Papers with model compression
TextPruner: A Model Pruning Toolkit for Pre-Trained Language Models (2022.acl-demo)
Copied to clipboard
| Challenge: | Large pre-trained language models have been used for many NLP tasks but computational resources are limited. |
| Approach: | They propose an open-source model pruning toolkit for pre-trained language models . they propose a self-supervised pruning method that can be applied without labeled data. |
| Outcome: | The proposed pruning method reduces model size without retraining the model and speeds up inference speed on the common CPU and GPU devices. |
Efficient Transformer Knowledge Distillation: A Performance Review (2023.emnlp-industry)
Copied to clipboard
| Challenge: | Pretrained transformer language models have been gaining popularity in the field of natural language processing . however, there is no study into the intersection of these two fields . |
| Approach: | They propose a method to extract knowledge from transformers to produce high-performing efficient attention models with low costs. |
| Outcome: | The proposed model compression method preserves up to 98.6% of original model performance across short-context tasks and up to 95.8% on long-concept Named Entity Recognition tasks while decreasing inference times by up to 57%. |
On Linearizing Structured Data in Encoder-Decoder Language Models: Insights from Text-to-SQL (2024.naacl-long)
Copied to clipboard
| Challenge: | Structured data is prevalent in tables, databases, and knowledge graphs, but there is a gap in our understanding of how these linearization-based methods handle structured data, which is inherently non-linear. |
| Approach: | They investigate the linear handling of structured data in encoder-decoder language models, specifically T5. |
| Outcome: | The proposed model can mimic human-designed processes such as schema linking and syntax prediction, and it can be compressed due to modality fusion redundancy. |
An Effective Post-training Embedding Binarization Approach for Fast Online Top-K Passage Matching (2022.aacl-short)
Copied to clipboard
| Challenge: | Existing models that learn semantic representations of passages are prone to performance degradation . embedding binarization is a promising branch of model compression . |
| Approach: | They propose an embedding binarization approach that can be used to optimize for online inference. |
| Outcome: | The proposed model can perform query-passage matching acceleration. |
YANMTT: Yet Another Neural Machine Translation Toolkit (2023.acl-demo)
Copied to clipboard
| Challenge: | Neural machine translation (NMT) is an end-to-end approach that provides stateof-the-art results for a variety of language pairs. |
| Approach: | They propose to build an open-source neural machine translation toolkit on top of HuggingFace's Transformers library and use it for pre-training and fine-tuning sequence-to-sequence models. |
| Outcome: | The proposed toolkit is built on top of the HuggingFace Transformers library and provides advanced features such as document/multi-source NMT, simultaneous NMT and mixtures-of-experts. |
MobileNMT: Enabling Translation in 15MB and 30ms (2023.acl-industry)
Copied to clipboard
| Challenge: | Existing work on NMT models is limited in storage, memory, computation and power consumption. |
| Approach: | They propose a mobile machine translation system that can translate in 15MB and 30ms on devices. |
| Outcome: | The proposed system can translate in 15MB and 30ms on mobile devices. |
BMCook: A Task-agnostic Compression Toolkit for Big Models (2022.emnlp-demos)
Copied to clipboard
Zhengyan Zhang, Baitao Gong, Yingfa Chen, Xu Han, Guoyang Zeng, Weilin Zhao, Yanxu Chen, Zhiyuan Liu, Maosong Sun
| Challenge: | Existing efforts to compress medium-sized models for specific tasks have limited results. |
| Approach: | They propose a task-agnostic compression toolkit for big models that implements quantization, pruning, distillation and MoEfication methods. |
| Outcome: | The proposed tool improves performance on a model with 3 billion parameters by 12x . it also outperforms the original model on three typical NLP benchmarks. |
Fast Vocabulary Transfer for Language Model Compression (2022.emnlp-industry)
Copied to clipboard
| Challenge: | Existing methods to reduce model size and size are expensive and inefficient for some applications. |
| Approach: | They propose a method that relies on vocabulary transfer to reduce model size and inference time while compromising on performance. |
| Outcome: | The proposed method reduces model size and inference time while compromising on performance. |
Speed Without Sacrifice: Fine-Tuning Language Models with Medusa and Knowledge Distillation in Travel Applications (2025.acl-industry)
Copied to clipboard
Daniel Zagyva, Emmanouil Stergiadis, Laurens Van Der Maas, Aleksandra Dokic, Eran Fainman, Ilya Gusev, Moran Beladev
| Challenge: | Rapid growth of digital applications has intensified the demand for real-time natural language processing (NLP) capabilities. |
