Challenge: Existing static vocabulary pruning designs that reduce memory usage suffer from rigid, one-size-fits-all designs that cause information loss during the prefill stage and lack flexibility.
Approach: They propose a decoupled dynamic vocabulary selection framework that addresses memory constraints through offloading embedding and implements a hybrid static-dynamic vocabulary selection strategy for LM Head.
Outcome: The proposed framework reduces memory usage by 99% with minimal or no degradation in performance.

Similar Papers

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.
XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models (2023.emnlp-main)

Copied to clipboard

Challenge: Large multilingual models rely on a single vocabulary shared across 100+ languages . this vocabulary bottleneck limits the representational capabilities of multilingual model XLM-R .
Approach: They propose a new approach for scaling to large multilingual vocabularies by de-emphasizing token sharing between languages with little lexical overlap and assigning vocabulary capacity to achieve sufficient coverage for each individual language.
Outcome: The proposed model outperforms XLM-R on all language tasks and is particularly effective on low-resource tasks.
Demystifying Small Language Models for Edge Deployment (2025.acl-long)

Copied to clipboard

Challenge: Small language models (SLMs) are a promising solution for resource-constrained devices such as smartphones and the Web of Things.
Approach: They propose to use SLMs to build and optimize a set of small language models that are publicly accessible.
Outcome: The proposed models outperform 7B models in general tasks, while their in-context learning capabilities remain limited and their efficiency has significant optimization potential.
TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference (2021.naacl-main)

Copied to clipboard

Challenge: Existing pre-trained language models (PLMs) are expensive in inference, making them impractical in resource-limited real-world applications.
Approach: They propose a dynamic token reduction approach to accelerate PLMs' inference by adapting the layer number of each token to avoid redundant calculation.
Outcome: The proposed approach speeds up BERT by 2-5 times and improves performance in long-text tasks with less computation.
SLlama: Parameter-Efficient Language Model Architecture for Enhanced Linguistic Competence Under Strict Data Constraints (2025.emnlp-main)

Copied to clipboard

Challenge: Large-scale language models (LLMs) have shown remarkable performance across a wide array of tasks.
Approach: They propose an architecture that preserves parameter efficiency of tied models without sacrificing representational benefits of untied embeddings.
Outcome: The proposed architecture achieves a 31.72% improvement in linguistic knowledge acquisition over the baseline model.
DisLoRA: Task-specific Low-Rank Adaptation via Orthogonal Basis from Singular Value Decomposition (2025.emnlp-main)

Copied to clipboard

Challenge: Extensive experiments on GLUE and Commonsense Reasoning benchmarks demonstrate that DisLoRA surpasses established PEFT methods, including LoRA, PiSSA, DoRA, LoRA-Dash, and SORSA.
Approach: They propose a framework that leverages singular value decomposition to decompose pretrained weight matrices into orthogonal backbone and task-specific subspaces.
Outcome: Extensive experiments on GLUE and Commonsense Reasoning benchmarks show that DisLoRA surpasses established PEFT methods, including LoRA, PiSSA, DoRA, LoRA-Dash, and SORSA.
Retrofitting Large Language Models with Dynamic Tokenization (2025.acl-long)

Copied to clipboard

Challenge: Current language models use a static tokenizer, which results in degraded efficiency and language capabilities.
Approach: They propose a way to dynamically decide on token boundaries based on input text . they merge frequent subword sequences in a batch and apply a hypernetwork to compute token embeddings .
Outcome: The proposed method reduces token sequence lengths by >20% across 14 languages while promoting fairness across languages.
Scaling Down, Serving Fast: Compressing and Deploying Efficient LLMs for Recommendation Systems (2025.emnlp-industry)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable performance across a wide range of industrial applications.
Approach: They propose two techniques for training and deploying small language models that deliver high performance for a variety of industry use cases.
Outcome: The proposed techniques retain much of the quality of larger models while reducing training/serving costs and latency.
LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin (2024.acl-long)

Copied to clipboard

Challenge: Experimental results show that, as the instruction data increases, LoRAMoE can significantly improve the ability to process downstream tasks, while maintaining the world knowledge stored in the LLM.
Approach: They propose a framework that introduces several low-rank adapters and integrates them by using a router network to freeze the backbone model and force a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks.
Outcome: The proposed framework freezes the backbone model and forces a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks, to alleviate world knowledge forgetting.
ModularMoE: Fast LLM Customization with Parameter-Sharing Mixture-of-Experts for Low-Resource Settings (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models impose significant computational and storage burdens on personal devices . existing customization approaches incur excessive computational costs or lead to suboptimal performance .
Approach: They propose a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment.
Outcome: The proposed training framework outperforms state-of-the-art training frameworks at the same sparsity level while delivering up to 2.71 inference speedup.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations