AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

NanoSpec: Accelerating Speculative Decoding using Minimalist In-Context Vocabularies

Zhiyang Chen, Daliang Xu, Yinyuan Zhang, Chenghua Wang, Mengwei Xu, Yun Ma

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace full 128k LM head with dynamic <3k in-context vocabularyEAGLE-3 · Llama-3.1-8B-Instructavg speed (tokens/s): 362.2 -> 430.0Improved
use NanoSpec dynamic vocabulary instead of full vocabularyEAGLE-2 · Llama-3.2-1B-Instructavg speed (tokens/s): 700.9 -> 900.7Improved
use dynamic vocabulary instead of static FR-Spec pruningNanoSpec · Llama-3.2-1B-Instructavg acceptance length: 2.70 -> 3.11Improved
prune vocabulary dynamically versus full vocabularyNanoSpec · Llama-3.1-8B-Instruct, EAGLE-2avg acceptance length: 3.80 -> 3.59Hurt
shrink draft LM head to dynamic vocabularyNanoSpec · Llama-3.1-8B-Instructdraft LM head latency (ms): 2.330 -> 0.237Improved
add asynchronous weight gathering with dual CUDA streams (ablation)NanoSpec · Llama-3.1-8B-Instructavg speed (tokens/s): 364.1 -> 392.7Improved
add decoding-time candidates (draft and verify tokens) to context-only vocabulary (ablation)NanoSpec · Llama-3.1-8B-Instructavg speed (tokens/s): 313.3 -> 351.4Improved
combine prompt-context and decoding-derived vocabulary sources (ablation)NanoSpec · Llama-3.1-8B-Instructavg speed (tokens/s): 351.4 -> 364.1Improved
increase window size W_max beyond 4096 up to 8192 (ablation)NanoSpec · Llama-3.1-8B-Instructend-to-end speedHurt
increase W_max from 1024 to 2048 (ablation)NanoSpec · Llama-3.1-8B-Instructavg acceptance lengthImproved
increase W_max beyond about 2048 (ablation)NanoSpec · Llama-3.1-8B-Instructavg acceptance lengthNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy