Papers with memory requirements
Efficient Content-Based Sparse Attention with Routing Transformers (2021.tacl-1)
Copied to clipboard
| Challenge: | Self-attention suffers from quadratic computation and memory requirements with respect to sequence length . despite its effectiveness, self-attention models suffer from quadratic computation and a limited set of locations . |
| Approach: | They propose to learn dynamic sparse attention patterns that avoid allocating computation and memory to attend to content unrelated to the query of interest. |
| Outcome: | The proposed model outperforms similar sparse attention models on language modeling and image generation on Wikitext-103 . |
EffEval: A Comprehensive Evaluation of Efficiency for MT Evaluation Metrics (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a recent surge of interest in developing evaluation metrics based on pretrained large language models (LLMs) can better cope with lexical variation. |
| Approach: | They propose to replace computation-intensive transformers with lighter alternatives and employ linear and quadratic approximations for alignment algorithms on top of LLM representations. |
| Outcome: | The proposed approach replaces computation-intensive transformers with lighter alternatives and employs linear and quadratic approximations for alignment algorithms on top of LLM representations. |
Divide, Reweight, and Conquer: A Logit Arithmetic Approach for In-Context Learning (2026.eacl-long)
Copied to clipboard
| Challenge: | In-Context Learning (ICL) is a key feature for Large Language Models (LLMs) but it faces challenges when dealing with increasing numbers of examples due to performance degradation and quadratic computational costs. |
| Approach: | They propose a Logit Arithmetic Reweighting Approach that uses logit-based ensembling of multiple demonstrations to enhance ICL. |
| Outcome: | The proposed framework outperforms baseline methods in accuracy and memory efficiency. |
JORA: JAX Tensor-Parallel LoRA Library for Retrieval Augmented Fine-Tuning (2024.acl-demos)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) face significant memory constraints when fine-tuning large prompt sequences. |
| Approach: | They propose a framework for PEFT-compatible fine-tuning of large language models, leveraging distributed training. |
| Outcome: | The proposed framework improves performance 12x compared to Hugging Face/DeepSpeed implementation with four GPUs while consuming less than half the VRAM per GPU. |
LinguaLinked: Distributed Large Language Model Inference on Mobile Devices (2024.acl-demos)
Copied to clipboard
| Challenge: | Recent research shows that large language models demonstrate enhanced capabilities in various language tasks. |
| Approach: | They introduce a system for decentralized, distributed LLM inference on mobile devices . they use optimized model assignment technique to segment LLMs and linear optimization to align segments with each device . |
| Outcome: | The proposed system performs well on high-end to low-end Android devices. |
Data Pruning for Efficient Model Pruning in Neural Machine Translation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large-scale pre-trained language models have demonstrated encouraging performance in various NLP tasks at the cost of over-parametrized networks and high memory requirements. |
| Approach: | They combine data pruning with movement pruning for Neural Machine Translation to enable efficient fine-pruning by leveraging cross-entropy scores of individual training instances. |
| Outcome: | The proposed pruning strategy outperforms other pruning methods on a translation task and shows that training cross-entropy scores can reduce the steps required for convergence and training time. |
Efficient Learning of Multiple NLP Tasks via Collective Weight Factorization on BERT (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to fine-tune a model for multiple tasks require a large amount of memory and computing power. |
| Approach: | They propose to factorize the weighs of a pre-trained Transformer model to improve training efficiency across multiple tasks by using BERT-Large as an instantiation of the Transformer and the GLUE as the evaluation benchmark. |
| Outcome: | The proposed method matches or improves the original fine-tuned model’s performance for each task while effectively decreasing parameter requirements by two orders of magnitude. |
Rethinking Token Reduction for State Space Models (2024.emnlp-main)
Copied to clipboard
Zheng Zhan, Yushu Wu, Zhenglun Kong, Changdi Yang, Yifan Gong, Xuan Shen, Xue Lin, Pu Zhao, Yanzhi Wang
| Challenge: | Existing methods for token reduction for SSMs lead to performance drops . a recent study shows that Mamba-2 improves the accuracy of the model by 5.7% to 13.1% . |
| Approach: | They propose a token reduction method that integrates token importance and similarity into SSMs and takes advantage of pruning and merging. |
| Outcome: | The proposed method improves accuracy by 5.7% to 13.1% on six benchmarks with Mamba-2 compared to existing methods while reducing computational demands and memory requirements. |
Sequence Shortening for Context-Aware Machine Translation (2024.findings-eacl)
Copied to clipboard
| Challenge: | Context-aware Machine Translation aims to improve translations of sentences by incorporating surrounding sentences as context. |
