Challenge: Current GQA configurations overlook how context length influences inference cost .
Approach: They propose a recipe for deriving cost-optimal GQA configurations that decouple the total head size from the hidden size and allow more flexible control over attention FLOPs.
Outcome: The proposed configurations reduce memory usage and FLOPs by more than 50% compared to Llama-3's GQA, with *no degradation in model capabilities*.

Similar Papers

Align Attention Heads Before Merging Them: An Effective Way for Converting MHA to GQA (2025.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks.
Approach: They propose a method for converting multi-head attention into grouped-query attention with any compression ratio of KV heads.
Outcome: The proposed method can compress up to 87.5% KV heads of LLaMA2-7B model and 75% Kv heads of Sheared-LLa MA-1.3B with acceptable performance degradation.
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints (2023.emnlp-main)

Copied to clipboard

Challenge: Multi-query attention (MQA) can lead to quality degradation and training instability . it may not be feasible to train separate models optimized for quality and inference.
Approach: They propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original training compute.
Outcome: The proposed model achieves comparable quality to multi-head attention with comparable speed.
LongHeads: Multi-Head Attention is Secretly a Long Context Processor (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models struggle to process lengthy inputs due to limited length generalization and attention’s quadratic computational demands.
Approach: They propose a training-free framework that allows each head to attend to important context chunks instead of allowing each head a full sentence .
Outcome: The proposed framework unlocks multi-head attention's untapped potential by allowing each head to attend to important context chunks instead of the full sentence.
Every Token Counts: Generalizing 16M Ultra-Long Context in Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: a recent study explores efficient ultra-long context modeling.
Approach: They propose to use Hierarchical Sparse Attention to achieve efficient ultra-long context modeling.
Outcome: The proposed model performs comparable to full-attention baselines on in-domain and out-of-domain tasks.
Cross-layer Attention Sharing for Pre-trained Large Language Models (2026.tacl-1)

Copied to clipboard

Challenge: Existing studies focus on compressing the Key-Value cache or grouping attention heads, while overlooking redundancy between layers.
Approach: They propose a lightweight substitute for self-attention in well-trained LLMs that uses feed-forward networks to align attention heads between adjacent layers and low-rank matrices to approximate differences in layer-wise attention weights.
Outcome: The proposed model reduces redundancy by sharing weights across layers while maintaining high response quality while reducing redundant calculations within 53% 84% of the total layers.
MATCH: Modulating Attention via In-Context Retrieval for Long-Context Transformers (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches to improve efficiency often enforce rigid structural constraints such as local attention windows.
Approach: They propose a framework that augments sparse-attention mechanisms with dynamically integrated in-context information through an efficient retrieval system.
Outcome: Empirical results show that MATCH significantly improves the performance of sparse-attention models on synthetic and real-world natural-language tasks.
DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Large reasoning models generate long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence.
Approach: They propose a training-free sparse attention mechanism that reduces inference cost by evicting entries from the key-value cache.
Outcome: The proposed model matches or surpasses full attention on reasoning benchmarks . it reduces the number of attended tokens by up to 4.25 and delivers 1.54 speedup .
Latent-Condensed Transformer for Efficient Long Context Modeling (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches address these bottlenecks separately: Multi-head Latent Attention (MLA) reduces the KV cache by projecting tokens into a low-dimensional latent space, while sparse attention reduces computation.
Approach: They propose a Latent-Condensed Attention mechanism that performs structured context condensation directly within MLA's latent space.
Outcome: The proposed approach reduces KV cache size and attention cost without adding parameters.
Towards Economical Inference: Enabling DeepSeek’s Multi-Head Latent Attention in Any Transformer-based LLMs (2025.acl-long)

Copied to clipboard

Challenge: Multi-head Latent Attention (MLA) is an innovative architecture designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector.
Approach: They propose a data-efficient fine-tuning method for transitioning from MHA to MLA using a latent vector cache.
Outcome: The proposed architecture reduces the KV cache size of Llama2-7B by 92.19%, with only 1% drop in LongBench performance.
Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for estimating attention importance for tokens are ineffective . dLLMs require bidirectional attention, which limits inference efficiency .
Approach: They propose a training-free attention sparsification framework for efficient long-context inference . they propose 'sink-aware pruning strategy' to accurately estimate and remove redundant computation .
Outcome: The proposed approach offers 29 lossless speedup under 32K context length.

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