Research map / Architectures and efficiency / Attention and transformers
KnapSpec: Self-Speculative Decoding via Adaptive Layer Selection as a Knapsack Problem
Seongjin Cha, Gyuwan Kim, Dongsu Han, Tao Yang, Insu Han
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use TPT instead of acceptance rate as the predictor of throughput (ablation) | KnapSpec · GovReport 12K-13K tokens | Pearson correlation with throughput: 0.538 -> 0.837 | Improved |
| add cosine-similarity pruning threshold tau=0.5 (ablation) | KnapSpec · GovReport | speedup / memory: 1.26x speedup retained, memory reduced by 31% | Improved |
| raise pruning threshold beyond 0.5 (ablation) | KnapSpec · Llama-3.1-8B | speedup | Hurt |
| increase optimization interval from 4 to 64 steps (ablation) | KnapSpec · Qwen3-8B | optimization overhead: 40.9% -> 7.5% | Improved |
| increase optimization interval beyond 64 steps (128, 256) (ablation) | KnapSpec · Qwen3-8B | throughput: peak 18.1 tok/s at 64 | Hurt |
| replace DEL/SWIFT/CLaSp with KnapSpec layer selection | self-speculative decoding · GovReport | speedup: 1.33x (SWIFT), 1.22x (CLaSp) -> 1.47x | Improved |
| replace greedy decoding with nucleus sampling | KnapSpec · AIME24 | speedup: 1.28 -> 1.13 | Mixed |
| use KnapSpec on other GPUs (RTX A6000 vs RTX 6000 Pro) | KnapSpec · GovReport Llama-3.1-8B | speedup: 1.27x -> 1.28x | No clear effect |
| use DEL early-exit on base models | self-speculative decoding · GovReport Llama3.1-70B | speedup: 1.00x -> 0.87x | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SWIFT — SwiFT: Swin 4D fMRI Transformer (NeurIPS 2023) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration (ICLR 2025 · similarity 0.93)
- TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees (NeurIPS 2026 · similarity 0.92)
- Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding (NeurIPS 2025 · similarity 0.91)
- SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting (NeurIPS 2026 · similarity 0.91)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.91)
- PEARL: Parallel Speculative Decoding with Adaptive Draft Length (ICLR 2025 · similarity 0.91)
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.
