During October 2025, Haozekun contributed backend development and machine learning expertise to the nvidia-cosmos/cosmos-rl repository, focusing on improving the Vision-Language Model’s handling of text-only prompts. He addressed a bug in the decode_vision_info function, enabling the model to gracefully skip multimedia processing when inputs contained only text. This Python-based solution reduced runtime errors and stabilized text-first workflows, allowing the Qwen2.5-VL model to process prompts without images or videos. By enhancing support for pure text prompts, Haozekun improved the model’s reliability and integration readiness, demonstrating depth in backend engineering and a practical approach to machine learning system robustness.