Breaking01 / AI
MITRA-Swarm lifts Sanskrit translation score by 0.63 points
The researchers reported roughly nine times the inference cost per verse compared with a single model.

Researchers reported that MITRA-Swarm improved Sanskrit translation quality by 0.63 points over a single-model baseline in their evaluation.
The paper was submitted to arXiv on 10 Nov 2026. It describes a system that connects an agent for splitting words fused by sound changes with several translation models. A further stage uses context to restore omitted sentence elements.
The researchers’ published benchmark figures put MITRA-Swarm’s MQM translation quality score at 85.90. The paper specifies an evaluation set of 1,000 verses.
The researchers used separate evaluation models for internal revisions and final reporting. The paper identifies gpt-4o-mini as the final reporting model.
In the researchers’ published benchmark figures, the single-model baseline scored 85.27 on MQM. They reported that MITRA-Swarm’s inference cost per verse was roughly nine times that of the baseline.
The researchers did not disclose the actual inference cost per verse in the abstract.