| Approach: | They propose to use latent representation of source sentence as context in a multi-encoder architecture to achieve higher accuracy on contrastive datasets. |
| Outcome: | The proposed architectures achieve comparable BLEU and COMET scores on contrastive datasets and comparable accuracies on the single- and multi-encoder approaches. |
Memory-efficient Temporal Moment Localization in Long Videos (2023.eacl-main)
Copied to clipboard
| Challenge: | Temporal Moment Localization is a multi-modal task that requires understanding the temporal relationships in the entire input video. |
| Approach: | They propose a stochastic sampling module that can process long videos at a constant memory footprint. |
| Outcome: | The proposed model can process videos as long as 18 minutes at a constant memory footprint and achieves faster and faster results than competing models. |
LSDC: An Efficient and Effective Large-Scale Data Compression Method for Supervised Fine-tuning of Large Language Models (2025.findings-naacl)
Copied to clipboard
Zhaoguang Long, Yuhao Zhou, Shangqing Zhao, Yupei Ren, Li Cai, Chenghao Jia, Zhe Chen, Zhe Fang, Yuxiang Song, Man Lan
| Challenge: | Large Language Models (LLMs) are expanding in scale and size, increasing computational costs . large-scale data compression techniques can reduce the size of training datasets while maintaining data integrity. |
| Approach: | They propose a large-scale data compression method to reduce the size of training data . they use a bifurcated quantization strategy to maximize the diversity of samples . |
| Outcome: | The proposed method significantly reduces the size of training data while maximizing the submodular gain. |
Wine is not v i n. On the Compatibility of Tokenizations across Languages (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained language models are used for performance and memory constraints, but there is little work that investigates the compatibility of tokenizations across languages. |
| Approach: | They propose a compatibility measure that reflects compatibility of tokenizations across languages. |
| Outcome: | The proposed measure prevents incompatible tokenizations, e.g., “wine” in English vs. “v i n” in French, which make it hard to learn good multilingual semantic representations. |
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. |
Game-theoretic Vocabulary Selection via the Shapley Value and Banzhaf Index (2021.naacl-main)
Copied to clipboard
| Challenge: | Using the full vocabulary results in less explainable and memory intensive models. |
| Approach: | They propose a vocabulary selection method that views words as members of a team trying to maximize the model's performance. |
| Outcome: | The proposed method outperforms baseline models on multiple tasks and datasets. |
Simple and Effective Unsupervised Redundancy Elimination to Compress Dense Vectors for Passage Retrieval (2021.emnlp-main)
Copied to clipboard
| Challenge: | Dense passage retrieval improves ranking accuracy in open-domain question answering but at the cost of large space and memory requirements. |
| Approach: | They propose a simple unsupervised pipeline that includes principal component analysis (PCA), product quantization, and hybrid search to improve space efficiency. |
| Outcome: | The proposed pipeline achieves good accuracy–space trade-offs, for example, 48 compression with less than 3% drop in top-100 retrieval accuracy on average or 96 compression without drop in space requirements. |
Teacher and Student Models of Offensive Language in Social Media (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to identify offensive language online use large pre-trained transformer models. however, the inference time, disk, and memory requirements of these models are prohibitively large. |
| Approach: | They propose to transfer knowledge from large transformer models to much smaller neural models to make predictions at the token- and post-level. |
| Outcome: | The proposed model performs 100 times better than transformer models but with 100 times less parameters and much less memory usage. |
LLaMAs Have Feelings Too: Unveiling Sentiment and Emotion Representations in LLaMA Models Through Probing (2025.acl-long)
Copied to clipboard
Dario Di Palma, Alessandro De Bellis, Giovanni Servedio, Vito Walter Anelli, Fedelucio Narducci, Tommaso Di Noia
| Challenge: | Large Language Models (LLMs) have become central to NLP, demonstrating their ability to adapt to various tasks through prompting techniques. |
| Approach: | They probe the hidden layers of Large Language Models to identify where sentiment features are most represented and to assess how this affects sentiment analysis. |
| Outcome: | The proposed approach enables sentiment tasks to be performed with memory requirements reduced by an average of 57%. |
Scaling Within Document Coreference to Long Texts (2021.findings-acl)
Copied to clipboard
Raghuveer Thirukovalluru, Nicholas Monath, Kumar Shridhar, Manzil Zaheer, Mrinmaya Sachan, Andrew McCallum
| Challenge: | Existing end-to-end coreference resolution models use expensive span representations and antecedent prediction mechanisms. |
| Approach: | They propose an approximation to end-to-end coreference resolution models which scales gracefully to documents of any length. |
