GanitLLM: Difficulty-Aware Bengali Mathematical Reasoning through Curriculum-GRPO

TL;DR

GanitLLM tackles the challenge of Bengali mathematical reasoning by grounding reasoning in Bengali and addressing reward sparsity through Curriculum-GRPO. The authors construct Ganit, a difficulty-tagged Bengali math corpus with rigorous filtering, deduplication, and decontamination, and train using a two-stage process (SFT followed by Curriculum-GRPO) to overcome cold-start issues in low-resource settings. Empirical results show that a 4B GanitLLM model can match or exceed larger baselines on Bengali math benchmarks, while producing predominantly Bengali reasoning and significantly shorter, more interpretable solutions. This work advances practical Bengali reasoning capabilities in low-resource contexts and provides a scalable training framework for language-grounded math reasoning in other underrepresented languages.

Abstract

We present a Bengali mathematical reasoning model called GanitLLM (named after the Bangla word for mathematics, "Ganit"), together with a new difficulty-aware Bengali math corpus and a curriculum-based GRPO pipeline. Bengali is one of the world's most widely spoken languages, yet existing LLMs either reason in English and then translate, or simply fail on multi-step Bengali math, in part because reinforcement learning recipes are tuned for high-resource languages and collapse under reward sparsity in low-resource settings. To address this, we construct Ganit, a rigorously filtered and decontaminated Bengali math dataset with automatic difficulty tags derived from the pass@k of a strong evaluator model. Building on this dataset, we propose Curriculum-GRPO, which combines multi-stage training (SFT + GRPO) with difficulty-aware sampling and verifiable rewards for format, numerical correctness, and Bengali reasoning. On Bn-MGSM and Bn-MSVAMP, GanitLLM-4B improves over its Qwen3-4B base by +8 and +7 accuracy points, respectively, while increasing the percentage of Bengali reasoning tokens from 14% to over 88% and reducing average solution length from 943 to 193 words.

Figures (4)

• Figure 1: Overview of our approach for a Bengali mathematical reasoning model. (Left) Current models reason in English even for Bengali questions, resulting in reduced interpretability for native speakers. (Center) Our solution combines the Ganit dataset with SFT to ground reasoning in Bengali, followed by Curriculum-GRPO for efficient RL training. (Right) Our approach achieves native Bengali reasoning (88% Bengali vs. 14%), reduces reasoning tokens by 79%, and improves accuracy from 69 to 76.
• Figure 2: Overview of the Ganit construction pipeline. Starting from $\sim$1.5M Bengali math problems, we apply multi-stage quality filtration, verification, deduplication, and decontamination to obtain Ganit-Train (SFT and RLVR) and Ganit-Dev.
• Figure 3: Evaluation curves comparing GRPO and Curriculum-GRPO on MGSM and MSVAMP benchmarks. Checkpoint-wise accuracy demonstrates that while both methods achieve comparable performance on the easier MGSM dataset (left), CGRPO substantially outperforms traditional GRPO on the harder MSVAMP dataset (right), where the cold-start problem causes GRPO to stagnate.
• Figure 4: Qualitative comparison of training configurations on an Olympiad-level problem from Ganit-Dev.