Fine-tuning Vision-Language Models (VLMs) is a common practice to adapt pre-trained models for downstream tasks. Yet existing fine-tuning methods consistently keep the pre-trained architecture intact, leaving open the question of whether maintaining such structural integrity is truly necessary for effective adaptation. To investigate this, we propose Divide-and-Adapt, a new fine-tuning paradigm that departs from preserving the full pre-trained structure. It first divides the pre-trained model into multiple downstream-oriented subnetworks, and then adapts them collaboratively through a routing-based coordination mechanism. This design facilitates the selection of structures that are most beneficial for downstream tasks and enables their collaborative integration. Remarkably, even without any task-specific training, our Divide-and-Adapt framework first divides the pre-trained model into downstream-oriented subnetworks and adaptively activates them for downstream inference, already outperforming the pre-trained model on downstream tasks. With further fine-tuning, Divide-and-Adapt surpasses both full fine-tuning and parameter-efficient baselines across multiple benchmarks, demonstrating its effectiveness. Our code will be made publicly available at https://github.com/haiiling/DivideAndAdapt.

Post-training quantization (PTQ) has emerged as a promising approach for reducing the memory footprint and computational cost of large language models (LLMs), enabling efficient deployment without full model retraining. However, existing PTQ methods struggle to simultaneously support weight–activation joint quantization and extreme low-bit weight quantization. This limitation primarily arises from the depth of LLMs and their strong cross-layer dependencies, which cause quantization errors to propagate and accumulate across layers, ultimately leading to significant performance degradation. In this paper, we present ACBQ, a simple yet effective framework that simultaneously addresses weight–activation joint quantization and extreme weight quantization. We first propose a granular quantization strategy that treats self-attention and FFN as separate quantization units with module-specific optimization objectives. To mitigate the propagation and accumulation of quantization errors across layers, we introduce an adaptive cross-block quantization strategy that explicitly accounts for cross-layer dependencies by encouraging consistency across blocks. Extensive experiments across diverse LLMs, including OPT and the LLaMA family, demonstrate that ACBQ achieves superior performance under both W4A4 and highly aggressive W2 settings, while incurring negligible additional computational overhead.

Foundation models learn highly transferable representations through large-scale pretraining on diverse data. An increasing body of research indicates that these representations exhibit a remarkable degree of similarity across architectures and modalities. In this survey, we investigate the representation potentials of foundation models, defined as the latent capacity of their learned representations to capture task-specific information within a single modality while also providing a transferable basis for alignment and unification across modalities. We begin by reviewing representative foundation models and the key metrics that make alignment measurable. We then synthesize empirical evidence of representation potentials from studies in vision, language, speech, multimodality, and neuroscience. The evidence suggests that foundation models often exhibit structural regularities and semantic consistencies in their representation spaces, positioning them as strong candidates for cross-modal transfer and alignment. We further analyze the key factors that foster representation potentials, discuss open questions, and highlight potential challenges.