Papers with memory footprint
Iterative Structured Pruning for Large Language Models with Multi-Domain Calibration (2026.eacl-industry)
Copied to clipboard
| Challenge: | Existing models with unstructured pruning often yield irregular sparsity patterns that necessitate specialized hardware or software support. |
| Approach: | They propose a structured pruning framework that eliminates entire architectural components and maintains compatibility with standard hardware accelerators. |
| Outcome: | The proposed model pruning framework achieves significant compression with minimal performance degradation on multiple models across diverse downstream tasks. |
Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models (2024.acl-long)
Copied to clipboard
Zhengxin Zhang, Dan Zhao, Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Qing Li, Yong Jiang, Zhihao Jia
| Challenge: | Existing methods to finetun large language models (LLMs) only update a small number of trainable parameters, or attempt to reduce the memory footprint during the training phase of the finetune process. |
| Approach: | They propose quantized side tuing (QST) which quantizes an LLM’s model weights into 4-bit to reduce the memory footprint of the original weights. |
| Outcome: | The proposed method reduces the memory footprint of the model weights, optimizer states, and intermediate activations while reducing the memory requirements. |
Easy and Efficient Transformer: Scalable Inference Solution For Large NLP Model (2022.naacl-industry)
Copied to clipboard
Gongzheng Li, Yadong Xi, Jingzhen Ding, Duan Wang, Ziyang Luo, Rongsheng Zhang, Bai Liu, Changjie Fan, Xiaoxi Mao, Zeng Zhao
| Challenge: | Recent studies show that transformer-based models are effective over many tasks, but they are expensive to deploy in the industrial application. |
| Approach: | They propose a transformer-based inference solution that optimizes kernels for long inputs and large hidden sizes and a flexible CUDA memory manager to reduce the memory footprint when deploying a large model. |
| Outcome: | The proposed solution achieves an average speedup of 1.40-4.20x on the transformer decoder layer with an A100 GPU. |
Unified Examination of Entity Linking in Absence of Candidate Sets (2024.naacl-short)
Copied to clipboard
| Challenge: | Entity linking systems depend on candidate sets for their performance, but a comprehensive comparative analysis of these systems is lacking. |
| Approach: | They propose a black-box benchmark and a method to evaluate all state-of-the-art entity linking methods. |
| Outcome: | The proposed approach reduces the inference time and memory footprint of some models. |
Generation-Distillation for Efficient Natural Language Understanding in Low-Data Settings (D19-61)
Copied to clipboard
| Challenge: | Recent research points to knowledge distillation as a potential solution for NLU tasks. |
| Approach: | They propose a training approach that distills large finetuned LMs into a small network using unlabeled training examples. |
| Outcome: | The proposed approach outperforms BERT training approaches while using 300 times fewer parameters. |
LightSeq: A High Performance Inference Library for Transformers (2021.naacl-industry)
Copied to clipboard
| Challenge: | Existing inference frameworks for natural language processing are not the best choice for online service of sequence processing problems. |
| Approach: | They propose a highly efficient inference library for Transformer models that includes GPU optimization techniques to streamline computation and reduce memory footprint. |
| Outcome: | The proposed library achieves 14x speedup compared with TensorFlow and 1.4x speed up compared to a concurrent CUDA implementation. |
Adaptive Attention Span in Transformers (P19-1)
Copied to clipboard
| Challenge: | We extend the maximum context size of a neural network called Transformer to 8k characters. |
| Approach: | They propose a self-attention mechanism that can learn its optimal attention span . this allows for models with longer context and the capability to catch longer dependencies. |
| Outcome: | The proposed model achieves state-of-the-art performance on text8 and enwiki8 using 8k characters with no loss of performance, and maintains control over memory footprint and computational time. |
Compressing and Debiasing Vision-Language Pre-Trained Models for Visual Question Answering (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on VQA models have found that they suffer from dataset biases and inefficient memory footprints. |
| Approach: | They investigate whether a VLP can be compressed and debiased simultaneously by searching sparse and robust subnetworks. |
| Outcome: | The proposed compression and debiasing pipelines outperform the debiased full VLPs on VQA tasks. |
Multi-CLS BERT: An Efficient Alternative to Traditional Ensembling (2023.acl-long)
Copied to clipboard
| Challenge: | ensembling BERT models often improves accuracy but at the cost of significantly more computation and memory footprint. |
| Approach: | They propose a new ensembling method for CLS-based prediction tasks that is almost as efficient as a single BERT model. |
| Outcome: | The proposed method outperforms existing BERT models on GLUE and SuperGLUE with 100 training samples. |
Memory-Efficient Backpropagation for Fine-Tuning LLMs on Resource-Constrained Mobile Devices (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Existing work on memory-efficient on-device fine-tuning of large language models with backpropagation has focused on approximating gradients with zeroth-order optimization (ZO). |
