Elamur Research⌕
☰
home / papers / 2026 / IndicSteer: Inference-Time Safety Steering for Indic LLMs

IndicSteer: Inference-Time Safety Steering for Indic LLMs

Paper · 2026 · Ruhaib Muhammad, Saahas Vijayalakshmi Rajaram and Suriya Priyan Durairaj

Summary

Most LLM safety resources are built on English data. They miss harms rooted in Indian society, such as caste-based abuse or communal incitement, and handle code-mixed or transliterated input poorly. The authors test whether inference-time activation steering, specifically Contrastive Activation Addition (CAA), can reduce such outputs in Indic LLMs without retraining. They create a synthetic, LLM-generated benchmark of harmful prompts covering 8 India-specific harm categories and 9 language settings: eight Indic languages plus Hinglish. For each slice, a steering vector is the difference between mean hidden states for unsafe and safe responses. This vector is scaled by α and subtracted during generation. An IndicBERT-based router selects the vector by language and category, and steering activates only above a risk threshold. Claude Sonnet 4.5 labels each output for harmfulness and coherence. Steering cuts harmful rates sharply for Sarvam-1 and OpenHathi, although some slices get worse and coherence suffers in certain categories. Linear CKA shows tight clustering among Dravidian languages. The study provides a first open testbed for culturally grounded safety steering in Indic languages.

Key points

  • Sarvam-1 results: at α = 12, across 72 matched slices, the harmful rate fell from 73.47% to 41.34%, a drop of 32.13 pp (43.73% relative). The sweep covered 12,960 judged instances (30 prompts × 9 languages × 6 α values × 8 categories) and used about 40 GPU-hours on NVIDIA T4 in fp32.
  • Variation by category:
    • Child safety and code-mixed toxicity responded best.
    • Gender-based violence stayed at 70.00% in every language.
    • Anti-minority sentiment got worse in Malayalam and Telugu.
    • Coherence ranged from 92.59% (caste discrimination) to 33.33% (gender-based violence).
    • In total, six slices showed higher harm after steering.
  • OpenHathi (Hindi): the harmful rate dropped steadily from 85.83% to 27.13% at α = 15, with no sign of saturation. The largest single step, 35.83 pp, came between baseline and α = 1.
  • CKA analysis:
    • Mean cross-family similarity was 0.609 for Sarvam-1 and 0.821 for Krutrim-2-Instruct.
    • Dravidian pairs were very cohesive (kn–ml: 1.00 and 0.99).
    • Response-level evaluation of Krutrim-2-Instruct is still pending.
  • Limitations and release:
    • Prompts have not been checked by native speakers.
    • Harmfulness was judged only by an LLM, with no human annotation.
    • Code, steering vectors and prompts are released in the GitHub repository SafeSteer_IN.

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

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSafety Research · 543 worksEngineering · 133 worksSafety, Risk, Reliability and Quality · 10 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksBias, fairness & safety · 517 works

Cite

Cite as Ruhaib Muhammad, Saahas Vijayalakshmi Rajaram and Suriya Priyan Durairaj. 2026. IndicSteer: Inference-Time Safety Steering for Indic LLMs. Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.stereacult-1.12

@inproceedings{muhammad2026indicsteer,
  title = {{IndicSteer: Inference-Time Safety Steering for Indic LLMs}},
  author = {Ruhaib Muhammad, Saahas Vijayalakshmi Rajaram and Suriya Priyan Durairaj},
  booktitle = {Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.stereacult-1.12},
  url = {https://doi.org/10.18653/v1/2026.stereacult-1.12}
}

References

13 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Somnath Banerjee, Sayan Layek, Pratyush Chatterjee, Animesh Mukherjee, and Rima Hazra. 2025. Soteria: Language-specific functional parameter steering for multilingual safety alignment. Preprint, arXiv:2502.11244. find
  2. Valerio Basile, Cristina Bosco, Elisabetta Fersini, Debora Nozza, Viviana Patti, Francisco Manuel Rangel Pardo, Paolo Rosso, and Manuela Sanguinetti. 2019. SemEval-2019 task 5: Multilingual detection of hate speech against immigrants and women in Twitter. In Proceedings of the 13th International Wor…
  3. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. Preprint, arXiv:2203.09509.
  4. Mohammed Khan, Priyam Mehta, Ananth Sankar, Umashankar Kumaravelan, Sumanth Doddapaneni, Suriyaprasaad B, Varun G, Sparsh Jain, Anoop Kunchukuttan, Pratyush Kumar, Raj Dabre, and Mitesh Khapra. 2024. Indicllmsuite: A blueprint for creating pre-training and fine-tuning datasets for indian languages.…
  5. Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. 2024. Llms-as-judges: A comprehensive survey on llm-based evaluation methods. Preprint, arXiv:2412.05579.
  6. Rhitabrat Pokharel, Ameeta Agrawal, and Tanay Nagar. 2026. Cross-lingual activation steering for multilingual language models. Preprint, arXiv:2601.16390. — [arXiv:2601.16390](https://arxiv.org/abs/2601.16390) find
  7. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering language models with activation engineering. Preprint, arXiv:2308.10248.
  8. Hui Wei, Shenghua He, Tian Xia, Fei Liu, Andy Wong, Jingyang Lin, and Mei Han. 2025. Systematic evaluation of llm-as-a-judge in llm alignment tasks: Explainable metrics and diverse prompt templates. Preprint, arXiv:2408.13006. — [arXiv:2408.13006](https://arxiv.org/abs/2408.13006) find
  9. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, and 2 others. 2025. Representa…
  10. Sumanth Doddapaneni, Rahul Aralikatte, Gowtham Ramesh, Shreya Goyal, Mitesh M. Khapra, Anoop Kunchukuttan, and Pratyush Kumar. 2023. Towards leaving no indic language behind: Building monolingual corpora, benchmark and models for indic languages. Preprint, arXiv:2212.05409.
  11. Jay Gala, Thanmay Jayakumar, Jaavid Aktar Husain, Aswanth Kumar M, Mohammed Safi Ur Rahman Khan, Diptesh Kanojia, Ratish Puduppully, Mitesh M. Khapra, Raj Dabre, Rudra Murthy, and Anoop Kunchukuttan. 2024. Airavata: Introducing hindi instruction-tuned llm. Preprint, arXiv:2401.15006.
  12. Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. 2022. *Figure 1: Online runtime pipeline for IndicSteer: prompt routing, risk prediction, and steering-hook application.* *Figure 2: Offline pipeline: activation extraction, contrastive steering-vector comput… find
  13. ∆ (pp) Child safety Code-mixed toxicity Financial scam Communal/rel. hate Caste discrimination Political misinfo. Gender-based violence Anti-minority sent. 96.30 79.63 94.07 84.81 50.74 49.26 82.59 50.37 33.70 24.07 55.19 46.67 29.26 31.48 70.00 40.37 62.59 55.56 38.89 38.15 21.48 17.78 12.59 10.00… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/muhammad-2026-indicsteer-inference-time-safety-steering/