Mixture-of-Experts (MoE) has become a key architecture for scaling large language models. Recent fine-grained MoE designs introduce hundreds of experts per layer, with multiple experts activated per token, enabling stronger specialization. However, during pre-training, routers are optimized mainly for stability and robustness: they converge prematurely and enforce balanced usage, limiting the full potential of model performance and efficiency at inference. In this work, we uncover two overlooked issues: (i) a few highly influential experts are underutilized due to premature and balanced routing decisions; and (ii) enforcing a fixed number of active experts per token introduces substantial redundancy. Instead of retraining or redesigning MoE architectures, we introduce Ban Pick, a training-free, plug-and-play strategy for smarter routing. Pick discovers and reinforces key experts, a small group with outsized impact on performance, leading to notable accuracy gains. Ban further dynamically prunes redundant experts based on layer and token sensitivity, delivering faster inference. Experiments on fine-grained MoE-LLMs (DeepSeek, Qwen3) across math, code, and general reasoning benchmarks demonstrate that Ban Pick delivers free performance gains and inference acceleration. For instance, on Qwen3-30B-A3B, it improves accuracy from 80.67 to 84.66 on AIME2024 and from 65.66 to 68.18 on GPQA-Diamond, while accelerating inference by 1.25× under the vLLM.

Parameter-efficient fine-tuning (PEFT) has become a prevalent approach for adapting large language models (LLMs). However, low-rank adaptation methods face an inherent trade-off: improving target task performance can compromise pre-trained world knowledge, while aggressively constraining updates to preserve world knowledge may hinder improvements in the target task. Furthermore, most current methods fail to account for layer-wise differences in adaptation sensitivity, resulting in suboptimal preservation of world knowledge and task adaptation. To address these challenge, we propose Fisher-Optimized Adaptive Low Rank and Singular-VectorSelection (FARSS), an effective framework for knowledge-preserving fine-tuning. This framework introduces two key innovations. First, we propose a Fisher-guided adaptive rank allocation strategy, which assigns smaller ranks to shallow layers that are critical for preserving world knowledge, and larger ranks to deep layers that are essential for task adaptation. Second, we introduce a task-aware initialization method that integrates singular value information with layer-specific second-order statistics estimated from activation and gradient covariances, enabling efficient and task-sensitive low-rank updates. We evaluated several models across various tasks, and the experimental results show that our approach outperforms existing PEFT methods, including LoRA, Corda, and KaSA, achieving a balance between preserving world knowledge and enhancing target task performance. The code is available at https://github.com/chenyehuang/FARSS.

Mixture-of-Experts Multimodal Large Language Models (MoE-MLLMs) offer remarkable performance but incur prohibitive GPU memory costs, making compression essential. Among PTQ methods, expert-level mixed-precision quantization has proven effective for MoE-LLMs, yet suffers notable degradation on MoE-MLLMs due to two overlooked biases in expert importance estimation. (1) At the cross-modal level, the numerical dominance of vision tokens causes expert selection frequency to be dominated by vision tokens, masking experts that are critical to the text modality; (2) at the intra-vision level, the large proportion of redundant vision tokens further skew frequency statistics, obscuring experts critical for informative visual content. To bridge gaps, we propose MODE, a modality-decomposed expert-level mixed-precision quantization framework for MoE-MLLMs that decomposes expert selection frequency by modality, filters redundant vision tokens to obtain denoised visual frequency, and further evaluates quantization sensitivity per modality as a complementary signal to frequency-based estimation. These signals are integrated into an Integer Linear Programming formulation to assign per-expert bit-widths under a given budget. Extensive experiments show that MODE is particularly well-suited for MoE-MLLMs, limiting average performance loss to within 2.9% at W3A16, with larger gains at the extreme 2-bit setting.

Large Language Models (LLMs) have revolutionized various domains but encounter substantial challenges in tackling optimization modeling tasks for Operations Research (OR), particularly when dealing with complex problem. In this work, we propose Step-Opt-Instruct, a framework that augments existing datasets and generates high-quality fine-tuning data tailored to optimization modeling. Step-Opt-Instruct employs iterative problem generation to systematically increase problem complexity and stepwise validation to rigorously verify data, preventing error propagation and ensuring the quality of the generated dataset. Leveraging this framework, we fine-tune open-source LLMs, including LLaMA-3-8B and Mistral-7B, to develop Step-Opt—a model that achieves state-of-the-art performance on benchmarks such as NL4OPT, MAMO, and IndustryOR. Extensive experiments demonstrate the superior performance of Step-Opt, especially in addressing complex OR tasks, with a notable 17.01% improvement in micro average accuracy on difficult problems. These findings highlight the effectiveness of combining structured validation with gradual problem refinement to advance the automation of decision-making processes using LLMs. The code and dataset are available at https://github.com/samwu-learn/Step.

