In a major advance for low-resource language code translation, researchers have launched BanglaCodeAct, a new framework that converts Bangla instructions into Python code. Powered by the Qwen3-8B model, it achieves 94.0% pass@1 accuracy on the development set and 71.6% on a blind test set. This breakthrough makes programming more accessible to Bangla speakers.

The Story

AI progress often favors English, leaving languages like Bangla behind. BanglaCodeAct changes that by focusing on Bangla-to-Python translation. This project highlights the need for linguistic diversity in AI and pushes for greater inclusivity in tech.

Its impact goes beyond code generation—it could transform educational tools and programming access in Bangla-speaking areas, opening doors for learners and developers alike.

The Context

BanglaCodeAct’s core innovation is its agent-based reasoning system. Instead of relying on heavy fine-tuning, it uses a Thought-Code-Observation loop with multi-agent prompting and iterative self-correction. This lets the system generate, test, and refine code dynamically from Bangla instructions.

This approach mirrors how humans solve problems—by testing and refining until the solution works. As noted by The Verge, this architecture shows how agent-based reasoning can boost code generation accuracy, especially for low-resource languages.

The Qwen3-8B model is key to this success. This open-source multilingual large language model handles complex language tasks well, making it ideal for BanglaCodeAct. The research team—Jahidul Islam, Md Ataullha, and Saiful Azad—benchmarked several small open-source LLMs on the mHumanEval dataset for Bangla NL2Code. Their results, published in an arXiv preprint, set a new benchmark for low-resource language code translation.

Transparency is central to the project. The team has shared their experimental scripts on GitHub, encouraging developers to contribute and improve the framework. This open-source approach ensures ongoing innovation and adaptability.

Key Takeaways

• Inclusivity in AI: BanglaCodeAct addresses the gap for low-resource languages.
• Agent-Based Reasoning: Multi-agent prompting and iterative self-correction improve accuracy.
• Qwen3-8B Model: Its language skills drive the framework’s success.
• Educational Impact: The framework could broaden programming access in Bangla-speaking regions.
• Open Source: Public scripts invite community collaboration and continuous development.

BanglaCodeAct is more than a technical feat. It marks a step toward a tech landscape where language barriers no longer hold back progress.