| Approach: | They propose a framework that combines Medusa and knowledge distillation to achieve compounded benefits in both model size and inference speed. |
| Outcome: | The proposed framework reduces inference latency by 10-20x while maintaining the student model’s performance quality. |
Efficient Vocabulary Reduction for Small Language Models (2025.coling-industry)
Copied to clipboard
| Challenge: | Large language models (LLMs) have high computational costs and energy consumption, making their deployment in industrial settings difficult. |
| Approach: | They propose a small language model that compresses the embedding layer and reduces model size without significant loss of performance. |
| Outcome: | The proposed model reduces the embedding layer while maintaining performance while improving accuracy and performance. |
On Attention Redundancy: A Comprehensive Study (2021.naacl-main)
Copied to clipboard
| Challenge: | Attention redundancy has been observed among attention heads but has not been deeply studied in the literature. |
| Approach: | They propose a multi-layer multi-head self-attention mechanism which is widely applied in modern neural language models. |
| Outcome: | The proposed model is useful for interpretation and model compression. |
Model Compression for Domain Adaptation through Causal Effect Estimation (2021.tacl-1)
Copied to clipboard
| Challenge: | Existing methods for compressing language representation models are not interpretable and do not consider the differences in the predictive power of various model components or the generalizability of the compressed models. |
| Approach: | They propose a model compression scheme that estimates the average treatment effect of a single layer on the model's predictions. |
| Outcome: | The proposed model compression scheme outperforms strong baselines on dozens of domain pairs across three text classification and sequence tagging tasks. |
Minimal Distillation Schedule for Extreme Language Model Compression (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for teacher assistant-based distillation require multiple trials to find the optimal teacher assistant. |
| Approach: | They propose a method that allows scheduling of an optimal teacher assistant in just one trial . they show that student performance is positively correlated with the scale-performance tradeoff . |
| Outcome: | The proposed method can select the optimal teacher assistant in just one trial . it can be used to compare performance of student and teacher assistants on GLUE benchmarks. |
Structured Pruning Learns Compact and Accurate Models (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models have high costs in terms of storage, memory, and computation time. |
| Approach: | They propose a task-specific structured pruning method CoFi which provides highly parallelizable subnetworks and matches distillation methods in both accuracy and latency. |
| Outcome: | The proposed method matches the distillation methods in accuracy and latency without resorting to unlabeled data. |
Robustness Challenges in Model Distillation and Pruning for Natural Language Understanding (2023.eacl-main)
Copied to clipboard
| Challenge: | Recent studies have focused on compressing pre-trained language models (PLMs) however, few studies have examined the impact of compression on generalizability and robustness of compressed models for out-of-distribution data. |
| Approach: | They propose to use knowledge distillation and pruning to reduce model generalization and generalization on out-of-distribution data. |
| Outcome: | The proposed compression techniques overfit on shortcut samples and generalize poorly on hard ones. |
Position-Aware Depth Decay Decoding (D3): Boosting Large Language Model Inference Efficiency (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. |
| Approach: | They propose a token-position aware layer skipping framework to save 1.5x times operations efficiently while maintaining performance. |
| Outcome: | The proposed algorithm achieves 1.5x speedup on large language models with no retraining and with comparable performance on the GSM8K and BBH benchmarks. |
Understanding the Effect of Model Compression on Social Bias in Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) trained with self-supervision on vast corpora of web text fit to the social biases of that text, leading to representational harm. |
| Approach: | They propose to use quantization and knowledge distillation to reduce the computational burden of LLMs to mitigate the effects of inappropriate social biases learned during pretraining. |
| Outcome: | The proposed methods reduce the computational burden of large language models by reducing their size and complexity. |
SWITCH: Studying with Teacher for Knowledge Distillation of Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) has emerged as a popular method for compressing large language models due to high inference costs and memory requirements. |
