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

Conference: ACL 2026 Findings
arXiv: 2601.06767
Code: Website
Area: Low-Resource Language Reasoning / Mathematical Reasoning
Keywords: Bengali Mathematical Reasoning, Curriculum Learning, GRPO Cold Start, Difficulty-Awareness, Low-Resource Languages

TL;DR

This paper introduces GanitLLM, the first model to perform mathematical reasoning genuinely in Bengali (rather than through translation or reasoning in English). Through the construction of Ganit, a difficulty-annotated Bengali mathematics dataset, and the proposal of Curriculum-GRPO to address the cold start problem in low-resource GRPO training, the 4B model achieves an 8-percentage-point accuracy gain on Bn-MGSM, while increasing Bengali reasoning tokens from 14% to 88%.

Background & Motivation

Background: LLMs have achieved significant progress in mathematical reasoning in high-resource languages like English (e.g., DeepSeek-R1, OpenAI o1), where RL methods such as GRPO have proven effective. However, reasoning in low-resource languages lags behind. Despite Bengali being the seventh most spoken language globally, existing LLMs either reason in English and translate the answer or fail entirely on Bengali math problems.

Limitations of Prior Work: (1) Even when explicitly prompted, existing LLMs tend to reason in English before outputting Bengali answers, which offers poor interpretability for native users; (2) Standard GRPO training encounters a "cold start problem" in low-resource settings—the policy model lacks sufficient target language capability to generate any correct solutions in a rollout group, resulting in zero rewards, zero gradients, and ineffective training; (3) Existing Bengali math datasets vary in quality and lack difficulty annotations and systematic quality filtering.

Key Challenge: GRPO requires at least some correct answers within a rollout group to compute effective advantage values. However, low-resource models often fail completely to generate correct solutions for difficult problems—a "chicken and egg" problem where the model needs basic competence to learn further.

Goal: To build a high-quality, difficulty-annotated Bengali mathematical dataset and design a training strategy that solves the cold start problem, enabling models to reason truly in Bengali.

Key Insight: The problem is decomposed into three steps: (1) Data: Constructing a quality-filtered and difficulty-annotated dataset; (2) SFT: Teaching the model how to reason step-by-step in Bengali (prioritizing language over correctness); (3) GRPO: Utilizing a curriculum learning strategy to train the model progressively from easy to hard.

Core Idea: Use Curriculum-GRPO to arrange training data by difficulty, ensuring the model can generate partially correct answers at each stage to obtain valid gradients and avoid cold start.

Method

Overall Architecture

The training consists of two stages: (1) SFT Stage: Teaching the model step-by-step reasoning in Bengali using CoT-SFT data, focusing on language consistency; (2) Curriculum-GRPO Stage: Training with GRPO on RL data sorted by difficulty, starting with simple problems. The Ganit dataset is derived from ~1.5M raw samples through multi-stage filtering and difficulty annotation; these difficulty signals guide the curriculum scheduling. The GRPO optimization is controlled by a three-dimensional reward function targeting both correctness and Bengali language usage.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Raw Corpus<br/>9 Public Datasets ~1.5M Samples"]
    subgraph GANIT["Difficulty-Aware Dataset (Ganit)"]
        direction TB
        B["Quality Filtering<br/>Accuracy >95% + Rules + De-duplication"]
        C["Difficulty Annotation<br/>Qwen3-32B Solves 32x per Sample<br/>Classified by pass@k (Easy/Medium/Hard/Olympiad)"]
        D["Benchmark De-contamination"]
        B --> C --> D
    end
    A --> GANIT
    GANIT -->|CoT-SFT Data| E["SFT Stage (Scaffold)<br/>Teach Bengali Step-by-Step Reasoning"]
    E --> F["Curriculum-GRPO Strategy<br/>60/40 Bucket Sampling, Easy to Hard"]
    G["3D Reward Function<br/>R = Format + Correctness + Bengali %"] --> F
    F --> H["GanitLLM<br/>True Bengali Math Reasoning Model"]

Key Designs

1. Difficulty-Aware Dataset (Ganit): Refining Raw Corpus into a Graduated Training Set

Existing Bengali math datasets vary in quality, and standard benchmarks (Bn-MGSM / Bn-MSVAMP) are too simple for modern LLMs (77-86% categorized as "Easy"), failing to train high-level capabilities. Ganit uses a multi-stage pipeline: starting with ~1.5M samples from 9 sources, it retains only datasets with accuracy >95% (~1.1M). It applies rule-based filtering (numerical answers, >99% Bengali characters, excluding MCQs) and deduplication (Fuzzy + MinHash). Most importantly, it performs difficulty annotation: Qwen3-32B generates 32 solutions per problem, classifying them into Easy, Medium, Hard, or Olympiad based on pass@k. Finally, decontamination is performed against evaluation benchmarks. This creates a dataset with continuous difficulty signals to serve as a curriculum scale.

