Challenge: RankDyna is a matrix decomposition method that can be used to compress Transformer-based language models.
Approach: They propose a matrix decomposition method that enables dynamic rank resource allocation . they say it can outperform current SOTA methods under various parameter budget levels .
Outcome: The proposed method outperforms current SOTA methods under various budget levels . the proposed method is more efficient with higher compression rates .

Similar Papers

Numerical Optimizations for Weighted Low-rank Estimation on Language Models (2022.emnlp-main)

Copied to clipboard

Challenge: Singular value decomposition (SVD) is one of the most popular methods for estimating a target matrix with smaller matrices.
Approach: They propose a method that approximates a target matrix with smaller matrices by two smaller . they also propose metric to predict when the SVD may introduce a significant performance drop.
Outcome: The proposed method can perform better than current SOTA methods in compressing Transformer-based language models.
Sensitivity-LoRA : Low-Load Sensitivity-Based Fine-Tuning for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) is a promising approach to adapting LLMs to specialized tasks . existing rank allocation techniques remain computationally inefficient and unstable .
Approach: They propose a low-rank adapted model that approximates model weight updates using low-ranked decomposition.
Outcome: The proposed method is limited by its uniform rank allocation to each incremental matrix . it leverages the second-order derivatives of the loss function to capture weight sensitivity .
Beyond Uniform SVD: Dual-Level Optimization across Columns and Modules for LLM Compression (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods for low-rank decomposition overlook decomposing errors and suboptimal approximation.
Approach: They propose a low-rank decomposition framework that integrates low-level optimization at column and module levels.
Outcome: The proposed framework outperforms state-of-the-art methods and baselines in SVD and pruning.
LightFormer: Light-weight Transformer Using SVD-based Weight Transfer and Parameter Sharing (2023.findings-acl)

Copied to clipboard

Challenge: Deploying Transformer networks on resource-constrained edge devices is challenging.
Approach: They propose a low-rank factorization initialized by SVD-based weight transfer and parameter sharing to compress and accelerate Transformer networks.
Outcome: The proposed method achieves similar performance to the baseline Transformer with 3.8 times and 1.8 times fewer parameters and achieves 2.3 times speedup and 1.5 times speed up respectively.
Adaptive Rank Selections for Low-Rank Approximation of Language Models (2024.naacl-long)

Copied to clipboard

Challenge: Singular Value Decomposition (SVD) or its weighted variants has progressed in compressing language models.
Approach: They propose a binary masking mechanism for optimizing the number of ranks in a differentiable framework.
Outcome: The proposed algorithm achieves much better accuracy than previous SVD and its weighted variants.
HyperAdaLoRA: Accelerating LoRA Rank Allocation During Training via Hypernetworks without Sacrificing Performance (2026.findings-acl)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) assumes a uniform rank r for each incremental matrix, not accounting for the varying significance of weight matrices across modules and layers.
Approach: They propose a framework that allows for faster convergence of low-rank adaptive models . they use a hypernetwork to prune the outputs of the hypernetworks to generate parameters .
Outcome: The proposed framework accelerates convergence of AdaLoRA by leveraging a hypernetwork.
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.
ALoRA: Allocating Low-Rank Adaptation for Fine-tuning Large Language Models (2024.naacl-long)

Copied to clipboard

Challenge: Low-rank adaptation (LoRA) has demonstrated commendable performance as a popular method . however, it is implemented with a fixed intrinsic rank that might not be ideal for downstream tasks.
Approach: They propose a method that estimates the importance score of each LoRA rank and prunes abundant LoRA ranks to improve performance.
Outcome: The proposed method outperforms baselines on a variety of tasks with comparable parameters.
DoRA: Enhancing Parameter-Efficient Fine-Tuning with Dynamic Rank Distribution (2024.acl-long)

Copied to clipboard

Challenge: Existing parameter-efficient fine-tuning methods such as Low-Rank Adaptation ignore the differential parameter budget requirements across weight matrices, which may lead to suboptimal fine-uning outcomes.
Approach: They propose a parameter-efficient low-rank Adaptation method that decomposes high-rank LoRA layers into structured single-rank components and allows dynamic pruning of parameter budget .
Outcome: The proposed method outperforms LoRA and LoRA with the same parameter budget and performance.
LRQ: Optimizing Post-Training Quantization for Large Language Models by Learning Low-Rank Weight-Scaling Matrices (2025.naacl-long)

Copied to clipboard

Challenge: Existing methods for quantizing weights and activations of large language models suffer from non-negligible accuracy drops, especially on massive multitask language understanding.
Approach: They propose a weight-activation quantization method that reconstructs the outputs of an intermediate Transformer block by leveraging low-rank weight-scaling matrices.
Outcome: The proposed method reduces the complexity of the weight-activation quantization techniques while achieving high throughput and reducing inference costs.

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