| Approach: | They propose a memory-efficient implementation of backpropagation on mobile devices that allows flexible trade-offs between memory usage and compute time while converging faster. |
| Outcome: | The proposed method can fine-tune LLMs with backpropagation using less than 1GB of memory while achieving better performance than the baseline. |
CTC-based Compression for Direct Speech Translation (2021.eacl-main)
Copied to clipboard
| Challenge: | Existing studies have shown that a dynamic phone-informed compression of the input audio is beneficial for speech translation (ST). |
| Approach: | They propose a method which performs a phone-informed compression of the input audio in direct ST models by exploiting the Connectionist Temporal Classification (CTC) they demonstrate that their method brings a 1.3-1.5 BLEU improvement over a strong baseline on two language pairs (English-Italian and English-German) |
| Outcome: | The proposed method brings a 1.3-1.5 BLEU improvement over a strong baseline on two language pairs (English-Italian and English-German) it reduces memory footprint by more than 10%, and is faster than previous approaches. |
ZoomR: Memory Efficient Reasoning through Multi-Granularity Key Value Retrieval (2026.acl-long)
Copied to clipboard
David H. Yang, Yuxuan Zhu, Mohammad Mohammadi Amiri, Keerthiram Murugesan, Tejaswini Pedapati, Subhajit Chaudhury, Pin-Yu Chen
| Challenge: | Large language models generate long chain of thoughts but memory footprint grows with output length . prior work on KV cache optimization focused on compressing long input context . |
| Approach: | They propose a new approach that compresses verbose reasoning thoughts into summaries . they use a dynamic KV cache selection policy that leverages these summary keys . |
| Outcome: | The proposed approach reduces memory usage while avoiding full-cache attention at each step. |
ELLA: Efficient Lifelong Learning for Adapters in Large Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to training Large Language Models (LLMs) suffer from catastrophic forgetting when adapted sequentially to new tasks in a continual learning (CL) setting. Existing methods are impractical and could potentially violate privacy. |
| Approach: | They propose a training framework built on the principle of selective subspace de-correlation that characterizes the structure of past updates and penalizes alignments along their high-energy, task-specific directions. |
| Outcome: | The proposed training framework achieves state-of-the-art CL performance on three popular benchmarks spanning both classification and generative tasks with relative accuracy gains of up to 9.6% and a 35 smaller memory footprint. |
Unlocking Data-free Low-bit Quantization with Matrix Decomposition for KV Cache Compression (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods to compress KV cache compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. |
| Approach: | They propose a low-bit quantization technique based on tensor decomposition to effectively compress KV cache. |
| Outcome: | The proposed method reduces memory footprint and performance by 75% . it is compared with existing methods that compromise precision or require extra data for calibration . |
Masking as an Efficient Alternative to Finetuning for Pretrained Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that our binary masked language models encode information necessary for solving downstream tasks. |
| Approach: | They propose an efficient method of utilizing pretrained language models where selective binary masks are learned instead of finetuning. |
| Outcome: | Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that the proposed method yields comparable performance to finetuning, but has a much smaller memory footprint when multiple tasks need to be solved. |
Multimodal Phased Transformer for Sentiment Analysis (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to model multimodal sentiment analysis are limited due to their complexity and memory footprint. |
| Approach: | They propose a multimodal Sparse Phased Transformer to reduce self-attention complexity and memory footprint. |
| Outcome: | The proposed method achieves comparable or superior performance with a 90% reduction in the number of parameters. |
Efficient Large-Scale Neural Domain Classification with Personalized Attention (P18-1)
Copied to clipboard
| Challenge: | Using a scalable neural model, we show that personalization improves domain classification accuracy in a setting with thousands of overlapping domains. |
| Approach: | They propose a scalable neural model architecture with a shared encoder that incorporates personalization information and domain-specific classifiers that solves the problem efficiently. |
| Outcome: | The proposed architecture achieves two orders of magnitude faster than full model retraining. |
LittleBird: Efficient Faster & Longer Transformer for Question Answering (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing models for BERT have a limitation dealing with long inputs due to its attention mechanism. |
| Approach: | They propose a model based on BigBird with improved speed and memory footprint . they propose 'pack and unpack attention' to replace global attention . |
| Outcome: | The proposed model can work on long inputs even after being pre-trained on short inputs. |
