Researchers have developed a new method called Cross-Scale Directional Parameter Injection (CDPI) to analyze how knowledge is transferred when combining different multimodal large language models (MLLMs). Their experiments, using Qwen3-VL model pairs across twelve benchmarks, indicate that this fusion process primarily enhances reasoning capabilities, especially high-level reasoning, while perception abilities remain largely unchanged. The study suggests that effective knowledge transfer occurs mainly in the language model component and is most pronounced when the fusion involves a small ratio of parameters, recasting the process as selective reasoning transfer rather than broad capability inheritance. AI
IMPACT Provides a new analytical tool to understand how multimodal models learn from each other, potentially guiding future fusion strategies.
RANK_REASON Academic paper detailing a new method for analyzing MLLM fusion. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
