Guardado en:
| Autores principales: | , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | https://arxiv.org/abs/2605.29873 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866914614218850304 |
|---|---|
| author | Jana, Soumyadeep Nishad, Sagar Singh, Sanasam Ranbir |
| author_facet | Jana, Soumyadeep Nishad, Sagar Singh, Sanasam Ranbir |
| contents | Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_29873 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation Jana, Soumyadeep Nishad, Sagar Singh, Sanasam Ranbir Artificial Intelligence Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency. |
| title | Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2605.29873 |