Vyoman Jain

Large language models’ substantial memory and compute requirements hinder deployment on individual resource-constrained devices. HyperCluster is a framework for fully decentralized collaborative inference over wireless networks, contributing a ring-based pipelined inference protocol in which nodes deterministically self-organize into a computational ring by device capability and pass intermediate states directly between peers, and a generalizable model sharding methodology built on HuggingFace Transformers that automatically partitions any dense LLM across heterogeneous devices according to available compute and memory.

ReviewArena is a large-scale peer-review dataset constructed from OpenReview venues (NeurIPS, ICLR, ICML, CoRL, COLM, EMNLP, TMLR) comprising 51,529 papers and 196,099 reviews across fourteen review fields, including full PDFs, reviewer scores and text, rebuttals, meta-reviews, and final decisions. We further derive ReviewArena-Eval, a 1,002-paper benchmark spanning six conferences with aligned, venue-specific evaluation protocols. Baselines with six open-weight LLMs show that current models are miscalibrated, compress rating scales, and weakly distinguish accepted from rejected papers.

Existing document retrieval systems are overwhelmingly English-centric. M3DR uses synthetic multilingual document data and generalizes across vision-language architectures to enable semantic search across languages, employing contrastive training to learn unified representations for text and document images. Evaluated on 22 typologically diverse languages across monolingual, multilingual, and mixed-language settings, the resulting NetraEmbed and ColNetraEmbed models achieve state-of-the-art performance with roughly 150% relative improvements on cross-lingual retrieval, supporting both dense-vector and token-level multi-vector retrieval.

Nayana is a synthetically generated dataset for advancing document-centric vision-language models across tasks and languages. It comprises three interconnected subsets: a base dataset of 3 million document images with hierarchical annotations covering layout, textual content, reading order, and element relationships; a multilingual variant spanning 22 languages that preserves original layout while translating text through contextual models; and an information-retrieval subset of roughly 250,000 image-query pairs per language. Together these enable training for layout detection, equation recognition, captioning, markdown conversion, multilingual OCR, and document retrieval.

Modern vision-language models remain constrained by the scarcity of training data in non-English languages. Nayana is a scalable framework for adapting VLMs to low-resource languages through a layout-aware synthetic data generation pipeline combined with parameter-efficient adaptation, enabling existing models to learn new languages from purely synthetic data. Using LoRA, the framework demonstrates capability across ten Indic languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu.

Understanding complex scientific concepts in dense research papers is a significant challenge. Manimator uses large language models to transform research papers and natural language prompts into explanatory animations via the Manim engine: an LLM interprets the input text or PDF to produce a structured scene description outlining key concepts, formulas, and visual elements, and a second code-focused LLM translates that description into executable Manim Python code. The open-source system serves as an educational tool for rapidly creating visual explanations of STEM topics.