, SIGIR '26","author":"Berahmand","year":""},{"key":"B4","first-page":"4356","article-title":"\u201cMan is to computer programmer as woman is to homemaker? debiasing word embeddings,\u201d","volume-title":"Proceedings of the 30th international conference on neural information processing systems (NIPS'16)","author":"Bolukbasi","year":"2016"},{"key":"B5","doi-asserted-by":"publisher","first-page":"183","DOI":"10.1126\/science.aal4230","article-title":"Semantics derived automatically from language corpora contain human-like biases","volume":"356","author":"Caliskan","year":"2017","journal-title":"Science"},{"key":"B6","doi-asserted-by":"publisher","first-page":"104278","DOI":"10.1016\/j.ipm.2025.104278","article-title":"Medscalere-pf: a prompt-based framework with retrieval-augmented generation, chain-of-thought, and self-verification for scale-specific relation extraction in Chinese medical literature","volume":"62","author":"Chen","year":"2025","journal-title":"Info. Proc. Manag"},{"key":"B7","article-title":"\u201cBias in natural language inference,\u201d","author":"Dev","year":"2020"},{"key":"B8","doi-asserted-by":"publisher","first-page":"330","DOI":"10.1145\/230538.230561","article-title":"Bias in computer systems","volume":"14","year":"1996","journal-title":"ACM Transac. Informat. Syst"},{"key":"B9","first-page":"58","article-title":"\u201cCobias: assessing the contextual reliability of bias benchmarks,\u201d","author":"Govil","year":"2025","journal-title":"Proceedings of the 17th ACM Web Science Conference (WebSci '25"},{"key":"B10","article-title":"Post-hoc popularity bias correction in GNN-based collaborative filtering","author":"Islam","year":"2025","journal-title":"arXiv"},{"key":"B11","article-title":"No LLM is free from bias: a comprehensive study of bias evaluation in large language models","author":"Kumar","year":"2025","journal-title":"arXiv"},{"key":"B12","first-page":"166","article-title":"\u201cMeasuring bias in contextualized word representations,\u201d","author":"Kurita","year":"2019","journal-title":"Gender Bias in NLP Workshop"},{"key":"B13","doi-asserted-by":"publisher","first-page":"9459","DOI":"10.48550\/arXiv.2005.11401","article-title":"\u201cRetrieval-augmented generation for knowledge-intensive NLP tasks. Adv","volume":"33","author":"Lewis","year":"2020","journal-title":"Neural Informat. Proc. Syst"},{"key":"B14","doi-asserted-by":"publisher","first-page":"106738","DOI":"10.1016\/j.autcon.2025.106738","article-title":"Enhancing llm-based building data query with chain-of-thought, retrieval-augmented generation, and fine-tuning","volume":"182","author":"Li","year":"2026","journal-title":"Automation in Construction"},{"key":"B15","doi-asserted-by":"publisher","first-page":"104369","DOI":"10.1016\/j.ipm.2025.104369","article-title":"Towards evidence-aware retrieval-augmented generation via self-corrective chain-of-thought","volume":"63","author":"Li","year":"2026","journal-title":"Informat. Proc. Manag"},{"key":"B16","author":"Manerba","year":"2024"},{"key":"B17","doi-asserted-by":"publisher","first-page":"4946","DOI":"10.1016\/j.procs.2025.09.621","article-title":"Enhancing clinical decision support with llms: a feasibility study integrating Cot, Rag, and Qlora","volume":"270","year":"2025","journal-title":"Procedia Computer Sci"},{"key":"B18","doi-asserted-by":"publisher","author":"Nadeem","year":"2021","DOI":"10.18653\/v1\/2021.acl-long.416"},{"key":"B19","first-page":"1953","article-title":"\u201cCrows-pairs: a challenge dataset for measuring social biases in masked language models,\u201d","author":"Nangia","year":"2020","journal-title":"Proceedings of EMNLP"},{"key":"B20","author":"Parrish","year":"2022"},{"key":"B21","doi-asserted-by":"publisher","author":"Wan","year":"2023","DOI":"10.18653\/v1\/2023.findings-emnlp.243"},{"key":"B22","doi-asserted-by":"publisher","first-page":"100630","DOI":"10.1109\/ACCESS.2025.3578192","article-title":"Ai-driven post-earthquake emergency material demand prediction: integrating rag with reasoning large language model","volume":"13","author":"Zhang","year":"2025","journal-title":"IEEE Access"},{"key":"B23","doi-asserted-by":"publisher","first-page":"114526","DOI":"10.1016\/j.knosys.2025.114526","article-title":"Treeqa: enhanced llm-rag with logic tree reasoning for reliable and interpretable multi-hop question answering","volume":"330","author":"Zhang","year":"2025","journal-title":"Knowledge-Based Systems"},{"key":"B24","volume-title":"Gender bias in coreference","author":"Zhao","year":"2018"}],"container-title":["Frontiers in Artificial Intelligence"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/frai.2026.1925701\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,9,22]],"date-time":"2026-09-22T05:40:36Z","timestamp":1790055636000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/frai.2026.1925701\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,22]]},"references-count":24,"alternative-id":["10.3389\/frai.2026.1925701"],"URL":"https:\/\/doi.org\/10.3389\/frai.2026.1925701","relation":{},"ISSN":["2624-8212"],"issn-type":[{"value":"2624-8212","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,9,22]]},"article-number":"1925701"}}