SCA: Selective Compression Attention for Efficiently Extending the Context Window of Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to compress the KV cache of large language models are expensive and limited in their context window and cost. |
| Approach: | They propose a method to expand the context window and reduce memory footprint by compressing the KV cache of large language models. |
| Outcome: | The proposed method can reduce memory footprint and expand context window of large language models without training. |
SwapMoE: Serving Off-the-shelf MoE-based Large Language Models with Tunable Memory Budget (2024.acl-long)
Copied to clipboard
| Challenge: | Mixture of experts (MoE) is a popular technique to improve capacity of Large Language Models (LLMs) but memory-constrained devices are a major concern in edge AI training and serving. |
| Approach: | They propose a framework for efficient serving of MoE-based large language models with tunable memory budgets. |
| Outcome: | Experiments show that SwapMoE can reduce memory consumption while maintaining reasonable accuracy. |
Token and Head Adaptive Transformers for Efficient Natural Language Processing (2022.coling-1)
Copied to clipboard
| Challenge: | Pre-trained language models like BERT have shown significant accuracy improvements on various tasks, but their computational cost and memory footprint are prohibitive. |
| Approach: | They propose to extend Length Adaptive Transformer to extend the model to a token and head pruning scheme to optimize pruning efficiency. |
| Outcome: | The proposed model can compress and accelerate BERT-based models by fine-tuning and a token and head pruning scheme. |
Seq1F1B: Efficient Sequence-Level Pipeline Parallelism for Large Language Model Training (2025.naacl-long)
Copied to clipboard
| Challenge: | Current PP methods face severe bottlenecks, including pipeline bubbles and memory footprint. |
| Approach: | They propose a sequence-level one-forward-one-backward (1F1B) PP method for training LLMs on long sequences with high throughput and memory efficiency. |
| Outcome: | The proposed method achieves 1.14X training throughput with half memory footprint compared to baseline methods . it trains an LLM with 30B parameters on sequences up to 64k tokens using 64X NVIDIA A100 GPUs . |
Do Emergent Abilities Exist in Quantized Large Language Models: An Empirical Study (2024.lrec-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) require significant computational resources for deployment and use. |
| Approach: | They propose to use low-bit quantization methods to reduce memory footprint and increase inference rate to improve performance of Large Language Models. |
| Outcome: | The proposed methods can reduce the memory footprint and increase the inference rate of LLMs. |
Revisiting Knowledge Distillation for Autoregressive Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Autoregressive language models (LMs) are expensive and memory intensive, preventing the development of industrial applications. |
| Approach: | They propose an adaptive teaching approach to improve the KD of autoregressive language models by distilling knowledge into a small student model. |
| Outcome: | The proposed method can achieve consistent and significant performance gains across all model types and sizes. |
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. |
PuMer: Pruning and Merging Tokens for Efficient Vision Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Large-scale vision language models use Transformers to perform cross-modal interactions . state-of-the-art models are memory intensive and expensive due to quadratic complexity . |
| Approach: | They propose a token reduction framework that uses text-informed Pruning and modality-aware Merging strategies to progressively reduce the tokens of input image and text. |
| Outcome: | The proposed framework improves inference speed and memory footprint on four vision language tasks. |
HyHTM: Hyperbolic Geometry-based Hierarchical Topic Model (2023.findings-acl)
Copied to clipboard
| Challenge: | Hierarchical Topic Models (HTMs) often produce hierarchies where lower-level topics are unrelated and not specific enough to their higher-level subjects. |
| Approach: | They propose a Hyperbolic geometry-based Hierarchical Topic Model that incorporates hierarchical information from hyperbolic geometrics to explicitly model hierarchies in topic models. |
| Outcome: | The proposed model is significantly faster and leaves a much smaller memory footprint than the best-performing baseline. |
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. |
Balancing Fidelity and Plasticity: Aligning Mixed-Precision Fine-Tuning with Linguistic Hierarchies (2026.findings-acl)
Copied to clipboard
Changhai Zhou, Shiyang Zhang, Yuhua Zhou, Jun Gao, Qian Qiao, Shichao Weng, Weizhong Zhang, Cheng Jin
| Challenge: | Existing quantization-aware fine-tuning methods decouple weight precision and adapter capacity, overlooking that a layer’s ability to adapt is constrained by the information preserved in its frozen weights. |
| Approach: | They propose a framework that jointly optimizes per-layer quantization bit-width and LoRA rank. |
| Outcome: | Experiments on LLaMA and Qwen models show that the proposed framework matches or approaches 16-bit baselines while using substantially less memory. |
Context Compression for Auto-regressive Transformers with Sentinel Tokens (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing Transformer-based LLMs have limited performance due to complexity of attention module . key-value cache is the major memory footprint and inference latency problem . |
