analiza aiqbench
This research challenges the standard approach in training-free KV-cache compression by demonstrating that evicting tokens before the model generates its first answer tokens leads to suboptimal performance. The authors introduce Draft-Guided Eviction, which defers cache management until after drafting initial tokens, allowing the model to anchor on the correct answer trajectory before discarding data. This timing adjustment significantly improves long-context benchmark scores without altering existing eviction scoring mechanisms or increasing computational budgets.
When to Evict, Not What to Keep: Draft-Guided Eviction for Training-Free KV-Cache Compression
arXiv:2609.33334v1 Announce Type: new Abstract: Training-free KV-cache compression methods such as SnapKV, H2O, and PyramidKV evict tokens at the end of prefill, aiming to preserve the attention mass that future queries are expected to use -optimizing what to keep. We show that this objective fails in two distinct ways. (1) Compensation: restoring the evicted attention mass can recover the attention-level target without recovering task quality. (2) Selection: covering more of the true decode-query mass can hurt quality when the recovered mass is fragmented rather than concentrated in coherent spans. These failures share a common cause: eviction occurs before the queries that determine the answer trajectory exist. We propose Draft-Guided Eviction (DGE), which defers eviction until after drafting the first k=2 answer tokens using the full cache - just one decode step beyond prefill. Because the draft is generated from the answer's own prefix, no cache entries are discarded before this trajectory signal becomes available. The per-head cache budget remains unchanged, and DGE can be applied directly to SnapKV, PyramidKV, H2O, and StreamingLLM without modifying their eviction scores. Unlike extra-pass methods, DGE changes when eviction occurs rather than what cache entries are selected. Extensive experiments demonstrate that DGE outperforms prior methods at every evaluated budget on five of six instruct-tuned backbones, achieving 44.2 on LongBench, nearly matching FullKV at 44.3. The timing-only control DGE-W achieves the same score, demonstrating that the gain comes from when eviction occurs rather than what is selected - an effect we term trajectory anchoring.
Abstract:Training-free KV-cache compression methods such as SnapKV, H2O, and PyramidKV evict tokens at the end of prefill, aiming to preserve the attention mass that future queries are expected to use -optimizing what to keep. We show that this objective fails in two distinct ways. (1) Compensation: restoring the evicted attention mass can recover the attention-level target without recovering task quality. (2) Selection: covering more of the true decode-query mass can hurt quality when the recovered mass is fragmented rather than concentrated in coherent spans. These failures share a common cause: eviction occurs before the queries that determine the answer trajectory exist. We propose Draft-Guided Eviction (DGE), which defers eviction until after drafting the first k=2 answer tokens using the full cache - just one decode step beyond prefill. Because the draft is generated from the answer's own prefix, no cache entries are discarded before this trajectory signal becomes available. The per-head cache budget remains unchanged, and DGE can be applied directly to SnapKV, PyramidKV, H2O, and StreamingLLM without modifying their eviction scores. Unlike extra-pass methods, DGE changes when eviction occurs rather than what cache entries are selected. Extensive experiments demonstrate that DGE outperforms prior methods at every evaluated budget on five of six instruct-tuned backbones, achieving 44.2 on LongBench, nearly matching FullKV at 44.3. The timing-only control DGE-W achieves the same score, demonstrating that the gain comes from when eviction occurs rather than what is selected - an effect we term trajectory anchoring.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2609.33334 [cs.LG]
(or arXiv:2609.33334v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.33334 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 · 341 cuvinte
Subiecte conexe
Deploying Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod with vLLM
AWS · 9 sept. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
[AINews] not much happened today
Latent Space · 10 sept. 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026