| Approach: | They propose a method that integrates the teacher model during the student's sequence generation to reduce misguidance from the teacher. |
| Outcome: | Experiments on three model families and five instruction-following datasets show that SWITCH surpasses traditional methods, especially in the generation of long sequential data. |
Extremely Small BERT Models from Mixed-Vocabulary Training (2021.eacl-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods cannot be directly applied to train student models with reduced vocabulary and embedding dimensions. |
| Approach: | They propose a method to align teacher and student embeddings via mixed-vocabulary training. |
| Outcome: | The proposed method compresses BERT-LARGE to a task-agnostic model with smaller vocabulary and hidden dimensions, which is an order of magnitude smaller than other distilled models. |
AlphaTuning: Quantization-Aware Parameter-Efficient Adaptation of Large-Scale Pre-Trained Language Models (2022.findings-emnlp)
Copied to clipboard
Se Jung Kwon, Jeonghoon Kim, Jeongin Bae, Kang Min Yoo, Jin-Hwa Kim, Baeseong Park, Byeongwook Kim, Jung-Woo Ha, Nako Sung, Dongsoo Lee
| Challenge: | Existing approaches to improve inference efficiency by accelerating model fine-tuning have not been thoroughly explored. |
| Approach: | They propose to combine parameter-efficient adaptation and model compression to accelerate model . they propose to freeze binary parameters and scale scaling factors for target tasks . |
| Outcome: | The proposed algorithm achieves >10x compression ratio under 4-bit quantization and >1,000x reduction in trainable parameters. |
BERT-EMD: Many-to-Many Layer Mapping for BERT Compression with Earth Mover’s Distance (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models have been proposed and applied to many NLP tasks, yielding state-of-the-art performance, but high storage and computational costs obstruct them to be effectively deployed on resource-constrained devices and real-time applications. |
| Approach: | They propose a BERT distillation method which allows each intermediate student layer to learn from any intermediate teacher layers. |
| Outcome: | The proposed method can learn from different teacher layers adaptively for different NLP tasks. |
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression (2020.coling-main)
Copied to clipboard
Yihuan Mao, Yujing Wang, Chufan Wu, Chen Zhang, Yang Wang, Quanlu Zhang, Yaming Yang, Yunhai Tong, Jing Bai
| Challenge: | Existing models that use knowledge distillation are memory-intensive and latency-prohibitive . Existing solutions that use this knowledge distilling framework are expensive . |
| Approach: | They propose a solution that uses weight pruning, matrix factorization and knowledge distillation to learn a smaller model. |
| Outcome: | The proposed model reduces the training overheads by an order of magnitude on public datasets while preserving state-of-the-art accuracy. |
Length-Adaptive Distillation: Customizing Small Language Model for Dynamic Token Pruning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to accelerate inference speed are model compression and dynamic computation (e.g., dynamic token pruning). |
| Approach: | They propose a two-stage knowledge distillation framework that produces a customized small language model for dynamic token pruning. |
| Outcome: | The proposed framework can make the small language model more customized for dynamic token pruning and achieve better speed-performance trade-off. |
Tiny Budgets, Big Gains: Parameter Placement Strategy in Parameter Super-Efficient Fine-Tuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods such as LoRA and VeRA use memory-efficient methods to fine-tune large language models. |
| Approach: | They propose a method that uses only 1–5% of the standard LoRA parameters and achieves state-of-the-art performance across a wide range of tasks. |
| Outcome: | The proposed method achieves state-of-the-art performance across a wide range of tasks using only 1–5% of the standard LoRA parameters. |
Compression of Generative Pre-trained Language Models via Quantization (2022.acl-long)
Copied to clipboard
| Challenge: | Existing methods to compress generative pre-trained language models fail on generative tasks due to homogeneous word embeddings and limited memory. |
| Approach: | They propose a token-level contrastive distillation method to learn distinguishable word embeddings and a module-wise dynamic scaling method to make quantizers adaptive to different modules. |
| Outcome: | The proposed method outperforms the state-of-the-art compression methods on generative PLMs by a clear margin. |
E-LANG: Energy-Based Joint Inferencing of Super and Swift Language Models (2022.acl-long)
Copied to clipboard
| Challenge: | E-LANG is an efficient method for building large and highly capable language models . existing methods are only applicable to encoder-only backbones and classification tasks . |
