This paper details the submission of team TIP- RBG-AI for the WMT26 Indic-MT shared task. Addressing severe data sparsity, unstandardized orthographies, and script-level discrepancies across ten Indic languages, we introduce a robust, resource-agnostic cross-lingual framework. Rather than relying on standard downstream parameter fine-tuning, our methodology optimizes zero-shot inference topologies across three massively multilingual machine translation architectures: NLLB-200, MADLAD-400, and IndicTrans2. To systematically resolve vocabulary mismatch and tokenizer fragmentation, we implement a three-tiered preprocessing pipeline comprising baseline untuned direct inference for native scripts, translinear normalization via algorithmic back-transliteration for Romanized textual representations, and phylogenetic script mapping for undocumented, scriptless vernaculars. Empirical evaluations demonstrate that structurally aligning orthographic decoding environments and exploiting cross-lingual family networks substantially enriches semantic fidelity. Without executing a single downstream parameter update, our purely zero-shot framework secured multiple podium placements, including second-place finishes in the Bodo and Manipuri (Meitei Mayek) to English tracks, validating the potential of frontend representation alignment for moderate low-resource translation, while also exposing clear limits at the most extreme end of data scarcity. The code used for reproducing the experiments is publicly available at https: //github.com/rbg-research/EMNLP-2026.

We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651.