analiza aiqbench
This arXiv preprint studies how to shrink the key-value cache that draft models use during speculative decoding, a common way to speed up LLM output. It identifies "decode staleness," where prefill-focused cache selection loses relevance as outputs grow, and reports that refreshing selection and training the drafter with sparse attention masks recovers most acceptance length under an 8K cache budget. For practitioners, the result is a benchmark-specific efficiency gain for one model pairing (Qwen3.5), so its value depends on whether it generalizes to other models and workloads.
RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training
arXiv:2610.31098v1 Announce Type: new Abstract: Speculative decoding is a popular decode acceleration method. As such, it is particularly relevant to long-output generation, which requires an increasingly large KV cache. KV cache compression offers a way to constrain costs in long-context drafting. However, compression methods that focus on prefill information can lose relevance as the output length grows. We term this phenomenon decode staleness and show that it degrades draft model accuracy under selection-based KV cache compression. We introduce RIFT, a family of selection-based KV cache compression strategies that balance prefill information and recent generated tokens. We show that refreshing the KV cache selection during token generation significantly improves the drafter's quality at the cost of additional decode latency. To alleviate the refresh costs, we propose an adaptive KV refresh strategy that reduces the number of score calculations while retaining the drafter's quality. We further demonstrate how a compressibility-driven drafter training phase, based on sparse attention masks, can improve the drafter's quality under a compressed KV cache. Combined with compressibility training, RIFT enables Qwen3.5-4B, drafting for Qwen3.5-122B, to retain over 98% of its full-cache acceptance length on 32K-token inputs, 4K-token outputs, and an 8K-token KV cache budget.
Abstract:Speculative decoding is a popular decode acceleration method. As such, it is particularly relevant to long-output generation, which requires an increasingly large KV cache. KV cache compression offers a way to constrain costs in long-context drafting. However, compression methods that focus on prefill information can lose relevance as the output length grows. We term this phenomenon decode staleness and show that it degrades draft model accuracy under selection-based KV cache compression. We introduce RIFT, a family of selection-based KV cache compression strategies that balance prefill information and recent generated tokens. We show that refreshing the KV cache selection during token generation significantly improves the drafter's quality at the cost of additional decode latency. To alleviate the refresh costs, we propose an adaptive KV refresh strategy that reduces the number of score calculations while retaining the drafter's quality. We further demonstrate how a compressibility-driven drafter training phase, based on sparse attention masks, can improve the drafter's quality under a compressed KV cache. Combined with compressibility training, RIFT enables Qwen3.5-4B, drafting for Qwen3.5-122B, to retain over 98% of its full-cache acceptance length on 32K-token inputs, 4K-token outputs, and an 8K-token KV cache budget.
Computation and Language (cs.CL)
arXiv:2610.31098 [cs.CL]
(or arXiv:2610.31098v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.31098 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 286 cuvinte
Subiecte conexe
GS-MoE: Training-Free MoE Compression via Groupwise Subspace Sharing and Pruning
arXiv · 27 oct. 2026
Format-Controlled Multi-Scale JPEG Compression Response Analysis for Image-Level Forgery Screening
arXiv · 23 iul. 2026
Learning Representative Port Subsets for Efficient Fluid Antenna Systems
arXiv · 27 oct. 2026
Beyond the Best Teacher: Expanding and Compressing the Reasoning Solution Manifold
arXiv · 31 iul. 2026