structured-pruning

提出LayerTrim——一种无需训练的结构化剪枝方法，通过保留注意力机制和解码器深度、选择性修剪冗余层的MLP特征，显著缓解了推理语言模型在长链推理任务上的性能崩溃，在Qwen3-14B上实现1.24倍预填充加速和19.7%峰值显存降低。

ReCal 是一种即插即用的校准方法，在结构化剪枝前通过教师-剪枝探针间的前向 KL 识别受损预测并重加权校准统计，从而提升剪枝模型在在线策略蒸馏（OPD）后的恢复效果，在 AIME 上最高提升 16.7 个百分点。

The paper proposes a correlation-aware structured pruning method for large language models, explicitly modeling cross-unit dependencies to improve pruning decisions and achieve competitive accuracy-efficiency trade-offs.

This paper introduces a layerwise decoupling approach for structured sparsification of fully connected layers in neural networks, demonstrating improved robustness and reduced over-pruning compared to joint methods.

LILA is a calibration-free structured pruning method for large language models that uses latent spectral geometry to score neuron importance, achieving competitive performance without calibration data.

This paper proposes a damage-aware multi-armed bandit method for structured post-training pruning of vision and language transformers, showing reduced performance degradation compared to baseline approaches in experiments across various models and datasets.

The paper proposes COEC, a training-free compensation framework for structured pruning of large language models that applies orthogonal rotations and calibration to reduce output error and improve accuracy after column removal.

This paper introduces Sniper, a two-stage structured pruning framework for LLMs that uses binary knapsack optimization to unify depth and width pruning, achieving near-exact compression ratio adherence and improved performance retention across multiple architectures.

TriSP introduces a tri-signal importance metric combining weight magnitude, activation norm, and gradient sensitivity for structured pruning of LLMs, achieving lowest perplexity and high throughput improvements on LLaMA-7B.

Proposes a two-stage structured pruning framework for LLMs that jointly optimizes latency and model size using multi-objective depth pruning and parallel Bayesian optimization, achieving favorable trade-offs for edge deployment.

ShortOPD proposes a short-to-long on-policy distillation schedule that recovers pruned LLMs for free-form generation by focusing training on effective prefixes, achieving up to 9x improvement over unrecovered models and matching long-horizon distillation with a quarter of the training time.

This paper introduces Maestro, a structured pruning framework for Mixture-of-Experts language models that uses Markov chains to model expert activation trajectories, achieving globally aware pruning and outperforming baselines by up to 10.61% under 50% compression.

This paper proposes a structured pruning method for LLMs that addresses distribution mismatch, sign-information loss, and outlier influence when adapting unstructured pruning techniques, achieving comparable accuracy with 1.56-1.57x speedup on models like Llama-3-8B and Vicuna-v1.5-13B.

This paper presents a cascaded multi-granularity pruning framework for deploying LLMs on Industrial IoT edge devices, achieving up to 13.8x compression with minimal accuracy loss on MHA+GELU architectures while exposing a collapse on GQA+SwiGLU designs.

This paper proposes a novel structured neuron pruning framework for deep neural networks using multi-armed bandit algorithms, demonstrating effectiveness on various tasks.

Proposes KOFF, a framework that decomposes pretrained LLMs into a sparse shared backbone and domain-specific external memories using structured pruning and LoRA adapters, achieving 12% sparsity without significant performance loss.