| Approach: | They propose an efficient dynamic inference approach which distributes inference between large accurate Super-models and light-weight Swift models. |
| Outcome: | The proposed method outperforms existing methods on GLUE, SuperGLUE and WMT with 3.3X computation speed and 2.9X computation cost. |
PromptKD: Distilling Student-Friendly Knowledge for Generative Language Models via Prompt Tuning (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have raised concerns about inference costs, increasing the need for research into model compression. |
| Approach: | They propose a method that utilizes prompt tuning to enable generative language models to transfer student-friendly knowledge. |
| Outcome: | Extensive experiments on instruction-following datasets show that PromptKD achieves state-of-the-art performance while adding only 0.0007% of the teacher’s parameters as prompts. |
Divergent Token Metrics: Measuring degradation to prune away LLM components – and optimize quantization (2024.naacl-long)
Copied to clipboard
Björn Deiseroth, Max Meuer, Nikolas Gritsch, Constantin Eichenberg, Patrick Schramowski, Matthias Aßenmacher, Kristian Kersting
| Challenge: | Large Language Models (LLMs) have reshaped natural language processing with impressive capabilities, but their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. |
| Approach: | This study introduces the Divergent Token Metrics (DTMs) that measure token divergences that allow deeper insights into the subtleties of model compression. |
| Outcome: | The proposed measures can identify outliers and improve performance in the sparseness of the LLMs. |
Enabling Lightweight Fine-tuning for Pre-trained Language Model Compression based on Matrix Product Operators (2021.acl-long)
Copied to clipboard
| Challenge: | Existing methods for pre-trained language models (PLMs) use parameter reduction techniques. |
| Approach: | They propose a pre-trained language model compression approach based on the matrix product operator from quantum many-body physics. |
| Outcome: | The proposed approach can decompose an original matrix into central tensors and auxiliary tenses . it can be applied to the original or compressed PLMs in a general way, with a lighter network . |
Enhancing Cross-Tokenizer Knowledge Distillation with Contextual Dynamical Mapping (2025.findings-acl)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) approaches focus on homogeneous architectures with identical tokenizers, constraining their applicability in cross-architecture scenarios. |
| Approach: | They propose a framework that uses contextual information to enhance sequence alignment precision and dynamically improves vocabulary mapping. |
| Outcome: | The proposed framework shows significant advantages over existing methods for model compression . it can be used across multiple model families and across multiple benchmarks . |
Knowledge Distillation with Reptile Meta-Learning for Pretrained Language Model Compression (2022.coling-1)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) can transfer knowledge from the original model into a compact model to achieve model compression. |
| Approach: | They propose a knowledge distillation method with reptile meta-learning to facilitate the transfer of knowledge from the teacher to the student. |
| Outcome: | Extensive experiments on the GLUE benchmark show the proposed method performs better than previous methods. |
Towards Understanding and Improving Knowledge Distillation for Neural Machine Translation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge distillation techniques for neural machine translation lack special treatment on the top-1 information, which is limiting the potential of KD. |
| Approach: | They propose a method to distill knowledge from top-1 predictions of teachers and a technique to infuse more additional knowledge by distilling on the data without ground-truth targets. |
| Outcome: | The proposed method outperforms the vanilla word-level KD and outperfies the existing methods on three different students with different capacity gaps. |
Understanding and Improving Knowledge Distillation for Quantization Aware Training of Large Transformer Encoders (2022.emnlp-main)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) has been used for quantization-aware training to improve the ability of a lightweight model with the transferred knowledge from the teacher. |
| Approach: | They propose two methods to improve attention recovery of quantized large Transformers by combining attention-map and attention-output losses. |
| Outcome: | The proposed methods achieve state-of-the-art accuracy for quantized large Transformer encoder models with sub-2-bit weight quantization. |
Structured Pruning of Large Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in language modeling have led to remarkable improvements on a variety of tasks. |
| Approach: | They propose a generic, structured pruning approach by parameterizing each weight matrix and adaptively removing rank-1 components during training. |
