Multimodal LLM Merging (Samsung Electronics)

Research from my AI Model Research Internship at Samsung Electronics (Summer 2025) on merging multimodal large language models.

Contributions.

• Designed layer-wise merge strategies that improved language (LLM) benchmark scores while keeping multimodal (LMM) performance above target (>64).
• Built a scalable safetensors merging pipeline supporting multiple models and efficient weight-integration workflows.
• Evaluated merged models with LLMeval and LMMeval, confirming language gains without multimodal regression.

Code is proprietary to Samsung Electronics and not publicly available.

Python · PyTorch · LLM/LMM · safetensors · model merging · evaluation