HodgeCover isolates the harmonic kernel of a simplicial Laplacian on an expert 2-complex to identify irreducible merge cycles and selects experts for aggressive compression, matching or exceeding baselines on open-weight MoE models.
hub Canonical reference
A Simple and Effective Pruning Approach for Large Language Models
Canonical reference. 70% of citing Pith papers cite this work as background.
74
Pith papers citing it
52
external citations · Pith
Background
70% of classified citations
abstract
As their size increases, Large Languages Models (LLMs) are natural candidates for network pruning methods: approaches that drop a subset of network weights while striving to preserve performance. Existing methods, however, require either retraining, which is rarely affordable for billion-scale LLMs, or solving a weight reconstruction problem reliant on second-order information, which may also be computationally expensive. In this paper, we introduce a novel, straightforward yet effective pruning method, termed Wanda (Pruning by Weights and activations), designed to induce sparsity in pretrained LLMs. Motivated by the recent observation of emergent large magnitude features in LLMs, our approach prunes weights with the smallest magnitudes multiplied by the corresponding input activations, on a per-output basis. Notably, Wanda requires no retraining or weight update, and the pruned LLM can be used as is. We conduct a thorough evaluation of our method Wanda on LLaMA and LLaMA-2 across various language benchmarks. Wanda significantly outperforms the established baseline of magnitude pruning and performs competitively against recent method involving intensive weight update. Code is available at https://github.com/locuslab/wanda.
hub tools
citation-role summary
background 8
method 2
citation-polarity summary
representative citing papers
VLA models from VLM adaptation can be pruned 12-30% via multi-module joint scheme based on divergence signals while keeping ~90% performance on LIBERO without post-pruning recovery, unlike standard criteria that collapse.
VLA language backbones show high redundancy on manipulation benchmarks, with half the LLM blocks removable and even two blocks sufficient to recover baseline performance after fine-tuning, unlike vision and action pathways.
Fisher-MoE prunes sparse intermediate dimensions in MoE FFNs ranked by Fisher importance, delivering 50% compression that preserves capability while cutting memory ~45% and raising throughput 21%.
EvoESAP uses evolutionary search guided by a speculative-decoding-inspired ESAP metric to discover non-uniform layer-wise sparsity allocations for MoE expert pruning, improving generation accuracy up to 19.6% at 50% sparsity.
Width pruning in Llama-3.2 models reduces parametric knowledge while enhancing instruction-following and preserving reasoning.
Low-rank compression preserves training-data privacy and improves adversarial robustness but weakens personal-information protection, reduces ethical behavior in zero-shot use, and harms fairness.
EMP maps importance scores to effective sample size N_eff and prunes the lowest N - N_eff components, with a derived lower bound on retained effective mass and upper bound on loss increase.
DiT-Pruning squares the weight term in the pruning saliency metric and adapts per-layer versus per-channel granularity, claiming strong quality retention for diffusion transformers at high sparsity.
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
Cascaded multi-granularity pruning reaches 13.8x compression on MHA+GELU LLMs for bearing fault diagnosis at 83.82% accuracy while causing ~74pp collapse on GQA+SwiGLU models that violate the formalized Structural Independence Assumption.
CascadeFormer tapers Transformer width with depth based on gradient fan-in asymmetry to match uniform baselines in perplexity while cutting latency 8.6%.
GRINQH introduces a graded input-based quantization hierarchy that dynamically assigns multi-precision weights using activation magnitudes as importance proxy, unifying quantization with sparsification to improve LLM decoding speed and quality trade-offs on Llama3 and Qwen3 models.
Compositionality emerges in neural networks only in a narrow depth-connectivity regime, with gradient descent converging to fractured solutions outside it.
A structural pruning framework for MoE models that solves channel-score coverage maximization via attribution approximation, preserving accuracy at 50% or 25% pruning plus 4-bit quantization on DeepSeek and Qwen models.
Introduces TQS metric and TQS-PTQ framework that uses dynamical-systems stability to enable a priori, calibration-free mixed-precision post-training quantization for time-series models.
Embedding and removing a dummy backdoor reduces unknown backdoor success in generative LLMs by targeting shared trigger-activated internal mechanisms.
Sparse autoencoders identify shared latent features across diverse backdoor attacks in LLMs that enable unified detection via classifiers, causal control via steering, and mitigation via ablation fine-tuning.