| Outcome: | The proposed method outperforms unstructured pruning and block pruning on language modeling tasks while achieving speedups during training and inference. |
Self-calibration for Language Model Quantization and Pruning (2025.naacl-long)
Copied to clipboard
| Challenge: | Quantization and pruning are fundamental approaches for model compression, but they require large computational resources. |
| Approach: | They propose to use model calibration data to generate synthetic calibrations to improve model performance. |
| Outcome: | The proposed method outperforms other methods using real data in a post-training setting. |
Towards Zero-Shot Knowledge Distillation for Natural Language Processing (2021.emnlp-main)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a common knowledge transfer algorithm used for model compression across a variety of deep learning based natural language processing (NLP) solutions. |
| Approach: | They propose to use teacher training data for model compression . they investigate six tasks and find they can achieve between 75% and 92% of the teacher’s classification score while compressing the model 30 times. |
| Outcome: | The proposed solution achieves between 75% and 92% of the teacher’s classification score while compressing the model 30 times. |
ESF: Efficient Sensitive Fingerprinting for Black-Box Tamper Detection of Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly utilized in diverse applications, including code generation, legal document analysis, medical diagnosis, and decision-making. |
| Approach: | They propose a fingerprinting method tailored for black-box tamper detection of large language models. |
| Outcome: | The proposed method detects tampering with a 99.2% detection rate using 5 fingerprint samples across state-of-the-art LLMs. |
BERT-of-Theseus: Compressing BERT by Progressive Module Replacing (2020.emnlp-main)
Copied to clipboard
| Challenge: | a novel approach to compress neural networks by progressive module replacement is proposed . a number of techniques have been proposed to compress pretraining and fine-tuning models . |
| Approach: | They propose a model compression approach that divides BERT into modules and builds their compact substitutes. |
| Outcome: | The proposed approach outperforms existing knowledge distillation approaches on GLUE benchmark . it is based on a model that divides the original BERT into several modules and builds their substitutes . |
SwiftPrune: Hessian-Free Weight Pruning for Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a novel post-training pruning method relies on the Hessian matrix to perform pruning . current pruning methods are computationally intensive and lack performance due to second-order derivative calculations. |
| Approach: | They propose a Hessian-free weight pruning method that reduces computational burden . they use an Exponentially Weighted Moving Average technique to bypass weight sorting . |
| Outcome: | The proposed method achieves hardware-efficient model compression by eliminating computational intensive calculations. |
PPC-GPT: Federated Task-Specific Compression of Large Language Models via Pruning and Chain-of-Thought Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are becoming more popular and are gaining widespread use in artificial intelligence. |
| Approach: | They propose a unified framework that addresses both privacy preservation and model compression in federated settings. |
| Outcome: | The proposed framework maintains competitive performance comparable to full-sized LLMs while ensuring robust privacy protection through its federated architecture. |
Differentially Private Knowledge Distillation via Synthetic Text Generation (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language models (LLMs) are achieving state-of-the-art performance in many downstream tasks, but data privacy is a major challenge for practitioners. |
| Approach: | They propose a differentially private knowledge distillation algorithm that exploits the knowledge of a teacher LLM and a student's output distribution. |
| Outcome: | The proposed algorithm significantly improves the utility over baselines on the Big Patent dataset, with strong privacy parameters, =2. |
Is Modularity Transferable? A Case Study through the Lens of Knowledge Distillation (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing approaches to modularity are limited to the case of pre-trained modules in a pre-training language model. |
| Approach: | They propose a method that allows the transfer of pre-trained PEFT modules between incompatible PLMs without any change in the inference complexity. |
| Outcome: | The proposed method allows the transfer of modules between incompatible PLMs without any change in the inference complexity. |
Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Prior work on compression prioritizes preserving perplexity, which is analogous to training loss. |
