| Challenge: | Empirical evaluations across various prominent LLMs and benchmarks show that key-favored allocations retain up to 98.3% accuracy compared to uniform allocations (e.g., 4-bit keys, 2-bit values). |
| Approach: | They propose two theorems that anchor mixed-precision KV quantization in the intrinsic geometry of Transformer models. |
| Outcome: | Empirical evaluations show that key-favored allocations retain up to 98.3% accuracy while conserving memory. |
Similar Papers
AsymKV: Enabling 1-Bit Quantization of KV Cache with Layer-Wise Asymmetric Quantization Configurations (2025.coling-main)
Copied to clipboard
| Challenge: | Large language models require substantial storage space to perform tasks such as text generation and video generation. |
| Approach: | They propose to compress large language models using integer replacements for floating-point numbers, in a process known as Quantization. |
| Outcome: | The proposed model allows for quantization of up to 75% decoder layers with 1 bit while maintaining performance levels comparable to those of the models with floating parameters. |
TaDA: Training-free recipe for Decoding with Adaptive KV Cache Compression and Mean-centering (2025.acl-industry)
Copied to clipboard
| Challenge: | key-value caches in large language models consume memory, posing a major challenge for scalable deployment. |
| Approach: | They propose a training-free recipe for KV cache compression with quantization precision that adapts to error sensitivity across layers and a mean centering to eliminate separate outlier handling. |
| Outcome: | The proposed technique reduces the KV cache memory footprint to 27% of the original 16-bit baseline while achieving comparable accuracy. |
More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression (2025.findings-emnlp)
Copied to clipboard
Jiebin Zhang, Dawei Zhu, Yifan Song, Wenhao Wu, Chuqiao Kuang, Xiaoguang Li, Lifeng Shang, Qun Liu, Sujian Li
| Challenge: | storing more tokens in the KV cache at lower precision can enhance the long-context performance of large language models. |
| Approach: | They propose a token-precision trade-off strategy to optimize KV cache compression . they also propose storing more tokens in the KV at lower precision . |
| Outcome: | The proposed method achieves an optimal point within the Information Bottleneck compared to standalone KV pruning or KV quantization. |
Accurate KV Cache Quantization with Outlier Tokens Tracing (2025.acl-long)
Copied to clipboard
Yi Su, Yuechi Zhou, Quantong Qiu, Juntao Li, Qingrong Xia, Ping Li, Xinyu Duan, Zhefeng Wang, Min Zhang
| Challenge: | Large Language Models (LLMs) require substantial computational resources during deployment. |
| Approach: | They propose a method to identify outlier tokens and exclude them from quantization . they find that the method can deliver a 6.4 times reduction in memory usage and a 2.5 times increase in throughput . |
| Outcome: | The proposed method delivers a 6.4 times reduction in memory usage and a 2.5 times increase in throughput under 2-bit quantization. |
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. |
KV Pareto: Systems-Level Optimization of KV Cache and Model Compression for Long Context Inference (2026.eacl-industry)
Copied to clipboard
| Challenge: | Long-context Large Language Models (LLMs) face significant memory bottlenecks due to the linear growth of key-value (KV) cache with sequence length. |
| Approach: | They propose a framework that maps the trade-off frontier between total memory consumption and task accuracy across three complementary optimization techniques. |
| Outcome: | The proposed model-specific configurations achieve 68-78% total memory reduction with minimal (1-3%) accuracy degradation on long-context tasks. |
MixKVQ: Query-Aware Mixed-Precision KV Cache Quantization for Long-Context Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. |
| Approach: | They propose a plug-and-play method that uses a key channel's intrinsic quantization difficulty and relevance to the query to identify and preserve critical key channels that need higher precision. |
| Outcome: | Experiments on complex reasoning datasets show that the proposed method outperforms low-bit methods at a substantially reduced memory footprint. |
HqeKV: Towards Hybrid Quantization and Eviction for KV Cache in Long-Context LLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | autoregressive inference requires repeated computation across transformer layers. |
| Approach: | They propose a hybrid compression framework built on both quantization and eviction . they propose varying importance metric and flexible conversion policies to reduce memory overhead . |
| Outcome: | The proposed framework outperforms state-of-the-art methods under memory constraints. |
A Systematic Study of Cross-Layer KV Sharing for Efficient LLM Inference (2025.naacl-short)
Copied to clipboard
| Challenge: | Recent studies have shown that sharing key-value (KV) cache across layers is effective in efficient inference of large language models. |
| Approach: | They propose a unified framework that covers several recent methods and their novel variants to investigate cross-layer KV sharing. |
| Outcome: | The proposed framework achieves higher throughput and better performance when reducing the size of the key-value cache by 2 while maintaining competitive performance. |
LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (2024.findings-acl)
Copied to clipboard
Zechun Liu, Barlas Oguz, Changsheng Zhao, Ernie Chang, Pierre Stock, Yashar Mehdad, Yangyang Shi, Raghuraman Krishnamoorthi, Vikas Chandra
| Challenge: | Several post-training quantization methods have been shown to perform well down to 8-bits. |
| Approach: | They propose a data-free distillation method that leverages generations produced by the pre-trained model to quantize any generative model independent of its training data. |
| Outcome: | The proposed method outperforms SoTA PTQ and LLaMA models at low bit precision. |