AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Draft-based Approximate Inference for LLMs

Kevin Galim, Ethan Ewer, Wonjun Kang, Minjae Lee, Hyung Il Koo, Kangwook Lee

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 6 citations (OpenAlex)

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
use SpecKV draft lookahead instead of SnapKVKV dropping Cmax=256 · LongBench All, Qwen2.5average score: 42.98 -> 46.06Improved
use SpecKV instead of LAQ++KV dropping Cmax=256 · LongBench All, Llama-3 70Baverage score: 48.43 -> 48.80Improved
use SpecKV instead of SnapKV on SingleQAKV dropping Cmax=256 · LongBench SingleQA, Llama-3 70Bscore: 55.88 -> 51.80Hurt
cascade SpecPC prompt compression to 2048 tokens before SpecKVSpecKV · LongBench All, Llama-3 70Baverage score: 48.80 -> 51.60Improved
cascade SpecPC before SpecKVSpecKV · LongBench All, Qwen2.5average score: 46.06 -> 46.48Improved
use SpecPC instead of SpecPrefillprompt compression Cmax=1024 · LongBench All, Qwen2.5average score: 42.70 -> 43.66Improved
use SpecPC instead of SpecPrefillprompt compression Cmax=1024 · LongBench All, Llama-3 70Baverage score: 48.37 -> 48.44No clear effect
use SpecPC instead of LLMLingua-2prompt compression Cmax=1024 · LongBench All, Qwen2.5average score: 30.90 -> 43.66Improved
KV dropping with SpecKV vs dense targetKV dropping Cmax=256 · LongBench All, Qwen2.5average score: 47.78 -> 46.06Hurt
use larger draft model (ablation)SpecKV / SpecPC · RULER-32Kdownstream scoreImproved
use draft lookahead tokens (ablation)SpecKV · RULERaccuracyImproved
increase lookahead tokens (ablation)SpecPC · RULERaccuracyNo 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