针对大模型CoT推理中“过度思考”的冗余步骤，用最优停止策略动态剪枝，降本增效不丢精度。
arXiv:2607.11089v1 Announce Type: new Abstract: Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks through Chain…
突破传统一刀切剪枝，PALS按每层百分位自适应稀疏度，精准提升大模型压缩效率。
arXiv:2607.07557v1 Announce Type: cross Abstract: One-shot pruning methods like Wanda and SparseGPT apply the same sparsity ratio to every layer of a …
扩散变换器虽强但计算开销大，后训练剪枝技术如何高效优化？这篇论文给出新方案。
arXiv:2607.00927v1 Announce Type: cross Abstract: Diffusion Transformers (DiTs) have demonstrated impressive performance in image generation but suffe…
研究修剪注意力层如何影响LLM的解释忠实性与置信度校准，揭示模型优化新视角。
arXiv:2606.24970v1 Announce Type: new Abstract: Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the netwo…
LLM压缩新突破：将块删除问题转化为约束二元优化，物理模型Ising玻璃降低下游性能损失
arXiv:2602.00161v2 Announce Type: replace Abstract: In this paper, we formulate the compression of large language models (LLMs) by optimally deleting …
剪枝后的大模型在多项选择基准测试中表现优异，却在开放回答任务中“原形毕露”，揭示评测陷阱。
arXiv:2606.17609v1 Announce Type: new Abstract: Compressing large language models reduces memory use and inference cost, but it can also create failur…
平衡三元量化（-1,0,1）将剪枝与量化合二为一，微软BitNet已证实在100B+参数上性能持平FP16。
Why Balanced Ternary {-1, 0, +1} Could Be the Future of AI Hardware** For 70 years, computing has been binary: 0 or 1. But AI workloads are fundamenta…
新论文提出多粒度节点剪枝方法，高效发现因果电路，提升模型可解释性。
arXiv:2512.10903v2 Announce Type: replace Abstract: Circuit discovery aims to identify minimal subnetworks that are responsible for specific behaviors…
一种实用的一次性剪枝方法，同时支持稀疏与密集GEMM运算，显著降低LLM推理成本。
arXiv:2606.10445v1 Announce Type: new Abstract: Semi-structured 2:4 sparsity is widely supported by modern accelerators, providing up to a 2x theoreti…
MoE大模型专家剪枝新方法，梯形剪枝保留路由拓扑，大幅压缩静态参数占用。
arXiv:2606.09885v1 Announce Type: new Abstract: Mixture-of-Experts large language models (LLMs) scale efficiently through sparse activation, yet their…
新方法通过交换感知的令牌剪枝，高效提升音频-视觉字幕生成性能，论文来自arXiv最新研究。
arXiv:2606.10533v1 Announce Type: new Abstract: Audio-visual captioning generates natural language descriptions from video and audio content. Multimod…
针对Agent推理场景的跨轮意图感知KV缓存剪枝方法，显著降低内存占用与延迟。
arXiv:2606.09916v1 Announce Type: cross Abstract: Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and in…
稀疏MoE大模型部署新突破：引入联盟感知策略的专家剪枝方法
arXiv:2606.09886v1 Announce Type: cross Abstract: Sparse Mixture-of-Experts (MoE) large language models achieve strong quality with low per-token comp…
长对话中动态修剪上下文，大幅提升LLM效率与性能，看完直呼“省省你的token”！
arXiv:2601.07994v5 Announce Type: replace Abstract: Large Language Models (LLMs) increasingly operate over long-form dialogues with frequent topic shi…
提出Hyperflux，一种基于L0的新型网络剪枝方法，将剪枝建模为连续演化系统，揭示剪枝过程中神经元重要性变化。
arXiv:2504.05349v4 Announce Type: replace-cross Abstract: Network pruning is used to reduce inference latency and power consumption in large neural ne…
联合结构化剪枝与混合精度量化，为LLM压缩提供高效新方案，兼顾精度与速度。
arXiv:2606.07819v1 Announce Type: cross Abstract: Recently, the efficiency of Large Language Models (LLMs) deployment has become a critical concern in…
提出拓扑感知层剪枝方法，有效降低大视觉语言模型的计算与内存成本，助力高效部署。
arXiv:2604.16502v2 Announce Type: replace Abstract: Large Language Models (LLMs) have demonstrated strong capabilities in natural language understandi…
压缩LLM时，准确率不是唯一指标——新基准用保形概率评估不确定性保留。
arXiv:2606.01850v1 Announce Type: new Abstract: Model compression techniques such as quantization and pruning are widely used to reduce the deployment…
大模型剪枝新突破：CRePE用卷积感知相对重要性实现高效后训练剪枝，大幅降低部署开销
arXiv:2606.01544v1 Announce Type: new Abstract: Deploying Large Language Models (LLMs) in practice incurs substantial memory and computational costs. …
在MacBook上也能轻松剪枝MoE专家模型了，Cerebras REAP方法完美移植到MLX。
Article URL: https://github.com/egesabanci/reap-mlx Comments URL: https://news.ycombinator.com/item?id=48364457 Points: 2 # Comments: 0