| Approach: | They examine the impact of model compression along four dimensions: degeneration harm, representational harm, dialect bias, and language modeling and downstream task performance. |
| Outcome: | The proposed compression methods can lead to unexpected consequences, the authors show . quantization preserves bias while pruning degrades quickly. |
Low-Rank Prune-And-Factorize for Language Model Compression (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods to reduce parameter redundancy in pre-processed language models fail to retain satisfactory performance under moderate to high compression rates. |
| Approach: | They propose to use network pruning to extract low-rank sparsity pattern desirable to matrix factorization. |
| Outcome: | The proposed method has a superior compression-performance trade-off compared to existing methods. |
TransferCVLM: Transferring Cross-Modal Knowledge for Vision-Language Modeling (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent large vision-language multimodal models pre-trained with huge amount of image-text pairs show remarkable performances in downstream tasks. |
| Approach: | They propose a method of efficient knowledge transfer that integrates pre-trained uni-modal models into a combined vision-language model without pre-training . they propose to fine-tune the model and transfer multimodal knowledge from a teacher vision-linguistic model to the CVLM for each task application. |
| Outcome: | The proposed method outperforms existing vision-language models in downstream tasks. |
Are Compressed Language Models Less Subgroup Robust? (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to reduce model size and latency while retaining overall performance are not known about their impact on subgroup robustness. |
| Approach: | They investigate the effects of model compression on subgroup robustness of BERT language models. |
| Outcome: | The proposed compression methods do not worsen the performance on minority subgroups. |
On the Way to Lossless Compression of Language Transformers: Exploring Cross-Domain Properties of Quantization (2024.lrec-main)
Copied to clipboard
Nikita Martynov, Aleksei Goncharov, Gleb Kumichev, Evgeniy Egorov, Stanislav Vladimirovich Pavlov, Mikhail Sergeevich Durinov, Aleksandr Sergeevich Zuev, Egor Anatolievich Filimonov
| Challenge: | Modern Natural Language Processing models have a huge capacity, but this makes it difficult to employ. |
| Approach: | They propose a method to quantize at least 95% of Transformer weights without access to task-specific data so the drop in performance does not exceed 0.02%. |
| Outcome: | The proposed method quantizes 95% of Transformer weights and corresponding activations to INT8 without access to task-specific data so the drop in performance does not exceed 0.02%. |
ClusComp: A Simple Paradigm for Model Compression and Efficient Finetuning (2025.findings-acl)
Copied to clipboard
| Challenge: | Weight-only quantization reduces model size but suffers from performance degradation at lower bit widths. |
| Approach: | They propose a weight-only quantization paradigm that clusters weight matrices into codebooks and finetunes them block-by-block. |
| Outcome: | The proposed paradigm outperforms quantization methods and fine tunes LLMs to 1-bit compression and fine tuning. |
GRASP: Replace Redundant Layers with Adaptive Singular Parameters for Efficient Model Compression (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have demonstrated that many layers are functionally redundant in large language models (LLMs), enabling model compression by removing these layers to reduce inference cost. |
| Approach: | They propose a framework that removes redundant layers to reduce inference cost by preserving sensitivity-aware singular values. |
| Outcome: | The proposed framework outperforms existing methods in 90% of the original model under a 20% compression ratio. |
Automated Fine-Grained Mixture-of-Experts Quantization (2025.findings-acl)
Copied to clipboard
| Challenge: | specialized quantization framework for Mixture of Experts architectures is inadequate for model compression. |
| Approach: | They propose a specialized quantization framework for Mixture of Experts architectures . they find that expert networks exhibit distinctive channel-wise outlier distributions ." |
| Outcome: | The proposed framework improves on the Mixtral-8x7b-v0.1 architecture while maintaining minimal computational overhead. |
Break Through the Compression Bottleneck: From Theory to Practice (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing compression methods suffer from bottleneck issues when compression ratio is increased. |
| Approach: | They propose a novel approach to combine low-rank decomposition and quantization methods to reduce the compression bottleneck. |
| Outcome: | The proposed method reduces the computational and memory overhead of existing methods while maintaining model accuracy. |