Efficient inference of large language models (LLMs) is hindered by an ever-growing key-value (KV) cache, making KV cache compression a critical research direction. Traditional methods selectively evict less important KV cache entries, which leads to information loss and hallucinations. Recently, merging-based strategies have been explored to retain more information by merging KV pairs that would be discarded; however, these existing approaches inevitably introduce inconsistencies in attention distributions before and after merging, causing degraded generation quality. To overcome this challenge, we propose KeepKV, a novel adaptive KV cache merging method designed to preserve performance under strict memory constraints, achieving single-step lossless compression and providing error bounds for multi-step compression. KeepKV introduces the Electoral Votes mechanism that records merging history and adaptively adjusts attention scores. Moreover, it further leverages a novel Zero Inference-Perturbation Merging method, compensating for attention loss resulting from cache merging. Extensive experiments on various benchmarks and LLM architectures demonstrate that KeepKV substantially reduces memory usage while successfully retaining essential context information, achieving over 2x inference throughput improvement and maintaining superior generation quality even with only 10% KV cache budgets.

Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV Cache compression methods mostly rely on attention score to reduce cache size, which makes them are incompatible with established efficient attention kernels (e.g., FlashAttention) and ignores the contribution of value vectors to the attention output. In this work, we revisit multimodal KV Cache compression from the perspective of the KV matrices' distribution. First, we observe that frequency-domain energy of multimodal KV matrices is predominantly concentrated in low-frequency and extract this principal energy via a low-pass filter. Further, we find that removing KV pairs that deviate substantially from this principal energy leads to a pronounced performance drop, which we define as Outlier KVs. Considering Outlier KVs are more likely to encode features critical for inference, we propose FlashCache, a frequency-domain-guided, Outlier-KV-aware KV Cache compression framework. First, we introduce an Outlier KV Recognition Module that models the principal component of multimodal KV matrices in the frequency domain and preferentially retains KV pairs that significantly deviate from it. Furthermore, Dynamic Budget Allocation Module is designed to adaptively determine the per-layer KV Cache size to retain more Outlier KVs. Experiments on multiple MLLMs and benchmarks demonstrate that FlashCache outperforms state-of-the-art multimoal KV compression methods, achieving up to 1.69 times faster decoding with 80% lower KV memory usage while maintaining task performance.

Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a Token Importance Recurrence phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose LazyEviction, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50%~70% while maintaining comparable accuracy, outperforming existing KV cache compression baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.

Large Language Models (LLMs) are increasingly deployed in multi-agent systems, where effective inter-model communication is crucial. Existing communication protocols either rely on natural language, incurring high inference costs and information loss, or on hidden states, which suffer from information concentration bias and inefficiency. To address these limitations, we propose KVComm, a novel communication framework that enables efficient communication between LLMs through selective sharing of KV pairs. KVComm leverages the rich information encoded in the KV pairs while avoiding the pitfalls of hidden states. We introduce a KV layer-wise selection strategy based on attention importance scores with a Gaussian prior to identify the most informative KV pairs for communication. Extensive experiments across diverse tasks and model pairs demonstrate that KVComm achieves comparable performance to the upper-bound method, which directly merges inputs to one model without any communication, while transmitting as few as 30% of layers' KV pairs. Our study highlights the potential of KV pairs as an effective medium for inter-LLM communication, paving the way for scalable and efficient multi-agent systems. Large Language Models (LLMs) have catalyzed a paradigm shift from isolated model capabilities towards collaborative multi-agent systems (Guo et al., 2024; Tran et al., 2025). CAMEL (Li et al., 2023), AutoGen (Wu et al., 2024), and ChatDev (Qian et al., 2023) have demonstrated the potential of LLMs to collaborate effectively in multi-agent systems, achieving impressive results in various tasks. These systems leverage the strengths of individual LLMs and enable them to work together to solve complex problems that are beyond the capabilities of a single model (Y ang et al., 2024a).

Memory consumption of the Key-Value (KV) cache represents a major bottleneck for efficient large language model inference. While attention-score-based KV cache pruning shows promise, it faces critical practical limitations: attention scores from future tokens are unavailable during compression, and modern implementations like Flash Attention do not materialize the full attention matrix, making past scores inaccessible. To overcome these challenges, we introduce $\textbf{Expected Attention, a training-free compression method}$ that estimates KV pairs importance by predicting how future queries will attend to them. Our approach leverages the distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair. These scores enable principled ranking and pruning of KV pairs with minimal impact on the residual stream, achieving effective compression without performance degradation. Importantly, our method operates seamlessly across both prefilling and decoding phases, consistently outperforming state-of-the-art baselines in both scenarios. Finally, $\textbf{we release KVPress, a comprehensive library to enable researchers to implement and benchmark KV cache compression methods, already including more than 20 techniques}$.

