Multilingual machine translation for endangered and severely underrepresented languages faces two interacting challenges: scarce parallel data and uneven subword vocabulary allocation across writing systems. We study these factors for seven Indian tribal languages, Gondi, Mundari, Bhili, Santali, Kui, Garo, and Kokborok, spanning four language families and four scripts. We release a speaker-curated Hindi–English–X tri-parallel corpus of approximately 20k sentences per language across six community-relevant domains, with all development and test references human-curated or independently speaker-verified. We evaluate prompted LLMs, fine-tuned IndicTrans2 and NLLB-600M, and byte-level ByT5 under corpus-Native, Devanagari-Unified, and Latin-Unified script conditions. Fine-tuning improves CHRF++ by 16-25 points over prompting, but script unification is strongly model-dependent: the same Latin conversion improves Kui under NLLB-600M while substantially degrading performance under IndicTrans2. We introduce Script Tokenization Parity and Script Sensitivity to analyse script–tokenizer interaction; byte-level modelling reduces Script Sensitivity from 14.2 to 2.1 CHRF++, localizing most cross-script variance in learned vocabulary allocation. Our findings show that script choice in endangered-language MT is not routine preprocessing, but a community-sensitive modelling decision that directly shapes linguistic representation, translation quality, and equitable digital access.

India’s tribal (*Janjatiya/Adivasi*) languages remain absent from public ASR evaluation, despite their importance for oral history, ecological knowledge, and cultural continuity. We present **AdiDhwani**, a preservation-oriented ASR benchmark for **Bhili** (ISO 639-3: ‘bhb‘), India’s largest non-scheduled language, and **Mundari** (ISO 639-3: ‘unr‘), a Vulnerable North Munda language. The benchmark comprises 56.5 hours of community-validated Bhili speech and 11.7 hours of Mundari speech with fixed evaluation splits and reproducible baselines. We evaluate multilingual ASR under zero-shot transfer, target-language fine-tuning, donor-language initialization, synthetic supervision, and KenLM-assisted decoding. Target-language supervision substantially improves WER, yet lower edit distance does not always reflect better documentation fidelity: models replace attested Bhili forms with Hindi cognates or delete grammatically meaningful Mundari suffixes while still achieving competitive scores. We introduce **Hindi-Normalization Bias (HNB)** and its alignment-constrained variant **(HNBA)**, **Morphological Fragmentation Rate (MFR)**, and **Suffix Deletion Rate (SDR)** as preservation-oriented diagnostics beyond WER. Human transcriber baselines show that these gaps are not fixed limits. **AdiDhwani** provides the first controlled ASR evaluation framework for these two under-documented Indian languages and supports language-preserving speech recognition research in endangered-language settings.

Bridging molecular structures and natural language is essential for controllable design. Autoregressive models struggle with long-range dependencies, while standard diffusion processes apply uniform corruption across positions, which can distort structurally informative tokens. We present BiMol-Diff, a unified diffusion framework for the paired tasks of text-conditioned molecule generation and molecule captioning. Our key component is a token-aware noise schedule that assigns position-dependent corruption based on token recovery difficulty, preserving harder-to-recover substructures during the forward process. On ChEBI-20 and M3-20M, BiMol-Diff improves molecule reconstruction with a 15.4% relative gain in Exact Match and achieves strong captioning results, attaining best BLEU and BERTScore among compared baselines. These results indicate token-aware noising improves fidelity in molecular structure-language modeling

The linguistic diversity of India poses significant machine translation challenges, especially for underrepresented tribal languages like Bhili, which lack high-quality linguistic resources. This paper addresses the gap by introducing Bhili-Hindi-English Parallel Corpus (BHEPC), the first and largest parallel corpus worldwide comprising 110,000 meticulously curated sentences across Bhili, Hindi, and English. The corpus was created with the assistance of expert human translators. BHEPC spans critical domains such as education, administration, and news, establishing a valuable benchmark for research in low resource machine translation. To establish a comprehensive Bhili Machine Translation benchmark, we evaluated a wide range of proprietary and open-source Multilingual Large Language Models (MLLMs) on bidirectional translation tasks between English/Hindi and Bhili. Comprehensive evaluation demonstrates that the fine-tuned NLLB-200 distilled 600M variant model outperforms others, highlighting the potential of multilingual models in low resource scenarios. Furthermore, we investigated the generative translation capabilities of multilingual LLMs on BHEPC using in-context learning, assessing performance under cross-domain generalization and quantifying distributional divergence. This work bridges a critical resource gap and promotes inclusive natural language processing technologies for low-resource and marginalized languages globally.