As the parameter size of language models becomes extremely large, fine-tuning them with limited resources has become a challenging task. Latest advancements in parameter-efficient fine-tuning (PEFT) techniques allow for adjustments to only a minor fraction of the parameters of these LLMs. Yet, most of PEFT methods may suffer from the following limitations: (1) As the rank decreases sharply, PEFT methods like LoRA and Adapter tuning will exhibit significant performance degradation in downstream tasks. (2) An accuracy gap between these methods and full fine-tuning (Full-FT) still exists. To tackle these problems, we propose a Low-Rank Direct Attention Adaptation (LoRaDA) method for efficient LLM fine-tuning. Specifically, we introduce a novel Low-rank Multi-head Attention Map Module (LMAM), which can bring negative attention to self-attention modules and learn low-rank attention weights directly, capturing the characteristics of downstream tasks. Furthermore, LMAM can serve as a plug-in to existing methods, such as LoRA and Adapter, providing state-of-the-art performance even with extreme low rank setting.Extensive experiments on various downstream tasks demonstrate the superior performance of our LoRaDA method. Specifically, LoRaDA even outperforms the full fine-tuning method by up to 2.1% on GLUE benchmark. As a plug-in, LMAM boosts the accuracy of LoRA by up to 27.7% with LLaMA-7B on Commonsense Reasoning benchmark.

Delta compression methods focus on efficiently serving multiple uniquely fine-tuned models, each tailored to specific tasks and user requirements. These approaches decompose a fine-tuned LLM into a base model and corresponding delta weights, which are compressed using low-rank or low-bit representations to reduce storage costs. However, their effectiveness is highly sensitive to the magnitude of the model deltas—a factor directly influenced by the scale of the training data. We propose the Residual Quantization Tree (RQT), a hierarchical quantization framework that automatically shares low-bit integer weights across similar fine-tuned models. The RQT construction employs a two-phase greedy algorithm: a bottom-up aggregation of models based on weight matrix similarity, and top-down residual quantization, in which each node optimizes the quantization parameters and then delegates residual errors to child nodes. We evaluate RQT on fine-tuned models across mathematics, coding, chatbot, and Chinese LLMs. The results show that RQT achieves an average accuracy degradation of approximately 3% (comparable to previous 4-bit post-training quantization) while maintaining an effective bitwidth of around 2 bits.

State Space Models (SSMs), such as Mamba, have recently demonstrated potential in language understanding tasks, positioning them as competitors to transformer architectures. However, our investigations reveal that the Mamba architecture still has room for further optimization—not only in linear projections but also in state caches, which contribute significantly to memory consumption, particularly after quantizing the former into low bits. After a theoretical analysis of the causes of outliers in states, we propose Decoupled Scale Quantization (DSQ), which mitigates outliers in both the state and channel dimensions by applying separate quantization scales. To preserve the selective ability of quantized Mamba, we introduce Efficient Selectivity Reconstruction (ESR), a novel quantization simulation scheme in block-wise reconstruction that enables fast parallel scan algorithms with the non-linear quantization function. We demonstrate the effectiveness of Q-Mamba across various quantization settings, model sizes, and both generation and zero-shot tasks. In particular, for Mamba2-2.7B with W8A8H4 (8-bit weights and activations, 4-bit state caches) quantization, Q-Mamba achieves a 50% reduction in memory consumption with only a 2.13% average accuracy degradation on zero-shot tasks.

Mixture-of-Experts (MoE) has demonstrated promising potential in scaling LLMs. However, it is hindered by two critical challenges: (1) substantial GPU memory consumption to load all experts; (2) low activated parameters cannot be equivalently translated into inference acceleration effects. In this work, we propose EAC-MoE, an Expert-Selection Aware Compressor for MoE-LLMs, which deeply aligns with the characteristics of MoE from the perspectives of quantization and pruning, and introduces two modules to address these two challenges respectively: (1) The expert selection bias caused by low-bit quantization is a major factor contributing to the performance degradation in MoE-LLMs. Based on this, we propose Quantization with Expert-Selection Calibration (QESC), which mitigates the expert selection bias by calibrating the routers within the MoE; (2) There are always certain experts that are not crucial for the corresponding tasks, yet causing inference latency. Therefore, we propose Pruning based on Expert-Selection Frequency (PESF), which significantly improves inference speed by pruning less frequently used experts for current task. Extensive experiments demonstrate that our approach significantly reduces memory usage and improves inference speed with minimal performance degradation.