@inproceedings{muhammad-etal-2026-indicsteer,
    title = "{I}ndic{S}teer: Inference-Time Safety Steering for {I}ndic {LLM}s",
    author = "Muhammad, Ruhaib  and
      Rajaram, Saahas Vijayalakshmi  and
      Durairaj, Suriya Priyan",
    editor = "Ma, Weicheng  and
      Vosoughi, Soroush  and
      Gillani, Nabeel  and
      Coto-Solano, Rolando",
    booktitle = "Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies ({S}tere{AC}u{LT} 2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.stereacult-1.12/",
    pages = "126--136",
    ISBN = "979-8-89176-408-8",
    abstract = "Safety controls for Indic language generation must account for multilingual variation and culturally grounded harm categories that are underrepresented in English-centric resources. We present IndicSteer, an initial study of inference-time activation steering for safety across 8 harm categories and 9 Indic language settings, based on contrastive directions computed from safe/unsafe response pairs. To the best of our knowledge, this is the first application of Contrastive Activation Addition (CAA) to Indic LLMs. Evaluation uses a structured LLM-as-a-judge protocol with strict isolation by category and alpha, covering $\approx$12,960 prompt-response pairs. We report harmful-response and coherence metrics for Sarvam-1 and OpenHathi (Hindi track), and present cross-lingual representation structure via linear CKA for Sarvam-1 and Krutrim-2-Instruct. On matched slices, Sarvam-1 at $\alpha=12$ reduces harmful rate from 73.47{\%} to 41.34{\%} (32.13 pp; 43.73{\%} relative) with no additional retraining. For OpenHathi Hindi, harmful rate falls monotonically from 85.83{\%} (baseline) to 27.13{\%} at $\alpha=15$, a 58.71 pp total reduction."
}