Resume

Publications

1. EMNLP 2025 (Main Conference)

   Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval-Augmented Generation Across Learning Styles

   Debdeep Sanyal, Agniva Maiti, Umakanta Maharana, Dhruv Kumar, Ankur Mali, C. Lee Giles, Murari Mandal

   Multi-agent teacher and student framework where genetic algorithms evolve teaching strategies across 100 heterogeneous student agents, plus Persona-RAG, a retrieval module reaching 88% accuracy on conceptual questions (HyDE 85%, vanilla RAG 74%), with human evaluation by 20 raters (Krippendorff's alpha 0.78-0.84).

2. WMT 2026

   SCE-KIIT: Fine-Tuning NLLB-200 for English–Nagamese Translation

   Agniva Maiti, Aarsh Muley, Sovan Kumar Sahoo

   Curated 2,280 English–Nagamese sentence pairs and fine-tuned NLLB-200 (600M) with surrogate language tokens; BLEU on our held-out split rose from 0.40 to 40.02.

3. WMT 2026

   SCE-KIIT: KokLLaMA: Cross-Task Adaptation for Low-Resource English–Kokborok Translation

   Agniva Maiti, Aarsh Muley, Sovan Kumar Sahoo

   QLoRA-tuned Llama-3.2-3B-Instruct on 11,428 Kokborok instruction pairs, with no parallel training data.

4. GlobalSouthML at ICML 2026 (poster)

   Bridging the Visual Gap: Resource-Efficient VLM Adaptation for Meitei Mayek using Synthetic Multi-Modal Mixtures

   Agniva Maiti, Aditya Tiwari, Dwarikanath Mahapatra, et al.

   Adapts PaliGemma-3B to the Meitei Mayek script with QLoRA and a five-part synthetic multimodal data mixture, raising ANLS from 8.62% to 21.62% and doubling VQA exact match from 8% to 16%, with 76% POPE accuracy.

5. GlobalSouthML at ICML 2026 (poster)

   Evaluating Indic Literary Creativity In LLMs through the Lens of the Shastras

   Agniva Maiti, Aarsh Muley, Aditya Tiwari, Dwarikanath Mahapatra, et al.

   Benchmarks 8 LLMs on Hindi and Bengali creative writing using Torrance-test dimensions and 6 classical rhetorical devices, and identifies an "Alankara over-stuffing" bias in LLM-as-a-judge scoring.

6. arXiv preprint, 2025

   NagaNLP: Bootstrapping NLP for Low-Resource Nagamese Creole with Human-in-the-Loop Synthetic Data

   Agniva Maiti, Manya Pandey, Murari Mandal

   10,018-pair conversational corpus and annotated POS/NER data (Cohen's κ 0.92 POS, 0.88 NER), a fine-tuned XLM-R reaching 93.81% POS accuracy and 0.75 NER F1, and NagaLLaMA (Llama-3.2-3B) with perplexity 3.85 vs 96.76 few-shot.

Education

Internships & Experience

Leadership & Volunteering

Selected Projects

Technical Skills

Languages

Certifications

Organizations