Reinforcement learning lifts Indic model's math but cuts cross-lingual question answering
A reported evaluation of a 30B-A3B mixture-of-experts model before and after GRPO reinforcement learning shows large MATH gains and mostly Indic improvements, alongside notable regressions.
In plain words
Researchers applied reinforcement learning, a training method that rewards desired behaviour, to a language model serving English and eight Indic languages. Scores improved on math and most Indic benchmarks, but the model answered far fewer questions in Indic languages on one test, shifting to English instead.
What was announced
The authors ran reinforcement learning with GRPO across four environments, starting from a supervised fine-tuned 30B-A3B mixture-of-experts model covering English and eight Indic languages, then compared before and after checkpoints across 34 benchmarks in a single evaluation campaign. MATH improved by 36.07 points and most Indic benchmarks also rose. However, XQuAD-Indic F1 fell 19.32 points as English answers rose from 8.7% to 91.7%, and IndicQA dropped 8.64, while English XQuAD gained 11.71 F1.
A second RL run from the same supervised initialization added an Indic environment plus language and repetition rewards. Compared with the released model, it improved passage-answering and instruction-following benchmarks and preserved MATH performance, but lost ground on reasoning and coding benchmarks. The authors note that several training choices changed together, so no single reward term's effect can be isolated, and adding a language reward did not ensure use of the requested script.
What it does not show
- The evidence is the paper's abstract only; full methodology, data and results are not covered.
- The authors state they cannot isolate the effect of any individual reward term because several training choices changed together.
- Script compliance remains unresolved according to the authors, so requested-script usage needs direct evaluation.
For security and IT teams
- Scope
- The findings cover one 30B-A3B mixture-of-experts model evaluated across 34 benchmarks in a single campaign, not a broad survey of RL outcomes.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
17 Nov 2026 · Updated wording and source-support review.
First published 13 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.