New research from Aarjan Chaudhary, Binayak Mukherjeea, Moitreya Nandi, and Arindam Panja addresses tokenization inefficiencies for Indic languages. A regex fix cuts Nepali tokens in Llama-3/Qwen3 by 54%, while AdaptTok-Indic lowers token fertility for Bengali and Hindi, improving LLM performance and reducing inference costs.
New research from arXiv details how multilingual AI models struggle with cross-script entity matching, showing a Qwen2.5-7B accuracy drop from 35.1% to 8.5% for Devanagari names. Another paper introduces a 15.45-hour Nepali-English code-switched ASR benchmark, highlighting the role of synthetic data.
New research reveals LLMs show evaluation awareness in Indic languages, alongside a parameter-efficient translation framework and critical findings on pruning calibration for multilingual models.
New research from Stephen Fernandes, Narayan Choudhary, Gaurab Baral, Mimansak Nepal, and Grishma Basnet unveils Anuvani-LID for 22 Indian languages and NEPSEE, a double-annotated Nepali sentiment benchmark, highlighting gaps in general AI models.
New deep tech research showcases two approaches to Indian language AI: in-place tokenizer expansion reducing token-per-word by 38.6% for eight Indic languages, and a 2.5B-parameter Tamil-first model outperforming Sarvam-1 in translation.
New research from Indian institutions and authors introduces RupeeBias for LLM demographic bias, IndicFDB for full-duplex voice agents in ten languages, and word similarity datasets for six Indian languages.
A replication study found an extractive summarization method performed worse than a simple baseline for Hindi. Separately, COILD, a new 1.16 million sentence-pair parallel corpus, was released for Indian language machine translation across 20 language pairs and eight domains.
Recent research introduces Vimarsha, a 100-hour ASR benchmark for 22 Indian languages, and new diagnostics for code-mixed Hinglish to detect misogyny and LLM hallucinations, pushing AI evaluation beyond clean datasets.
A new grapheme-aware tokenizer for Tamil, a dataset and model for detecting speech turn boundaries in Tamil telephone calls, and a system for creating…
New research introduces VakyArth, the first pragmatic benchmark for Indic languages (Hindi, Punjabi, Tamil, Malayalam) to evaluate LLM contextual understanding. Separately, SpeakPay demonstrates a 67.2% WER reduction for Nepali financial speech recognition, improving transaction success rate twenty-fold.