Researchers from the University of California, Berkeley, have published a paper proposing a new approach to training multimodal large language models. Instead of fine-tuning both the model's backbone and projector, they suggest only training the projector between the backbone and modality-specific encoder. Experiments on 3D models show that this method achieves strong performance compared to baseline models and joint training methods. Moreover, it avoids undesirable drift in existing capabilities and has higher training sample throughput. The findings are validated across various benchmarks. --- Why it matters: This study is important for AI researchers as it challenges the conventional approach to multimodal model training. By showing that projector-only training can achieve comparable results with less computational overhead, it opens up new possibilities for efficient and effective model adaptation in real-world applications. Source: https://arxiv.org/abs/2608.19726