Worked on the nvidia-cosmos/cosmos-rl repository to enhance the Vision-Language Model’s handling of text-only prompts in Qwen2.5-VL. Addressed a bug in the backend by updating the decode_vision_info function to bypass multimedia processing when inputs contain only text, thereby reducing runtime errors and improving stability for text-based workflows. This adjustment broadened the model’s applicability to scenarios where no media is present, supporting more robust downstream integrations and demos. Utilized Python and applied backend development and machine learning skills to ensure the model could reliably process pure text prompts, ultimately improving reliability for text-first use cases across the project.