International Journal of Scientific Research and Engineering Development

Neural machine translation (NMT) research pipelines involve numerous loosely coupled stages — data curation, preprocessing, model training, hyperparameter search, evaluation, error diagnosis, and reporting — that are typically orchestrated manually or through rigid, script-based automation. This rigidity slows iteration, obscures provenance, and makes it difficult to reuse knowledge across experiments. We propose MANTIS, a multi-agent architecture in which specialized large language model (LLM) driven agents collaborate under a central orchestrator to plan, execute, evaluate, and report NMT training runs. MANTIS combines a supervisor agent for task decomposition and routing, a shared three-tier memory (short-term, long-term, and vector memory for retrieval-augmented generation), and a closed feedback loop that lets evaluation and error-analysis signals trigger automatic retraining or hyperparameter adjustment. We detail the agent roles, communication protocol, scheduling and conflictresolution algorithms, and a reference implementation built on open frameworks. We describe an experimental protocol using WMT and FLORES200 benchmark corpora (Costa-jussà et al., 2022), propose an evaluation methodology spanning translation quality, task completion rate, latency, cost, and scalability, and present an ablation design for isolating the contribution of each agent. The architecture targets a gap between singleagent LLM pipelines, which do not scale across concurrent experiments, and hand-orchestrated multi-tool pipelines, which do not adapt or reason about failures.

Jahnavi Somaraju, Nandyala Mamatha, "MANTIS: A Multi-Agent Architecture for Training and Benchmarking Neural Machine Translation Models" International Journal of Scientific Research and Engineering Development, V9(4): Page(328-335) May-June 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.