# SelKV: Selective KV Cache Merging with Per-Token Merge-or-Drop and Attention Compensation

arXiv: https://arxiv.org/abs/2607.16213 · HTML: https://mlsanity.com/paper/2607.16213/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in KV cache optimization, one of 12 directions in Efficient LLM inference and serving.
- Builds on 16 of its 18 references that are in the map.
- First posted on arXiv 19 May 2026.

## Placed in

- Efficient LLM inference and serving / [KV cache optimization](https://mlsanity.com/inference/kv-cache-optimization/)

## Builds on (most cited first, 16 of 16)

- [Efficient Streaming Language Models with Attention Sinks](https://mlsanity.com/paper/2309.17453/) (2023)
- [LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding](https://mlsanity.com/paper/2308.14508/) (2023)
- [SnapKV: LLM Knows What You are Looking for Before Generation](https://mlsanity.com/paper/2404.14469/) (2024)
- [KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache](https://mlsanity.com/paper/2402.02750/) (2024)
- [PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling](https://mlsanity.com/paper/2406.02069/) (2024)
- [Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Tasks](https://mlsanity.com/paper/2407.08454/) (2024)
- [LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference](https://mlsanity.com/paper/2406.18139/) (2024)
- [ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification](https://mlsanity.com/paper/2405.14256/) (2024)
- [DiffRate : Differentiable Compression Rate for Efficient Vision Transformers](https://mlsanity.com/paper/2305.17997/) (2023)
- [Attention Score is not All You Need for Token Importance Indicator in KV Cache Reduction: Value Also Matters](https://mlsanity.com/paper/2406.12335/) (2024)
- [ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference](https://mlsanity.com/paper/2502.00299/) (2025)
- [D2O: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models](https://mlsanity.com/paper/2406.13035/) (2024)
- [LAVa: Layer-wise KV Cache Eviction with Dynamic Budget Allocation](https://mlsanity.com/paper/2509.09754/) (2025)
- [EvolKV: Evolutionary KV Cache Compression for LLM Inference](https://mlsanity.com/paper/2509.08315/) (2025)
- [EMS: Adaptive Evict-then-Merge Strategy for Head-wise KV Cache Compression Based on Global-Local Importance](https://mlsanity.com/paper/2412.08521/) (2024)
- [WeightedKV: Attention Scores Weighted Key-Value Cache Merging for Large Language Models](https://mlsanity.com/paper/2503.01330/) (2025)

## Abstract (arXiv)

Large Language Models (LLMs) generate text autoregressively, relying on a key-value (KV) cache whose memory footprint grows linearly with context length, creating a major bottleneck. Recent compression methods mitigate this cost via token merging; however, these approaches often rely on indiscriminate aggregation, which degrades representations and introduces attention sag, a mismatch where merged tokens receive the same softmax mass as individual tokens despite encoding multiple inputs. We propose a training-free, dual-component framework for KV cache compression that addresses these limitations. First, a soft cosine gate adaptively modulates merging decisions based on value-vector similarity, suppressing or discarding dissimilar tokens to preserve semantic fidelity. Second, we introduce an attention-ratio compensation mechanism that applies a decoding-time logit bias derived from prefill attention statistics, correcting the softmax imbalance induced by merging. Evaluated on LongBench (16 English datasets) while retaining only 25% of the KV cache, our framework achieves strong compressed performance against representative one-shot baselines. It is especially robust on the evaluated grouped-query attention (GQA) models, maintaining nearlossless generation quality. Furthermore, the method outperforms the full-cache baseline on complex multi-document QA tasks and delivers a 3.3x decoding speedup at 100k tokens.