Strong Drafts Need Compact Memories: Long-Context Speculative Decoding with Compressed KV Cache

arXiv:2608.30252v1 Announce Type: new Abstract: Long-context LLM applications such as document summarization and multi-turn agents require generation from prefixes spanning tens of thousands of tokens, making decoding latency a major bottleneck. Speculative decoding (SD) reduces latency without changing model outputs, but its speedup depends on both accepted draft tokens and draft-step latency: Lightweight drafts are fast but lack the capacity to capture long-range dependencies, whereas strong independent drafts recover acceptance but incur growing KV-access cost at long prefixes. We introduce memory-augmented drafting for long-context SD, equipping a strong independent draft with compressed draft-side KV memory: A lightweight adaptor constructs and incrementally updates this memory to retain distant information and exact recent context. The target verifier retains its full KV cache and applies the standard accept/reject rule, preserving SD's lossless guarantee. Experiments on Llama~3.1-8B and 70B targets at prefix lengths up to 32K show that our method reduces draft-side memory by over 70%. It achieves speedups of up to 2.08x and 3.33x , respectively, over autoregressive decoding.

Related

• SpecPV: Improving Self-Speculative Decoding for Long-Context Generation via Partial Verification
• ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models
• MomentKV: Closing the Directional Gap in KV Cache Eviction for Long-Context Inference
• DynaCalKV: Key-Value Cache Compression via Head Grouping and Adaptive Rank Allocation
• RAP: KV-Cache Compression via RoPE-Aligned Pruning

Source: arXiv cs.LG | 2026-09-01