| Approach: | They propose a plug-and-play approach that incrementally compresses token activation into compact ones . they also profile the benefit of context compression on improving the system throughout . |
| Outcome: | The proposed approach reduces memory footprint and inference latency by compressing tokens into compact ones. |
Pseudo-Relevance for Enhancing Document Representation (2022.emnlp-main)
Copied to clipboard
| Challenge: | a novel approach to document retrieval can be used to encode documents as vectors . a few query-relevant terms can be pruned out to reduce index overhead . |
| Approach: | They propose to enhance the document representation for the bi-encoder approach in dense document retrieval. |
| Outcome: | The proposed solution reduces latency and memory footprint up to 8- and 3-fold . it is validated on MSMARCO and real-world search query logs . |
Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings (2025.findings-acl)
Copied to clipboard
Yubo Ma, Jinsong Li, Yuhang Zang, Xiaobao Wu, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Haodong Duan, Jiaqi Wang, Yixin Cao, Aixin Sun
| Challenge: | Visualized Document Retrieval (VDR) uses large vision-language models to encode document pages into embeddings. |
| Approach: | They evaluate methods to reduce patch embeddings per page while minimizing performance degradation. |
| Outcome: | The proposed method maintains 98.2% of retrieval performance with only 11.8% of original memory usage and preserves 94.6% effectiveness at 2% memory footprint. |
CARVQ: Corrective Adaptor with Group Residual Vector Quantization for LLM Embedding Compression (2025.findings-emnlp)
Copied to clipboard
Dayin Gou, Sanghyun Byun, Nilesh Malpeddi, Gabrielle De Micheli, Prathamesh Vaste, Jacob Song, Woo Seong Chung
| Challenge: | Large Language Models typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. |
| Approach: | They propose a corrective Adaptor with group Residual Vector Quantization that can be used to compress the embedding layer without requiring specialized hardware. |
| Outcome: | The proposed corrective adaptor can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. |
SADA: Bridging In-Context Learning and Fine-Tuning via State-Aligned Distillation Adapters (2026.acl-long)
Copied to clipboard
| Challenge: | Prompt-based in-context learning and parameter fine-tuning are dominant paradigms for incorporating external information into large language models, but they incur high inference costs or require expensive retraining. |
| Approach: | They propose to convert prompts into temporary adapter weights to bridge this gap by converting prompts to temporary adapters. |
| Outcome: | The proposed model outperforms baselines on long-context language modeling and downstream NLU and summarization benchmarks while significantly reducing memory footprint and latency. |
Memory-Efficient Fine-Tuning of Transformers via Token Selection (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning require caching of intermediate activations to update weights during the backward pass. |
| Approach: | They develop a method to reduce memory usage in fine-tuning of transformers by backpropagating through just a subset of input tokens. |
| Outcome: | The proposed method reduces memory usage and memory footprint on large transformer models . it can be easily combined with existing methods like LoRA, reducing memory cost . |
TALAS: Teacher-Anchored Layer Alignment with Adaptive Sharpness-Aware Minimization for Embedding Distillation (2026.acl-long)
Copied to clipboard
Quoc Phong Dao, Hoang Son Nguyen, Pham Khanh Chi, Linh Ngo Van, Nguyen Thi Ngoc Diep, Thien Huu Nguyen, Trung Le
| Challenge: | Existing methods that require a student to strictly mimic the teacher’s sentence embeddings or internal features often incur prohibitive computational costs and yield suboptimal performance due to the inherent capacity gap. |
| Approach: | They propose a Teacher-Anchored mechanism that selectively distills final sentence embeddings only into the student’s upper layers, thereby reducing overhead while respecting capacity constraints. |
| Outcome: | Empirical results show that TALAS outperforms existing methods while maintaining high performance. |
Linear-Time and Constant-Memory Text Embeddings Based on Recurrent Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing work on recurrent models for text embedding is limited to small task-specific models. |
| Approach: | They propose a vertically chunked inference strategy that enables fast embedding generation with memory usage that becomes constant in the input length once it exceeds the vertical chunk size. |
| Outcome: | The proposed architectures achieve competitive performance across benchmarks while maintaining a substantially smaller memory footprint compared to transformer-based models. |
ACBQ: Adaptive Cross-Block Quantization of Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for post-training quantization struggle to support weight–activation joint quantization and extreme low-bit weight quantization. |
| Approach: | They propose a framework that addresses weight–activation joint quantization and extreme weight quantization. |
| Outcome: | The proposed framework achieves superior performance under both W4A4 and highly aggressive W2 settings while incurring negligible additional computational overhead. |