Elamur Research⌕
☰
home / papers / 2026 / IHLC@LT-EDI 2026: Steering Toward Inclusivity - A Representation Engineering for Gender-Neutral Rewriting

IHLC@LT-EDI 2026: Steering Toward Inclusivity - A Representation Engineering for Gender-Neutral Rewriting

Paper · 2026 · Akhil Rajeev P and Manoj Balaji Jagadeeshan

Summary

The IHLC team describes its entries to the LT-EDI 2026 shared task on gender-inclusive language generation, English track only. For Subtask A, which rewrites gendered sentences into neutral ones, the authors used standard LoRA fine-tuning. Their main contribution is Subtask B, which generates counter-narratives to overtly gender-biased claims without updating any weights. They collect Gemma-3-4B-it hidden states for paired biased and inclusive sentences and apply PCA to the activation differences. During generation, a forward hook adds the leading component, scaled by a coefficient α, to an intermediate layer. Strict one-line prompts and low-temperature decoding with a repetition penalty keep the outputs concise. Scoring used an LLM-as-a-judge with human spot checks. Subtask A scored 80.00% (3rd place), and Subtask B scored 78.12% (6th place). Subtask B outputs were rated polite and coherent but weaker on quality and relevance. A manual error study shows that steering often rewrites more than necessary. The work offers a low-compute alternative to fine-tuning for bias mitigation and documents where it breaks down.

Key points

  • Subtask A (LoRA): GA, GN and QR were all 80%, ranking 3rd of 9 teams.
  • Subtask B (activation steering): PR was 84.89%, CCNC 84.79% and QR 64.68%, for an average of 78.12% and a rank of 6th of 7.
  • Steering setup:
    • The steering vector is the first principal component of last-token difference vectors, injected at layer 16 of Gemma-3-4B-it.
    • The recommended α range is 0.6–1.2; elsewhere the paper cites 0.7–1.5.
    • Excessive α produced looping text and fused words.
  • Manual analysis of 47 outputs found five failure modes:
    • faithfulness–edit trade-off: 53%
    • layer-specific entanglement: 23%
    • residual bias leakage: 11%
    • over-steering instability: 9%
    • repetition/degeneration: 4%
  • Every steered output opened with the same generic template phrase, which departs from the minimal-edit ideal. Code is released as a Jupyter notebook (GitHub: manojbalaji1/IHLC-Gender-Inclusive).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Indian Institute of Technology KharagpurKharagpur · India24 works · 714 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksGender Studies · 52 works

Cite

Cite as Akhil Rajeev P and Manoj Balaji Jagadeeshan. 2026. IHLC@LT-EDI 2026: Steering Toward Inclusivity - A Representation Engineering for Gender-Neutral Rewriting. Proceedings of the Sixth Workshop on Language Technology for Equality, Diversity, Inclusion.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.ltedi-1.19

@inproceedings{p2026ihlc,
  title = {{IHLC@LT-EDI 2026: Steering Toward Inclusivity - A Representation Engineering for Gender-Neutral Rewriting}},
  author = {Akhil Rajeev P and Manoj Balaji Jagadeeshan},
  booktitle = {Proceedings of the Sixth Workshop on Language Technology for Equality, Diversity, Inclusion},
  year = {2026},
  doi = {10.18653/v1/2026.ltedi-1.19},
  url = {https://doi.org/10.18653/v1/2026.ltedi-1.19}
}

References

18 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29.
  2. Helena Bonaldi, Serra Sinem Tekiroğlu, and Marco Guerini. 2022. Human-machine collaboration for generating counter-narratives. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4279–4292. find
  3. Bharathi Raja Chakravarthi, Shunmuga Priya Muthusamy Chinnan, Paul Buitelaar, Meghann L. Drury-Grogan, Miguel Ángel García Cumbreras, Salud María Jiménez Zafra, Thomas Mandl, Sylvia Jaki, Rahul Ponnusamy, Anand Kumar M, Dhanalakshmi V, Bharathi B, Premjith B, Senthil Kumar B, and Sathiyaraj T. 2026.… find
  4. Yi-Ling Chung, Serra Sinem Tekiroğlu, and Marco Guerini. 2021. Towards knowledge-grounded counter narrative generation for hate speech. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 899–914. find
  5. Gemma Team. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  6. Shunmuga Priya Muthusamy Chinnan, Meghann Drury-Grogan, and Bharathi Raja Chakravarthi. 2025. Gender inclusive language generation framework: A reasoning approach with rag and cot. Knowledge-Based Systems, 328:114092. find
  7. Jing Qian, Anna Bethke, Yinyin Liu, Elizabeth Belding, and William Yang Wang. 2019. A benchmark dataset for learning to intervene in online hate speech. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 4755–4764.
  8. Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. arXiv preprint arXiv:1909.01326.
  9. Nishant Subramani, Nithya Suresh, and Matthew E Peters. 2022. Extracting latent steering vectors from pretrained language models. Findings of the Association for Computational Linguistics: ACL 2022, pages 566–581.
  10. Tony Sun, Andrew Gaut, Tang Shirlyn, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. arXiv preprint arXiv:1906.08976.
  11. Serra Sinem Tekiroğlu, Yi-Ling Chung, and Marco Guerini. 2020. Generating counter narratives against online hate speech: Data and strategies. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1177–1190.
  12. Alex Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Maciej MacDiarmid. 2023. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248.
  13. Eva Vanmassenhove, Christian Hardmeier, and Andy Way. 2021. Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender neutral alternatives. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8940–8948. find
  14. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In International Conference on Learning Representations.
  15. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Hao, Zhanghao Wu, Joseph E Ba, Hao Zhuang, Zi Lin, Zhuohan Li, Eric Xing, and 1 others. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36.
  16. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36.
  17. Andy Zou, Long Phan, Sarah Chen, James Campbell, Pengfei Xia, Darren Lin, Minqi Jiang Wang, Danqi Yin, Mantas Woodside, and 1 others. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405.
  18. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359–17372. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/p-2026-ihlc-lt-edi-2026-steering/