Research map / Large language models / Reasoning and chain of thought
Mixture-of-Trees: Learning to Select and Weigh Reasoning Paths for Efficient LLM Inference
Yangbo Wei, Zhen Huang, Shaoqiang Lu, Junhong Qian, Dongge Qin, Ting Jung Lin, WEI W. XING, Chen Wu, Lei He
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 Mixture-of-Trees (MoT), which combines sparse expert activation via a learned gate with tree-based reasoning so that only the most relevant reasoning-tree experts, of varying model capacity, are run per problem for efficient LLM inference.
Problem
Efficient LLM inference for reasoning tasks, where running full tree-structured reasoning with large models on every problem is costly.
Approach
Applies sparse mixture-of-experts style gating at the level of reasoning trees: a learned gate activates only the most relevant expert trees per problem, and experts use models of different capacities matched to task complexity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sparse activation of expert trees (ablation) | MoT · HotpotQA | LLM calls: 98.45 -> 33.67 | Improved |
| add sparse activation of expert trees (ablation) | MoT · HotpotQA | accuracy: 82.8 -> 81.4 | Hurt |
| add early stopping with shared baseline tree (ablation) | MoT · HotpotQA | LLM calls: 33.67 -> 23.91 | Improved |
| add early stopping with shared baseline tree (ablation) | MoT · HotpotQA | accuracy: 81.4 -> 79.2 | Hurt |
| add specialized expert trees (ablation) | MoT · HotpotQA | accuracy: 79.2 -> 81.1 | Improved |
| replace FoT with MoT (gated experts, debate) | MoT · AIME 2024 (GPT-4o-mini) | accuracy: 33.3 -> 40.0 | Improved |
| replace FoT with MoT | MoT · GSM8K (GPT-4o-mini) | LLM calls: 32.32 -> 19.21 | Improved |
| replace FoT with MoT | MoT · MMLU (GPT-4o-mini) | accuracy: 91.6 -> 93.9 | Improved |
| lower activation threshold tau from 0.9 to 0.7 (ablation) | MoT · GSM8K/AIME | accuracy | Improved |
| lower activation threshold tau from 0.7 to 0.5 (ablation) | MoT · GSM8K/AIME | accuracy | No clear effect |
| scale expert pool from 8 to 64 (ablation) | MoT · GSM8K, MATH, MMLU | accuracy | Mixed |
Experimental setup
Builds on: mixture of experts tree-of-thoughts
Closest papers by meaning
- Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning (ICML 2025 · similarity 0.90)
- AlphaZero-Like Tree-Search can Guide Large Language Model Decoding and Training (ICML 2024 · similarity 0.88)
- DISC: Dynamic Decomposition Improves LLM Inference Scaling (NeurIPS 2025 · similarity 0.87)
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.87)
- VideoTree: Adaptive Tree-based Video Representation for LLM Reasoning on Long Videos (CVPR 2025 · similarity 0.86)
- Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees (NeurIPS 2024 · similarity 0.86)
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.
