Research map / Large language models / Reasoning and chain of thought
SMART: When is it Actually Worth Expanding a Speculative Tree?
Lifu Wang, Pan ZHOU
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 |
|---|---|---|---|
| add SMART marginal benefit-cost tree expansion to MSD | MSD, MLLMs · MLLM benchmarks T=0 | avg speedup: +20.2% avg across three MLLMs (LLaVA-1.5 7B 1.18 -> 1.53) | Improved |
| add SMART to MSD | MSD, MLLMs · MLLM benchmarks T=1 | avg speedup: +19.9% avg; LLaVA-1.5 7B 1.77 -> 2.28 | Improved |
| add SMART to EAGLE-3 | EAGLE-3, LLMs · MT-bench/HumanEval/GSM8K T=0 | avg speedup: +19.6% avg; LLaMA-3.3 70B 2.69 -> 3.35 | Improved |
| add SMART to EAGLE-3 | EAGLE-3, LLMs · MT-bench/HumanEval/GSM8K T=1 | avg speedup: +15.4% avg; LLaMA-3.1 8B 2.11 -> 2.38 | Improved |
| add SMART to MSD at batch size 1 (ablation) | MSD · LLaVA-1.5 RTX Pro 6000 batch 1 | speedup: 2.20 -> 2.17 | No clear effect |
| add SMART to MSD at batch size 16 (ablation) | MSD · LLaVA-1.5 RTX Pro 6000 batch 16 | speedup: 1.22 -> 1.58 | Improved |
| add SMART to MSD at batch size 32 (ablation) | MSD · LLaVA-1.5 RTX Pro 6000 batch 32 | speedup: 0.82 -> 1.39 | Improved |
| add SMART to MSD at batch size 12 (ablation) | MSD · LLaVA-1.5 L40S batch 12 | speedup: 0.90 -> 1.40 | Improved |
| increase verification token budget from 200 to 300-400 (ablation) | SMART · RTX Pro 6000 batch 16, T=0 | avg speedup: 1.58 -> 1.28 (300), 1.27 (400) | Hurt |
| decrease token budget from 200 to 100 (ablation) | SMART · RTX Pro 6000 batch 16, T=0 | avg speedup: 1.58 -> 1.43 | Hurt |
| set discount factor alpha=1.0 instead of 0.8 (ablation) | SMART · RTX Pro 6000 batch 16, T=0 | avg speedup: 1.56 -> 1.51 | Hurt |
| vary discount factor alpha within 0.5-0.9 (ablation) | SMART · RTX Pro 6000 batch 16, T=0 | avg speedup: 1.54 to 1.56 | No clear effect |
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.
- EAGLE-3 — EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test (NeurIPS 2025) · better on 5 of 5 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
- TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees (NeurIPS 2026 · similarity 0.91)
- RheoSampling: Resolving the One-Hot Dilemma in Stochastic Dynamic-Tree Speculative Decoding (NeurIPS 2026 · similarity 0.91)
- SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting (NeurIPS 2026 · similarity 0.91)
- STree: Speculative Tree Decoding for Hybrid State Space Models (NeurIPS 2025 · similarity 0.91)
- SpecBranch: Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism (ICLR 2026 · similarity 0.90)
- BASTION: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting (NeurIPS 2026 · similarity 0.90)
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.
