Research map / Large language models / Reasoning and chain of thought
LEAP: Zone-Aware MCTS for LLM Self-Speculative Decoding
LeiQuan Zheng, Yuan Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LEAP, a plug-and-play method that frames draft-model construction for LLM self-speculative decoding as a layer-configuration search solved with zone-aware Monte Carlo Tree Search.
Problem
Self-speculative decoding builds a draft model from a subset of the target LLM's layers, and choosing which layers to keep (the layer configuration) is the key challenge for speed-up.
Approach
Formulates draft model construction as an optimization problem over layer configurations and solves it with zone-aware Monte Carlo Tree Search, as stated in the abstract; the specific zone mechanism beyond that is not described in the available text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add zone partitioning (ablation) | LEAP · Overall (GSM8K, MT-Bench, WMT14), LLaMA-3-8B | speedup: 1.67× -> 2.10× | Improved |
| add layer grouping (ablation) | LEAP · Overall (GSM8K, MT-Bench, WMT14), LLaMA-3-8B | speedup: 1.63× -> 2.10× | Improved |
| use MCTS instead of random search (ablation) | LEAP · Overall (GSM8K, MT-Bench, WMT14), LLaMA-3-8B | speedup: 1.56× -> 2.10× | Improved |
| add zone partitioning (ablation) | LEAP · MT-Bench, LLaMA-3-8B | speedup: 1.64× -> 2.07× | Improved |
| add layer grouping (ablation) | LEAP · MT-Bench, LLaMA-3-8B | speedup: 1.48× -> 2.07× | Improved |
| add layer grouping (ablation) | LEAP · GSM8K, LLaMA-3-8B | speedup: 1.89× -> 2.12× | Improved |
| use MCTS instead of random search (ablation) | LEAP · WMT14, LLaMA-3-8B | speedup: 1.51× -> 2.11× | Improved |
| raise sampling temperature from 0 to 1 | LEAP · Overall 5 datasets, LLaMA-3-8B | speedup: 2.05× -> 1.68× | Hurt |
| raise sampling temperature from 0 to 1 | SWIFT · Overall 5 datasets, LLaMA-3-8B | speedup: 1.13× -> 0.76× | Hurt |
| raise sampling temperature from 0 to 1 | CLaSP · Overall 5 datasets, LLaMA-3-8B | speedup: 1.27× -> 1.06× | Hurt |
| scale target model from 8B to 70B | SWIFT · Overall 5 datasets, T=0 | speedup: 1.13× -> 1.41× | Improved |
| scale target model from 8B to 70B | LEAP · Overall 5 datasets, T=0 | speedup: 2.05× -> 1.94× | No clear effect |
Closest papers by meaning
- SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration (ICLR 2025 · similarity 0.92)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.90)
- Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding (NeurIPS 2025 · similarity 0.90)
- TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees (NeurIPS 2026 · similarity 0.89)
- SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting (NeurIPS 2026 · similarity 0.89)
- Reward-Guided Speculative Decoding for Efficient LLM Reasoning (ICML 2025 · similarity 0.89)
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.