Neural networks are compressed by lumping neurons with approximately matching dynamics in a polynomial ODE encoding, yielding substantial size reduction with preserved accuracy on synthetic and regression tasks.
A systematic MoE-to-dense conversion via expert scoring, grouping, and distillation yields +6.3 pp average accuracy over dense-to-dense pruning at matched parameter count on tested models.
PARE applies structure-aware head pruning and timestep/content-conditioned block routing to compress video DiTs, reducing per-step compute while preserving quality on Wan2.1-14B.
MCWC aligns permutation-symmetric blocks across layers to enable sequential prediction and residual entropy coding, improving rate-accuracy tradeoffs versus quantization and prior codecs on language and vision models.
LAPLEX introduces exact, phased Laplace-kernel operators defined by learnable anchors with FFT scaling for trainable large-scale dense-like layers and Gram matrices in neural networks.
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
citing papers explorer
-
HodgeCover: Higher-Order Topological Coverage Drives Compression of Sparse Mixture-of-Experts
HodgeCover isolates the harmonic kernel of a simplicial Laplacian on an expert 2-complex to identify irreducible merge cycles and selects experts for aggressive compression, matching or exceeding baselines on open-weight MoE models.
-
Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation
VLA models from VLM adaptation can be pruned 12-30% via multi-module joint scheme based on divergence signals while keeping ~90% performance on LIBERO without post-pruning recovery, unlike standard criteria that collapse.
-
Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?
VLA language backbones show high redundancy on manipulation benchmarks, with half the LLM blocks removable and even two blocks sufficient to recover baseline performance after fine-tuning, unlike vision and action pathways.
-
Less is MoE: Trimming Experts in Domain-Specialist Language Models
Fisher-MoE prunes sparse intermediate dimensions in MoE FFNs ranked by Fisher importance, delivering 50% compression that preserves capability while cutting memory ~45% and raising throughput 21%.
-
EvoESAP: Non-Uniform Expert Pruning for Sparse MoE
EvoESAP uses evolutionary search guided by a speculative-decoding-inspired ESAP metric to discover non-uniform layer-wise sparsity allocations for MoE expert pruning, improving generation accuracy up to 19.6% at 50% sparsity.
-
Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2
Width pruning in Llama-3.2 models reduces parametric knowledge while enhancing instruction-following and preserving reasoning.
-
Decomposed Trust: Privacy, Adversarial Robustness, Ethics, and Fairness in Low-Rank LLMs
Low-rank compression preserves training-data privacy and improves adversarial robustness but weakens personal-information protection, reduces ethical behavior in zero-shot use, and harms fairness.
-
Effective Model Pruning: Measure The Redundancy of Model Components
EMP maps importance scores to effective sample size N_eff and prunes the lowest N - N_eff components, with a derived lower bound on retained effective mass and upper bound on loss increase.
-
Post-Training Pruning for Diffusion Transformers
DiT-Pruning squares the weight term in the pruning saliency metric and adapts per-layer versus per-channel granularity, claiming strong quality retention for diffusion transformers at high sparsity.
-
When AI Reviews Its Own Code: Recursive Self-Training Collapse in Code LLMs
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
-
Cascaded Multi-Granularity Pruning for On-Device LLM Inference in Industrial IoT
Cascaded multi-granularity pruning reaches 13.8x compression on MHA+GELU LLMs for bearing fault diagnosis at 83.82% accuracy while causing ~74pp collapse on GQA+SwiGLU models that violate the formalized Structural Independence Assumption.
-
CascadeFormer: Depth-Tapered Transformers Motivated by Gradient Fan-in Asymmetry
CascadeFormer tapers Transformer width with depth based on gradient fan-in asymmetry to match uniform baselines in perplexity while cutting latency 8.6%.
-
GRINQH: Graded Input-based Quantization Hierarchy for Efficient LLM Generation
GRINQH introduces a graded input-based quantization hierarchy that dynamically assigns multi-precision weights using activation magnitudes as importance proxy, unifying quantization with sparsification to improve LLM decoding speed and quality trade-offs on Llama3 and Qwen3 models.
-
Compositionality Emerges in a Narrow Depth-Connectivity Regime: Architecture Constraints and Solution Manifolds
Compositionality emerges in neural networks only in a narrow depth-connectivity regime, with gradient descent converging to fractured solutions outside it.
