MITRA-Swarm: Injecting Explicit Linguistic Structure into Foundation Models for Classical Machine Translation
Authors: Kayshav Bhardwaj, Pragun Seth, Kush Bhardwaj, Vijay Nagasamy, Kurt Keutzer
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv: 2611.14675 [cs.CL] · Submitted 10 Nov 2026 · Accepted to NeurIPS 2026 LP4FM Workshop (17 pages)
Overview
Classical Sanskrit poses three challenges that single-call, monolithic models leave unresolved: morphophonemic fusion (Sandhi), dense nominal compounding (Samāsa), and discourse-level zero-anaphora forced by poetic meter. MITRA-Swarm addresses these with a multi-agent architecture that injects explicit linguistic structure into an otherwise monolithic translation pipeline.
In 2026, as foundation models saturate near the 85–86 MQM range on classical-language benchmarks, the remaining gains increasingly depend on structured decomposition rather than raw model scale. MITRA-Swarm is a case study in this shift: it combines specialized agents with independent evaluation to produce small but statistically robust improvements.
Architecture
The swarm comprises four components:
- A Sandhi-splitting agent that segments morphophonemically fused input before translation.
- A parallel array of translators combining the generalist
gemma-4-26B-A4B-itmodel with domain-specializedgemma-2-mitra-ittranslators. - A discourse-aware Ellipsis Restorer that consults corpus-derived register tags, avoiding the invention of active-voice subjects in impersonal legal and philosophical registers.
- Two separated judges: a routing judge (
Jroute), drawn from the same model family as the generators, drives an internal critique loop; an independent reporting judge (Jreport=gpt-4o-mini, with no shared vendor lineage) scores every output exactly once.
Experimental Setup
Evaluation uses a stratified benchmark of 1,000 verses spanning five registers, yielding 23,990 scored verdicts across a pre-registered 12-condition ablation. A preflight screen of a 192,151-entry retrieval index against all 12 evaluation works verifies zero reference leaks across 269 runtime checks.
Results
| Configuration | MQM |
|---|---|
| Full swarm | 85.90 |
| Monolithic baseline | 85.27 |
| Call-count-matched self-refinement baseline | 85.39 |
The full swarm achieves +0.63 MQM over the monolithic baseline (padj = 0.0018, 95% CI [+0.30, +0.97]). Isolated discourse restoration alone accounts for +0.44 MQM (padj = 0.0288) and outperforms the call-count-matched self-refinement baseline by +0.51 MQM (p_adj = 0.0288). Severe per-verse failures (MQM < 80) fall from 9.0% to 7.6%, at roughly 9× the per-verse inference cost of the baseline.
Key Findings
- Judge separation matters.
Jroutescores its own family's outputs 2–7 MQM higher thanJreporton every one of the 12 conditions, empirically motivating the two-judge split. - Discourse restoration is the highest-value component. The Ellipsis Restorer delivers the largest isolated contribution, suggesting that register-aware discourse handling is a productive target for classical-language MT.
- Gains are modest but robust. The authors are candid that these are small improvements on a metric saturating near 85–86, achieved at substantially higher inference cost.
Conclusion
MITRA-Swarm demonstrates that explicit linguistic structure—Sandhi segmentation, compound handling, and discourse-aware ellipsis restoration—can be injected into foundation-model pipelines to yield statistically significant, if modest, quality gains for classical Sanskrit translation. The work also offers a methodological contribution: an evaluation design that separates routing and reporting judges to guard against self-preference bias.
Cite as: arXiv:2611.14675 [cs.CL]
via ArXiv CL+LG