IndicSteer: Inference-Time Safety Steering for Indic LLMs

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

IndicSteer introduces an inference-time activation steering method designed to enhance safety in Indic Large Language Models (LLMs). This initial study addresses the critical need for safety controls that account for multilingual variation and culturally grounded harm categories often underrepresented in English-centric resources. IndicSteer applies Contrastive Activation Addition (CAA), computing contrastive directions from safe/unsafe response pairs, across 8 harm categories and 9 Indic language settings. Evaluated using an LLM-as-a-judge protocol covering approximately 12,960 prompt-response pairs, the method shows significant harm reduction without retraining. For Sarvam-1, an 𝛼=12 setting reduced the harmful rate from 73.47% to 41.34% (a 32.13 percentage point or 43.73% relative reduction). OpenHathi Hindi saw its harmful rate decrease monotonically from 85.83% to 27.13% at 𝛼=15, a 58.71 percentage point total reduction.

Key takeaway

For NLP engineers or AI ethicists developing or deploying LLMs for Indic languages, you should consider integrating inference-time safety steering like IndicSteer. This approach offers a powerful, retraining-free method to significantly reduce culturally specific harmful responses, as demonstrated by reductions from 73.47% to 41.34% for Sarvam-1 and 85.83% to 27.13% for OpenHathi Hindi. Implementing such techniques can enhance model trustworthiness and user safety in diverse linguistic contexts.

Key insights

IndicSteer applies inference-time activation steering to enhance safety in Indic LLMs by addressing cultural and linguistic nuances.

Principles

• Safety controls require cultural and multilingual context.
• Contrastive Activation Addition (CAA) enables LLM safety steering.
• Inference-time steering offers harm reduction without retraining.

Method

IndicSteer uses Contrastive Activation Addition (CAA) based on contrastive directions from safe/unsafe response pairs, applied at inference-time across 8 harm categories and 9 Indic languages.

In practice

• Implement CAA for culturally specific safety.
• Utilize LLM-as-a-judge for safety evaluation.
• Tune 𝛼 parameter for optimal harm reduction.

Topics

• Indic LLMs
• Inference-Time Steering
• Contrastive Activation Addition
• LLM Safety
• Multilingual NLP
• Cultural Harm Categories

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models — FineSteer offers fine-grained, adaptive inference-time steering for LLMs, improving safety and truthfulness with minimal utility loss.
• Food Noise & False Safety: A Systematic Evaluation of How LLMs Fail to Adapt to Eating Disorder Queries with Clinician Feedback — LLMs uncritically adapt to eating disorder queries, increasing unsafe response likelihood with specific linguistic cues.
• Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation — Activation steering on early LLM layers improves synthetic data diversity and downstream performance for low-resource languages.
• Beyond English and Evasion: A Human-Annotated Multi-Domain Benchmark for High-Stakes LLM Safety Evaluation in Chinese — Chinese LLM safety requires culturally specific benchmarks to counter linguistic and cultural evasion techniques.
• Week Ending 2.8.2026 — AI research is advancing model robustness, efficiency, and real-world applicability across diverse domains.
• Breaking the Chain: A Causal Analysis of LLM Faithfulness to Intermediate Structures — LLMs' apparent faithfulness to intermediate reasoning structures is fragile; they often fail to update predictions after causal intervention.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.