Post-training quantization (PTQ) for multimodal large language models (MLLMs) faces a distinct optimization challenge: text and vision tokens can induce conflicting gradients when learning a shared channel-wise transformation. We identify this issue as modality gradient conflict, characterized by opposing update directions and severe magnitude imbalance between modality-specific gradients, and propose Modality-Coordinated Transformation Quantization (MCTQ) to address it. MCTQ combines whitening-based preconditioning, modality-decomposed reconstruction, and modality-coordinated optimization to learn a single shared transformation that serves both modalities without introducing modality-specific parameters or inference overhead. Across multiple MLLMs, benchmarks, and precision settings, MCTQ consistently outperforms existing channel-wise PTQ methods, with the largest gains under aggressive low-bit quantization. These results show that coordinating modality-specific gradients is critical for learning effective shared transformations in multimodal PTQ. Code and reproduction materials are available at https://github.com/jtsdw/UniMQ.

Pruning is essential for the efficient deployment of Large Language Models (LLMs); however, it causes severe performance degradation due to the structural distortion induced by sparsity.Existing recovery strategies, such as LoRA, predominantly employ global fine-tuning, often overlooking the mechanistic root of this degradation: the layer-wise accumulation and amplification of local errors. To address this limitation, we propose LaCo(Layer-wise Compensation), a framework that reorients the recovery paradigm from global adaptation to hierarchical representation alignment.By sequentially optimizing each layer to reconstruct the model’s hidden states, LaCo effectively intercept the error propagation chain at its source.Extensive experiments demonstrate that LaCo surpasses parameter-efficient baselines in both perplexity reduction and zero-shot reasoning.Notably, it reduces recovery-time memory usage to approximately 1/7 of the baseline and requires only 2,048 unlabeled samples to match a LoRA model trained on 50k examples—achieving a ∼25× improvement in data efficiency.