| Outcome: | The proposed model reduces training and inference time and memory costs compared to current models with minimal loss in accuracy. |
SlimFit: Memory-Efficient Fine-Tuning of Transformer-based Models Using Training Dynamics (2024.naacl-long)
Copied to clipboard
Arash Ardakani, Altan Haan, Shangyin Tan, Doru Thom Popovici, Alvin Cheung, Costin Iancu, Koushik Sen
| Challenge: | SlimFit reduces the memory requirements of transformer-based models by analyzing their training dynamics and freezing less-contributory layers during fine-tuning. |
| Approach: | They propose a tool that analyzes transformer-based models and freezes less-contributory layers during fine-tuning to reduce the overall on-device memory usage. |
| Outcome: | SlimFit reduces the memory requirements of transformer-based models by analyzing their training dynamics and freezing less-contributory layers during fine-tuning. |
Data-Informed Global Sparseness in Attention Mechanisms for Deep Neural Networks (2024.lrec-main)
Copied to clipboard
| Challenge: | Attention pruning techniques have been developed to identify and exploit sparseness . previous work has taken pioneering steps to discover and explain the sparsity in attention patterns . |
| Approach: | They propose a framework that observes attention patterns in a fixed dataset and generates a global sparseness mask. |
| Outcome: | The proposed approach saves 90% of computations and maintains quality of results. |
Natural Logic-guided Autoregressive Multi-hop Document Retrieval for Fact Verification (2022.emnlp-main)
Copied to clipboard
| Challenge: | Recent evidence retrieval approaches rely on heuristics and assume hyperlinks between documents. |
| Approach: | They propose a retrieval method that combines a retriever and a proof system that reranks documents and reorders them . |
| Outcome: | The proposed method exceeds or is on par with the current state-of-the-art on FEVER, HoVer and FEVEROUS-S while using 5 to 10 times less memory than competing systems. |
VPTQ: Extreme Low-bit Vector Post-Training Quantization for Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent research has focused on pushing weight-only quantization to extremely low-bit due to numerical representation limitations. |
| Approach: | They propose a vector-based quantization approach that pushes LLMs to extremely low-bit . they propose scalar-based weight quantization that reduces memory requirements and optimizes storage costs . |
| Outcome: | The proposed method reduces model quantization perplexity by 0.01-0.34 on LLaMA-2, 0.38-0.68 on mistral-7B, 4.41-7.34, on llaMA-3 on QA tasks on average. |
Does quantization affect models’ performance on long-context tasks? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models support context windows exceeding 128K tokens, but this comes with significant memory requirements and high inference latency. |
| Approach: | They present the first systematic evaluation of quantized LLMs on tasks with long inputs and long-form outputs. |
| Outcome: | The proposed method preserves accuracy, while 4-bit methods lead to substantial losses . the results highlight the importance of a careful evaluation before deploying quantized LLMs . |
XQuant: Achieving Ultra-Low Bit KV Cache Quantization with Cross-Layer Compression (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks. however, their extensive memory requirements present significant challenges for deployment in resource-constrained environments. |
| Approach: | They propose a training-free framework that achieves ultra-low equivalent bit-width KV cache quantization. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on TruthfulQA and LongBench. |
EfficientQAT: Efficient Quantization-Aware Training for Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Quantization-aware training (QAT) is a low-bit training solution that requires substantial training resources. |
| Approach: | They propose an algorithm that reduces memory consumption by low-bit representations with minimal accuracy loss. |
| Outcome: | EfficientQAT achieves 2-bit Llama-2-70B model on single GPU in 41 hours . compared to previous methods, it obtains model with less than 3 points accuracy degradation . |
SSH: Sparse Spectrum Adaptation via Discrete Hartley Transformation (2025.naacl-long)
Copied to clipboard
| Challenge: | Low-rank adaptation (LoRA) has been demonstrated effective in reducing the trainable parameter number when fine-tuning a large foundation model (LLM). |
| Approach: | They propose a low-rank adaptation approach that reduces the number of trainable parameters while enhancing model performance. |
| Outcome: | The proposed approach outperforms existing parameter-efficient fine-tuning methods while achieving substantial reductions in computational cost and memory requirements. |
LPC: A Logits and Parameter Calibration Framework for Continual Learning (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to solve catastrophic forgetting problem are varied . current approaches to learn continuous learning are based on replay-based methods . |
| Approach: | They propose to calibrate parameters and logits so that preserving old parameters and generalized learning on new concepts can be solved simultaneously. |
| Outcome: | The proposed model achieves state-of-the-art performance in all scenarios. |