2. Curriculum-GRPO Strategy: Avoiding Cold Start via Difficulty-Mixed Sampling

Standard GRPO requires rollout groups to contain correct answers to compute advantages. A weak Bengali model failing 100% of hard problems results in zero gradients. Curriculum-GRPO utilizes the fine-grained difficulty signals (1-32 correct generations): problems are placed in buckets. Each training batch samples 60% from the current difficulty bucket and 40% from the remaining 31 buckets (3 samples each). The curriculum advances from the easiest bucket to the hardest. This 60/40 design ensures the model gains correct experience on simple tasks for non-zero gradients, while the 40% mixed samples prevent overfitting and forgetting.

3. Three-Dimensional Reward Function: Rewarding Bengali Reasoning Specifically

Traditional GRPO focuses only on answer correctness, which leads models to "shortcut"—reasoning in English and only translating the final answer. This work adds rewards for target language consistency:

\[R = R_{format} + R_{correctness} + R_{bengali}\]

Where \(R_{format} \in \{0,1\}\) checks format compliance, \(R_{correctness} \in \{0,1,2\}\) rewards correctness (with a bonus point for Bengali responses), and \(R_{bengali} \in \{0,1\}\) rewards instances where Bengali tokens exceed 80% of the reasoning process. The 80% threshold allows for language-independent mathematical symbols and formulas. This explicitly optimizes for "thinking in the target language," increasing the Bengali reasoning ratio from 14% to 88%.

Loss & Training

The SFT stage uses standard cross-entropy loss. The GRPO stage uses standard GRPO loss with an ultra-long sequence filter and token-level penalties. The base model is Qwen3-4B.

Key Experimental Results

Main Results

Model	Bn-MGSM	Bn-MSVAMP	Bengali %	Avg Length (Words)
Qwen3-4B (Base)	69	78	14%	943
+ SFT only	73	81	82%	210
+ Curriculum-GRPO	77	84	88%	193
Qwen3-8B	76	83	18%	876
GPT-5-mini	82	88	45%	520

Ablation Study

Training Strategy	Bn-MGSM	Cold Start Rate
Shuffled GRPO	72	35%
Fully Sorted (Easy→Hard)	74	12%
Curriculum-GRPO (60/40)	77	5%

Key Findings

• Curriculum-GRPO reduces the cold start rate from 35% to 5%, proving essential for low-resource RL.
• The SFT stage is critical for language switching; GRPO rewards alone struggle to flip the reasoning language from English to Bengali.
• The 4B model achieves accuracy comparable to an 8B base model while reducing reasoning tokens by 79.5%.
• The Ganit-Dev set offers a more balanced difficulty distribution (21-29% per level) compared to standard sets (77-86% Easy), providing more discriminative evaluation.

Highlights & Insights

• Identifying and solving the "cold start problem" provides a reference for RL training across all low-resource languages.
• The 3D reward function design is elegant—optimizing for correctness while explicitly incentivizing reasoning in the target language.
• The 80% Bengali threshold demonstrates domain awareness by accounting for language-agnostic mathematical symbols.

Limitations & Future Work

• Validated only on 4B models; the cold start problem might manifest differently at larger scales.
• The 60/40 curriculum ratio is empirically tuned and lacks theoretical derivation.
• Difficulty labels depend on the capabilities of Qwen3-32B and may need updates as model capabilities evolve.
• Only validated on mathematical reasoning; applicability to logical or commonsense reasoning remains unknown.

Related Work & Insights

• vs. Confucius3-Math: Chinese K-12 math models use standard RL; GanitLLM must resolve cold start issues specific to smaller Bengali data scales.
• vs. mCoT: mCoT uses multilingual CoT tuning but does not enforce target language reasoning; GanitLLM achieves 88% native reasoning through specific rewards.
• vs. MathOctopus: Uses parallel corpora but reasoning remains in English; GanitLLM achieves true native-language reasoning.

Rating

• Novelty: ⭐⭐⭐⭐ The identification of the cold start problem and Curriculum-GRPO are novel contributions.
• Experimental Thoroughness: ⭐⭐⭐⭐ Detailed ablation, dataset quality analysis, and language ratio statistics.
• Writing Quality: ⭐⭐⭐⭐ Clear problem definition and detailed data construction process.
• Value: ⭐⭐⭐⭐ Provides a practical solution for RL training in low-resource languages.

Related Papers

• [ICLR 2026] Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation
• [ACL 2026] Budget-Aware Anytime Reasoning with LLM-Synthesized Preference Data
• [ACL 2026] DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning
• [ACL 2026] Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding (CoRD)
• [NeurIPS 2025] Curriculum Abductive Learning