Heterogeneous Pruning Scores

Depth pruning improves the deployment efficiency of large language models (LLMs) by identifying and removing redundant layers. A widely accepted standard for this identification process is to measure the similarity between layers using cosine distance. However, we find that methods relying solely on this one-dimensional heuristic can exhibit unpredictable performance and even catastrophic collapse across different architectures. To address this issue, we propose SimDiff, a novel layer importance criterion that jointly evaluates layers from two orthogonal perspectives: representational similarity and transformation difference. The difference is quantified using two distinct metrics: MSSD, which is sensitive to outliers and identifies layers that make decisive corrections, and MASD, which robustly measures a layer's average contribution. Extensive experiments on multiple models ranging from 0.5B to 13B parameters demonstrate that SimDiff significantly outperforms state-of-the-art baselines across various pruning ratios. Notably, our method retains over 91% of LLaMA2-7B's performance at a 25% pruning ratio and achieves up to a 1.49x inference speedup when pruning 12 layers on LLaMA3.1-8B. We also show that pruned models can be effectively recovered with minimal fine-tuning.

Large language models (LLMs) are expensive to serve because model parameters, attention computation, and KV caches impose substantial memory and latency costs. We present GRASPrune, a structured pruning framework applied after pretraining that jointly prunes FFN channels and KV head groups under a single global budget. Instead of learning importance scores without constraints and applying the budget only after training, GRASPrune learns lightweight gate scores with a projected straight-through estimator that enforces a hard mask satisfying the budget at every step while keeping the backbone weights frozen. After the mask is fixed, we calibrate scaling factors on the retained units to mitigate scale mismatch caused by pruning, and fold these factors into the pruned weights to obtain a smaller dense checkpoint with no extra parameters at inference. On LLaMA-2-7B, GRASPrune removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five benchmarks, using four epochs on 512 unlabeled calibration sequences on a single NVIDIA A100 80GB GPU without any full model fine-tuning.

Merging multiple Low-Rank Adaptation (LoRA) experts into a single backbone is a promising approach for efficient multi-task deployment. While existing methods strive to alleviate interference via weight interpolation or subspace alignment, they rest upon the implicit assumption that all LoRA matrices contribute constructively to the merged model. In this paper, we uncover a critical bottleneck in current merging paradigms: the existence of $\textit{negative modules}$ -- specific LoRA layers that inherently degrade global performance upon merging. We propose $\textbf{E}$volutionary $\textbf{N}$egative $\textbf{M}$odule $\textbf{P}$runing ($\textbf{ENMP}$), a plug-and-play LoRA pruning method to locate and exclude these detrimental modules prior to merging. By leveraging an evolutionary search strategy, ENMP effectively navigates the discrete, non-differentiable landscape of module selection to identify optimal pruning configurations. Extensive evaluations demonstrate that ENMP consistently boosts the performance of existing merging algorithms, achieving a new state-of-the-art across both language and vision domains. Code is available at https://github.com/CaoAnda/ENMP-LoRAMerging.

Existing feature-interpretation pipelines typically operate on uniformly sampled units or exhaustive feature sets, incurring massive costs on units irrelevant to target behaviors. To address this, we introduce the first CLT-native end-to-end pruning framework, PIE, which pioneers the paradigm of pruning first and interpreting later. PIE connects Pruning, automatic Interpretation, and interpretation Evaluation, establishing a comprehensive benchmarking environment to systematically measure behavioral fidelity and downstream interpretability under pruning. Within this framework, we adapt strong relevance baselines and propose Feature Attribution Patching (FAP), a patch-grounded attribution method that scores CLT features by aggregating gradient-weighted write contributions. Furthermore, we introduce FAP-Synergy, a systematic synergy-aware reranking procedure. We evaluate pruning using KL-divergence behavior retention and assess interpretation quality with FADE-style metrics across IOI and Doc-String datasets. Across budget constraints of K in {50, 100, 200, 400, 800}, our rigorous benchmarking reveals distinct operational regimes: while base FAP and adapted baselines perform robustly at relaxed budgets, FAP-Synergy excels in highly constrained, strict-budget regimes. Crucially, we demonstrate a practical "Effective Budget" advantage: on the IOI task for both Llama-3.2-1B and Gemma-2-2B, FAP-Synergy at K=50 functionally matches the behavioral fidelity of baseline circuits at K=75. Because downstream evaluation costs scale linearly per feature, Synergy effectively grants the pipeline 25 "free" features, achieving K=75 fidelity while reducing interpretation costs by 33%.