Is the Brain Mechanism for Hierarchical Structure Building Universal Across Languages? An fMRI Study of Chinese and English (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown that the brain builds hierarchical syntactic structures, but it is unknown whether they are universal across languages. |
| Approach: | They analyze the working memory requirements when applying parsing strategies to two languages: Chinese and English. |
| Outcome: | The proposed method shows that the brain adopts parsing strategies with less memory load according to different language structures. |
HashFormers: Towards Vocabulary-independent Pre-trained Transformers (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing pre-trained language models are vocabulary-dependent, mapping by default each token to its corresponding embedding. |
| Approach: | They propose a family of vocabulary-independent pre-trained transformers that support unlimited vocabulary . they propose to map each token to its corresponding embedding by default . |
| Outcome: | The proposed models are more memory efficient than existing models while achieving comparable performance on multiple text classification tasks. |
EasyQuant: An Efficient Data-free Quantization Algorithm for LLMs (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent work has shown that large language models are superior to conventional methods in various tasks. |
| Approach: | They propose a data-independent quantization algorithm that leaves outliers in the weight and quantization ranges . they find the algorithm runs over 10 times faster than the data-dependent methods . |
| Outcome: | The proposed method runs over 10 times faster than the data-dependent methods. |
Dynamic-prototype Contrastive Fine-tuning for Continual Few-shot Relation Extraction with Unseen Relation Detection (2025.coling-main)
Copied to clipboard
| Challenge: | Existing approaches to learn relations from labeled data overlook task interference in continual learning and memory requirements for different relations. |
| Approach: | They propose a framework to learn new relations from limited labeled data while preserving knowledge about previously learned relations. |
| Outcome: | The proposed framework is more practical and comprehensive for real-world scenarios. |
Continual Quantization-Aware Pre-Training: When to transition from 16-bit to 1.58-bit pre-training for BitNet language models? (2025.findings-acl)
Copied to clipboard
| Challenge: | Quantization-aware training of large language models reduces the precision of model parameters and reduces memory usage and energy consumption at inference time. |
| Approach: | They propose a method where models are first trained with 16-bit precision and then transition to 1.58-bit quantization-aware training. |
| Outcome: | The proposed training strategy reduces memory and energy consumption while maintaining model accuracy while reducing memory and inference time. |
Pit One Against Many: Leveraging Attention-head Embeddings for Parameter-efficient Multi-head Attention (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing pre-trained language models have produced performance gains in various tasks but come with large computational requirements. |
| Approach: | They propose an alternative module that uses only a single shared projection matrix and multiple head embeddings (MHE) they demonstrate that MHE attention is substantially more memory efficient compared to alternative attention mechanisms. |
| Outcome: | The proposed model is more memory efficient compared to the current model while achieving high retention ratio on several downstream tasks. |
AdaLomo: Low-memory Optimization with Adaptive Learning Rate (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models require substantial memory for training, thereby setting a high hardware threshold. |
| Approach: | They propose a low-memory optimization technique that reduces memory footprint . they propose an adaptive learning rate for each parameter and a grouped update normalization to stabilize convergence . |
| Outcome: | The proposed low-memory optimization performs better than the prevailing algorithm for large language models, AdamW. |
Angular Dispersion Accelerates k-Nearest Neighbors Machine Translation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | k-nearest neighbors machine translation is a well-established strategy for increasing translation performance . high computational cost and memory requirements make it difficult to perform exhaustive searches in large data stores . |
| Approach: | They propose to augment neural machine translation with external memory at decoding time . they propose to encourage angular dispersion of the neural hidden representations of contexts . |
| Outcome: | The proposed method is expensive and high memory-intensive . it is based on the performance properties of k-NN lookup data structures. |
Train Once, Deploy Anywhere: Matryoshka Representation Learning for Multimodal Recommendation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | integrating rich multimodal knowledge into recommender systems remains a challenge . despite performance improvements, different recommendation scenarios often require varying granularities. |
| Approach: | They propose a framework that captures item features at different granularities and learns informative representations for efficient recommendation across multiple dimensions. |