-
Attribution-Guided and Coverage-Maximized Pruning for Structural MoE Compression
A structural pruning framework for MoE models that solves channel-score coverage maximization via attribution approximation, preserving accuracy at 50% or 25% pruning plus 4-bit quantization on DeepSeek and Qwen models.
-
Quantizing Time-Series Models As Dynamical Systems: Trajectory-Based Quantization Sensitivity Score
Introduces TQS metric and TQS-PTQ framework that uses dynamical-systems stability to enable a priori, calibration-free mixed-precision post-training quantization for time-series models.
-
Dummy Backdoor as a Defense: Removing Unknown Backdoors via Shared Internal Mechanisms for Generative LLMs
Embedding and removing a dummy backdoor reduces unknown backdoor success in generative LLMs by targeting shared trigger-activated internal mechanisms.
-
Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMs
Sparse autoencoders identify shared latent features across diverse backdoor attacks in LLMs that enable unified detection via classifiers, causal control via steering, and mitigation via ablation fine-tuning.
-
Neural Network Compression by Approximate Differential Equivalence
Neural networks are compressed by lumping neurons with approximately matching dynamics in a polynomial ODE encoding, yielding substantial size reduction with preserved accuracy on synthetic and regression tasks.
-
Pruning and Distilling Mixture-of-Experts into Dense Language Models
A systematic MoE-to-dense conversion via expert scoring, grouping, and distillation yields +6.3 pp average accuracy over dense-to-dense pruning at matched parameter count on tested models.
-
PARE: Pruning and Adaptive Routing for Efficient Video Generation
PARE applies structure-aware head pruning and timestep/content-conditioned block routing to compress video DiTs, reducing per-step compute while preserving quality on Wan2.1-14B.
-
Motion-Compensated Weight Compression
MCWC aligns permutation-symmetric blocks across layers to enable sequential prediction and residual entropy coding, improving rate-accuracy tradeoffs versus quantization and prior codecs on language and vision models.
-
LAPLEX: The FFT of Learnable Laplace Kernels
LAPLEX introduces exact, phased Laplace-kernel operators defined by learnable anchors with FFT scaling for trainable large-scale dense-like layers and Gram matrices in neural networks.
-
Search Your Block Floating Point Scales!
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
-
ADMM-Q: An Improved Hessian-based Weight Quantizer for Post-Training Quantization of Large Language Models
ADMM-Q is a new post-training quantization method using ADMM operator splitting that reduces WikiText-2 perplexity compared to GPTQ on Qwen3-8B across W3A16, W4A8, and W2A4KV4 settings.
-
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO is a sparse MoE architecture with ReLU-based routing, learnable expert scaling, and NormSiLU activation that matches dense Transformer performance at 20% expert activation and delivers 2.93x speedup on Jetson AGX Orin.
-
XPERT: Expert Knowledge Transfer for Effective Training of Language Models
XPERT extracts and reuses cross-domain expert knowledge from pre-trained MoE LLMs via inference analysis and tensor decomposition to improve performance and convergence in downstream language model training.
-
Different Prompts, Different Ranks: Prompt-aware Dynamic Rank Selection for SVD-based LLM Compression
PARSE trains a prompt-aware linear router on dense-model outputs to select dynamic SVD ranks, improving accuracy up to 10% at 0.6 compression ratio on LLaMA-7B while delivering 2.5x prefill and 2.4x decode speedups.
-
SparseForge: Efficient Semi-Structured LLM Sparsification via Annealing of Hessian-Guided Soft-Mask
SparseForge achieves 57.27% zero-shot accuracy on LLaMA-2-7B at 2:4 sparsity using only 5B retraining tokens, beating the dense baseline and nearly matching a 40B-token SOTA method.
-
Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models
CAP scores attention heads via interventional masking on reasoning calibration data and converts those scores into weight pruning decisions, reporting up to 61% relative accuracy gains over Wanda at 20% sparsity on ARC-Challenge.
-
Temporally Extended Mixture-of-Experts Models
Temporally extended MoE layers using the option-critic framework with deliberation costs cut switching rates below 5% while retaining most capability on MATH, MMLU, and MMMLU.
-
SLaB: Sparse-Lowrank-Binary Decomposition for Efficient Large Language Models
SLaB compresses LLM weights via sparse-lowrank-binary decomposition guided by activation-aware scores, achieving up to 36% lower perplexity than prior methods at 50% compression on Llama models.
