Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs

• URL: http://arxiv.org/abs/2512.03324v1
• Date: Wed, 03 Dec 2025 00:20:35 GMT
• Title: Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs
• Authors: Ngoc Bui, Shubham Sharma, Simran Lamba, Saumitra Mishra, Rex Ying,
• Abstract summary: We propose a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate.<n>We show that it consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes.<n>It even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization.
• Score: 26.951325519894525
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBench and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

Related papers

• G-KV: Decoding-Time KV Cache Eviction with Global Attention [57.47409249054187]
  Large language models (LLMs) excel in complex tasks but encounter significant computational and memory challenges due to long sequence lengths.<n> KV cache compression has emerged as an effective approach to greatly enhance the efficiency of reasoning.<n>We propose G-KV, a KV cache eviction method that employs a global scoring mechanism, combining local and historical attention scores to more accurately assess token importance.
  arXiv  Detail & Related papers  (2025-11-29T14:21:33Z)
• Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction [53.83828564664595]
  Large language models (LLMs) utilize key-value ( KV) cache to store historical information during sequence processing.<n>Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction.<n>We propose Judge Q, a novel training method which incorporates a soft token list.
  arXiv  Detail & Related papers  (2025-09-13T03:34:12Z)
• ZSMerge: Zero-Shot KV Cache Compression for Memory-Efficient Long-Context LLMs [15.76582272387931]
  We propose ZSMerge, a dynamic KV cache compression framework for efficient cache management.<n>ZSMerge significantly enhances memory efficiency and inference speed with negligible performance degradation.
  arXiv  Detail & Related papers  (2025-03-13T03:36:03Z)
• LLMs Know What to Drop: Self-Attention Guided KV Cache Eviction for Efficient Long-Context Inference [16.83202690345235]
  We propose Self-Attention Guided Eviction(SAGE-KV), a simple and effective KV eviction cache method for long-context inference.<n>After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache.<n>SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.
  arXiv  Detail & Related papers  (2025-03-11T20:45:02Z)
• Compressing KV Cache for Long-Context LLM Inference with Inter-Layer Attention Similarity [24.118503938098307]
  textscPoD allocates memory according to token importance.<n>textscPoD reduces KV cache memory usage by up to 35% without compromising performance.
  arXiv  Detail & Related papers  (2024-12-03T08:29:27Z)
• In-context KV-Cache Eviction for LLMs via Attention-Gate [12.732519329131392]
  The KV-Cache technique has become the standard for the inference of large language models (LLMs)<n>This paper enables a novel dynamic KV-Cache eviction policy by injecting a lightweight module called Attention-Gate to the model.<n>We empirically evaluate the proposed approach across multiple scenarios, showing that effective eviction of redundant tokens can not only improve efficiency but also enhance performance.
  arXiv  Detail & Related papers  (2024-10-15T05:01:19Z)
• ThinK: Thinner Key Cache by Query-Driven Pruning [63.13363917871414]
  Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications.<n>This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference.<n>We propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels.
  arXiv  Detail & Related papers  (2024-07-30T17:59:08Z)
• Efficient Inference of Vision Instruction-Following Models with Elastic Cache [76.44955111634545]
  We introduce Elastic Cache, a novel strategy for efficient deployment of instruction-following large vision-language models. We propose an importance-driven cache merging strategy to prune redundancy caches. For instruction encoding, we utilize the frequency to evaluate the importance of caches. Results on a range of LVLMs demonstrate that Elastic Cache not only boosts efficiency but also notably outperforms existing pruning methods in language generation.
  arXiv  Detail & Related papers  (2024-07-25T15:29:05Z)
• Training-Free Exponential Context Extension via Cascading KV Cache [49.608367376911694]
  We introduce a novel mechanism that leverages cascading sub-cache buffers to selectively retain the most relevant tokens.<n>Our method reduces prefill stage latency by a factor of 6.8 when compared to flash attention on 1M tokens.
  arXiv  Detail & Related papers  (2024-06-24T03:59:17Z)
• Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference [78.65321721142624]
  We focus on a memory bottleneck imposed by the key-value ( KV) cache. Existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs. We propose LESS, a simple integration of a constant sized cache with eviction-based cache methods.
  arXiv  Detail & Related papers  (2024-02-14T18:54:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.