| Outcome: | The proposed framework achieves superior performance over state-of-the-art models on multiple benchmark datasets. |
SparseGrad: A Selective Method for Efficient Fine-tuning of MLP Layers (2024.emnlp-main)
Copied to clipboard
Viktoriia Chekalina, Anna Rudenko, Gleb Mezentsev, Aleksandr Mikhalev, Alexander Panchenko, Ivan Oseledets
| Challenge: | High-performance methods for parameter-efficient fine-tuning (PEFT) typically work with Attention blocks and overlook dense MLP blocks, which contain about half of the model parameters. |
| Approach: | They propose a selective PEFT method that performs well on MLP blocks by converting layer gradients into a sparse structure and reducing the number of updated parameters. |
| Outcome: | The proposed method outperforms LoRA and MeProp, robust state-of-the-art PEFT approaches. |
LoNAS: Elastic Low-Rank Adapters for Efficient Large Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) reach hundreds of billions of parameters and require resources for training and inference stages. |
| Approach: | They propose a low-rank adapter to reduce the number of trainable parameters in a model and reduce memory requirements. |
| Outcome: | The proposed approach reduces memory and compute requirements while preserving performance. |
When Compression Meets Model Compression: Memory-Efficient Double Compression for Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit excellent performance in various tasks, but memory requirements present a challenge when deploying on memory-limited devices. |
| Approach: | They propose a framework to compress LLM after quantization further, achieving about 2.2x compression ratio. |
| Outcome: | The proposed model can achieve 40% reduction in memory size with negligible loss in accuracy and inference speed. |
MaCP: Minimal yet Mighty Adaptation via Hierarchical Cosine Projection (2025.acl-long)
Copied to clipboard
| Challenge: | MaCP is a new adaptation method for large foundation models that requires minimal parameters and memory for fine-tuning. |
| Approach: | They propose a method that exploits the superior energy compaction and decorrelation properties of cosine projection to improve model efficiency and accuracy. |
| Outcome: | The proposed method improves model efficiency and accuracy across a wide range of single-modality tasks including natural language understanding, natural language generation, text summarization, and multi-modalities such as image classification and video understanding. |
A Simple and Effective L_2 Norm-Based Strategy for KV Cache Compression (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to reduce the KV cache size involve fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce sequence length. |
| Approach: | They find a correlation between the L2 norm and attention scores over cached KV pairs . they compress the KV cache based on the L1 norm of key embeddings . |
| Outcome: | The proposed approach reduces the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. |
CHESS: Optimizing LLM Inference via Channel-Wise Thresholding and Selective Sparsification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for activation sparsification do not capture the relationship between activation and model performance. |
| Approach: | They propose a general activation sparsification approach using channel-wise thresholding and selective sparsifying to capture the relationship between activation and model performance. |
| Outcome: | The proposed approach reduces the number of activated neurons during inference by 1.27x over eight downstream tasks while activating fewer parameters than existing methods. |
Multi-matrix Factorization Attention (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing variants for Multi-Head Attention (MHA) fail to maintain strong performance under stringent Key-Value cache (KV cache) constraints. |
| Approach: | They propose to use multi-matrix factorization attention and MFA-Key-reuse attention architectures to increase model capacity under tight KV cache constraints. |
| Outcome: | The proposed architecture outperforms existing methods while reducing KV cache usage by 56% and 93.7% in large-scale experiments. |
Scaling Laws and Efficient Inference for Ternary Language Models (2025.acl-long)
Copied to clipboard
Tejas Vaidhya, Ayush Kaushal, Vineet Jain, Francis Couture-Harpin, Prashant Shishodia, Majid Behbahani, Yuriy Nevmyvaka, Irina Rish
| Challenge: | Large language models (LLMs) are increasingly used across research and industry applications, yet their inference efficiency remains a challenge. |
| Approach: | They propose ternary language models that employ quantization-aware training to significantly reduce memory requirements. |
| Outcome: | The proposed ternary language models demonstrate sustained performance gains at scale. |
HydraOpt: Navigating the Efficiency-Performance Trade-off of Adapter Merging (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods that produce a fixed trade-off between storage size and performance are often ineffective due to the growing size of large language models. |
| Approach: | They propose a model merging technique that capitalizes on similarities between low-rank adapters to reduce storage costs and improve performance. |
| Outcome: | The proposed method significantly reduces storage size (48% reduction) while outperforms existing merging techniques in terms of performance (0.2-1.8% drop). |