Parallel reasoning enhances Large Reasoning Models (LRMs) but incurs prohibitive costs due to futile paths caused by early errors. To mitigate this, path pruning at the prefix level is essential, yet existing research remains fragmented without a standardized framework. In this work, we propose the first systematic taxonomy of path pruning, categorizing methods by their signal source (internal vs. external) and learnability (learnable vs. non-learnable). This classification reveals the unexplored potential of learnable internal methods, motivating our proposal of STOP (Super TOken for Pruning). Extensive evaluations across LRMs ranging from 1.5B to 20B parameters demonstrate that STOP achieves superior effectiveness and efficiency compared to existing baselines. Furthermore, we rigorously validate the scalability of STOP under varying compute budgets - for instance, boosting GPT-OSS-20B accuracy on AIME25 from 84% to nearly 90% under fixed compute budgets. Finally, we distill our findings into formalized empirical guidelines to facilitate optimal real-world deployment. Code, data and models are available at https://bijiaxihh.github.io/STOP

Visual token pruning methods effectively mitigate the quadratic computational growth caused by processing high-resolution images and video frames in vision-language models (VLMs). However, existing approaches rely on predefined pruning configurations without determining whether they achieve computation-performance optimality. In this work, we introduce , a novel framework that formulates visual token pruning as a Pareto configuration optimization problem to automatically identify optimal configurations. Our approach employs continuous relaxation and straight-through estimators to enable gradient-based search, solved via the Augmented Lagrangian method. Extensive experiments across 8 visual benchmarks demonstrate that effectively approximates the empirical Pareto frontier obtained through grid search and generalizes well across various pruning methods and VLM architectures. Furthermore, through learnable kernel functions, we investigate layer-wise pruning patterns and reveal that multi-step progressive pruning captures VLMs' hierarchical compression structure, achieving superior accuracy-efficiency trade-offs compared to single-layer approaches.

Deploying deep neural networks on edge devices requires balancing accuracy, latency, and resource constraints under realistic execution conditions. To fit models within these constraints, two broad strategies have emerged: static compression techniques such as pruning and quantization, which permanently reduce model size, and dynamic approaches such as early-exit mechanisms, which adapt computational cost at runtime. While both families are widely studied in isolation, they are rarely compared under identical conditions on physical hardware. This paper presents a unified deployment-oriented comparison of static compression and dynamic early-exit mechanisms, evaluated on real edge devices using ONNX based inference pipelines. Our results show that static and dynamic techniques offer fundamentally different trade-offs for edge deployment. While pruning and quantization deliver consistent memory footprint reduction, early-exit mechanisms enable input-adaptive computation savings that static methods cannot match. Their combination proves highly effective, simultaneously reducing inference latency and memory usage with minimal accuracy loss, expanding what is achievable at the edge.

Pruning is a core technique for compressing neural networks to improve computational efficiency. This process is typically approached in two ways: one-shot pruning, which involves a single pass of training and pruning, and iterative pruning, where pruning is performed over multiple cycles for potentially finer network refinement. Although iterative pruning has historically seen broader adoption, this preference is often assumed rather than rigorously tested. Our study presents one of the first systematic and comprehensive comparisons of these methods, providing rigorous definitions, benchmarking both across structured and unstructured settings, and applying different pruning criteria and modalities. We find that each method has specific advantages: one-shot pruning proves more effective at lower pruning ratios, while iterative pruning performs better at higher ratios. Building on these findings, we advocate for patience-based pruning and introduce a hybrid approach that can outperform traditional methods in certain scenarios, providing valuable insights for practitioners selecting a pruning strategy tailored to their goals and constraints. Source code is available at https://github.com/janumiko/pruning-benchmark.

Network pruning is used to reduce inference latency and power consumption in large neural networks. However, most methods focus on empirical results at the expense of understanding the pruning process. We introduce Hyperflux, a novel $L_0$ method which models pruning as a continuously evolving system determined by flux, the gradient response to a weight's removal, and pressure, a global regularization driving weights toward pruning. By exploiting this model, Hyperflux's pruning behavior becomes understandable at both microscopic (weight regrowth/pruning) and macroscopic (sparsity convergence, etc.) levels. We also introduce a novel pressure scheduler that reliably targets desired sparsities. Hyperflux achieves competitive results with ResNet-50, VGG-19 and DeiT-T/S on CIFAR-10, CIFAR-100 and ImageNet datasets.

Large Language Models (LLMs) have achieved significant success across various NLP tasks. However, their massive computational costs limit their widespread use, particularly in real-time applications. Structured pruning offers an effective solution by compressing models and directly providing end-to-end speed improvements, regardless of the hardware environment. Meanwhile, different components of the model exhibit varying sensitivities towards pruning, calling for non-uniform model compression. However, a pruning method should not only identify a capable substructure, but also account for post-compression training. To this end, we propose DarwinLM, a method for training-aware structured pruning. DarwinLM builds upon an evolutionary search process, generating multiple offspring models in each generation through mutation, and selecting the fittest for survival. To assess the effect of post-training, we incorporate a lightweight, multistep training process within the offspring population, progressively increasing the number of tokens and eliminating poorly performing models in each selection stage. We validate our method through extensive experiments on Llama-2-7B, Llama-3.1-8B and Qwen-2.5-14B-Instruct, achieving state-of-the-art performance for structured pruning. For instance, DarwinLM surpasses ShearedLlama while requiring 5x less training data during post-compression training. Code is at: https://github.com/IST-DASLab/DarwinLM