Linear attention is an efficient alternative that maintains a constant memory footprint, even on infinite context lengths. While this is a potential candidate for lifelong learning, it falls short in memory capacity. In this paper, we propose LoLA, a training-free augmentation to linear attention that boosts associative recall. LoLA distributes past key-value pairs from context into three memory systems: (i) recent pairs in a local sliding window cache; (ii) difficult-to-memorize pairs in a sparse, global cache; and (iii) generic pairs in the recurrent hidden state of linear attention. We show through ablations that our self-recall error metric is crucial to efficiently manage long-term associative memories. On pass-key retrieval tasks, LoLA improves the base model's performance from 0.6% to 97.4% accuracy. This is achieved with a 4.6 smaller cache than Llama-3.1 8B on 4K context length. LoLA also outperforms other 1B and 8B parameter subquadratic models on zero-shot commonsense reasoning tasks. Transformer-based large language models (LLMs) rely on storing all past tokens in an ever-growing key-value (KV) cache (V aswani et al., 2017). This allows future query tokens to access past memories with associative recall, which enables in-context learning (Olsson et al., 2022). Since no previous information is discarded, the KV cache continues to grow with context length. This eventually leads to a memory bottleneck on long context tasks, such as lifelong in-context learning. Alternative architectures to transformers have been proposed--such as Mamba (Gu & Dao, 2024), DeltaNet (Schlag et al., 2021), linear attention (Katharopoulos et al., 2020), and others (Y ang et al., 2024a; Behrouz et al., 2024; Sun et al., 2024)--to reduce the compute complexity from quadratic to linear. Additionally, these approaches reduce the memory cost from linear to constant. In particular, linear attention removes the exponential dot product in softmax (Katharopoulos et al., 2020).

Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by $3$-$4\times$ and FlashAttention decoding latency by approximately $2\times$, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1, Qwen2.5, and Gemma3, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90% cache budget ratio under multi-query scenarios.

As large language models (LLMs) continue to support increasingly longer contexts, the memory demand for key-value (KV) caches during decoding grows rapidly, becoming a critical bottleneck in both GPU memory capacity and PCIe bandwidth. Sparse attention mechanisms alleviate this issue by computing attention weights only for selected key-value pairs. However, their indexing computation typically requires traversing all key vectors, resulting in significant computational and data transfer overhead. To reduce the cost of index retrieval, existing methods often treat each decoding step as an independent process, failing to exploit the temporal correlations embedded in historical decoding information. To this end, we propose LFPS(Learn From the Past for Sparse Indexing), an acceleration method that dynamically constructs sparse indexing candidates based on historical attention patterns. LFPS captures two prevalent trends in decoder attention -vertical patterns (attending to fixed positions) and slash patterns (attending to relative positions) -and incorporates a positional expansion strategy to effectively predict the Top-k indices for the current step. We validate LFPS on challenging long-context benchmarks such as LongBench-RULER, using Llama-3.1-8B-Instruct as the base model. Experimental results show that LFPS achieves up to 22.8$\times$ speedup over full attention and 9.6$\times$ speedup over exact Top-k retrieval on an RTX 4090 GPU and a single CPU core of a Xeon Gold 6430, respectively, while preserving generation accuracy. These results demonstrate that LFPS offers a practical and efficient solution for decoding optimization in long-context LLM inference.

Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.

Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial resources as their multimodal Key-Value (KV) caches grow with increasing input lengths, challenging inference efficiency. Existing methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, thus often adopting uniform or progressive reduction strategies for layer-wise cache allocation. In this work, we propose MEDA, a dynamic layer-wise KV cache allocation method for efficient multimodal long-context inference. As its core, MEDA utilizes cross-modal attention entropy to determine the KV cache size at each MLLMs layer. Given the dynamically allocated KV cache size at each layer, MEDA also employs a KV pair selection scheme to identify which KV pairs to select and a KV pair merging strategy that merges the selected and non-selected ones to preserve information from the entire context. MEDA achieves up to 72% KV cache memory reduction and 2.82 times faster decoding speed, while maintaining or enhancing performance on various multimodal tasks in long-context settings, including multi-images and long-video scenarios. Our code is released at https://github.com/AIoT-MLSys-Lab/MEDA.