-
Compressed-Sensing-Guided, Inference-Aware Structured Reduction for Large Language Models
A unified compressed-sensing framework enables dynamic, task- and token-adaptive structured reduction of LLMs with formal sample-complexity bounds.
-
Double: Breaking the Acceleration Limit via Double Retrieval Speculative Parallelism
Double achieves up to 5.3x inference speedup on 70B LLMs via synchronous double retrieval speculative parallelism that is lossless and outperforms trained baselines like EAGLE-3.
-
OmniZip: Audio-Guided Dynamic Token Compression for Fast Omnimodal Large Language Models
OmniZip introduces an audio-guided dynamic token compression framework that achieves 3.42X inference speedup and 1.4X memory reduction for omnimodal LLMs without any training.
-
CoSpaDi: Compressing LLMs via Calibration-Guided Sparse Dictionary Learning
CoSpaDi uses calibration-guided sparse dictionary learning to compress LLM weight matrices, beating SVD-based and structured pruning baselines in accuracy and perplexity at 20-40% compression ratios.
-
Forget What's Sensitive, Remember What Matters: Token-Level Differential Privacy in Memory Sculpting for Continual Learning
PeCL applies token-level dynamic differential privacy and privacy-guided memory sculpting to achieve superior privacy-utility balance in continual learning.
-
Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction
A pruning technique called Reasoning-Aware Compression (RAC) jointly reconstructs input and chain-of-thought activations to preserve reasoning performance better than standard methods when compressing models like DeepSeek-R1.
-
MaskPro: Linear-Space Probabilistic Learning for Strict (N:M)-Sparsity on LLMs
MaskPro learns categorical distributions over groups of M weights to generate exact (N:M) sparsity via N-way sampling without replacement and stabilizes training with a moving average tracker of loss residuals.
-
A3 : an Analytical Low-Rank Approximation Framework for Attention
A3 splits Transformer layers into QK, OV, and MLP components and derives analytical low-rank approximations that reduce hidden dimensions while minimizing each component's functional loss, yielding better perplexity than prior low-rank methods on LLaMA models.
-
Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning
CD-MoE condenses fine-grained MoE layers with shared experts into dense layers, retaining 90% accuracy with 27.5% memory cut and 1.26x speedup on DeepSeekMoE-16B, recovering 98% via brief fine-tuning.
-
NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models
NV-Embed achieves first place on the MTEB leaderboard across 56 tasks by combining a latent attention layer, causal-mask removal, two-stage contrastive training, and data curation for LLM-based embedding models.
-
H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
H2O evicts non-heavy-hitter tokens from the KV cache using a dynamic submodular policy, retaining recent and frequent-co-occurrence tokens to reduce memory while preserving accuracy.
-
PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning
Activation-percentile-based per-layer sparsity allocation improves LLaMA-2-7B pruning by 15% perplexity over uniform Wanda, while gradient-based allocation fails worse than random.
-
CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse
CreativityNeuro applies contrastive weight steering to LLMs, yielding up to 14 percentile gains on the Divergent Association Task and improved originality in human-rated tests while reducing mode collapse.
-
Scaling Up Thermodynamic AI Models
A backpropagation training method for deep conv nets enables thermodynamic inference on Ising hardware with reported CIFAR accuracies plus theory bounding inference cost versus accuracy.
-
Activation- and Influence-Aware Ranks (AIR): Function-Preserving SVD Compression for LLMs
AIR augments activation-aware SVD compression of LLMs with an influence metric and a closed-form ALS update, claiming >18% perplexity improvement at 60% parameter retention and 90% less calibration data than SVD-LLM(W).
-
LLM Parameters for Math Across Languages: Shared or Separate?
Mechanistic analysis of LLMs finds partial overlap in math-associated parameters across languages, concentrated in middle layers, with systematic language-dependent differences.
-
Multi-Bitwidth Quantization for LLMs Using Additive Codebooks
Drop-by-Drop uses additive codebooks and Matryoshka-style training to produce one LLM model whose ordered codebook subsets give accurate reconstructions at successively higher bitwidths under a weighted MSE distortion.
-
EinSort: Sorting is All We Need for Tensorizing LLM
Sorting tensor indices enables an adaptive tensorization method that discovers low-rank structure in LLM weights and KV caches, yielding better reconstruction quality than baselines.