Research map / Large language models / Agents, code and math
LEGO: An LLM-Enabled Hierarchical Optimizer for Tensor Computation Graphs with Structure-Aware Search and Compositional Synthesis
Ruiyuan Xu, Shuoming Zhang, Guangli Li, Qiuchu Yu, Rui Zhang, Yangyu Zhang, Hao Qian, Chunwei Xia, Jiacheng Zhao, Chenxi Wang, Xiaobing Feng, Jingling Xue and 1 more
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 bottom-up speculative fusion (mutation) to top-down construction (ablation) | LEGO · mHC-MoE, H100 | relative speedup: 1.20× -> 11.29× | Improved |
| add bottom-up speculative fusion (ablation) | LEGO · SD3-MMDiT, H100 | relative speedup: 0.85× -> 2.86× | Improved |
| allow deep recursive decomposition instead of depth-1 (ablation) | LEGO · Mamba2, SD3-MMDiT, Qwen3-8B, H100 | codegen success | Improved |
| allow deep recursive decomposition instead of depth-1 (ablation) | LEGO · ResNet18, H100 | relative speedup: 0.93× -> 1.02× | No clear effect |
| use structural decomposition instead of direct monolithic codegen (ablation) | LEGO · 5 workloads, H100 | codegen success | Improved |
| use LEGO instead of other LLM kernel agents | LEGO · Table 1 | speedup | Improved |
| use LEGO instead of TorchInductor max-autotune | LEGO · mHC-MoE, H100 | speedup vs Eager: 0.82× -> 9.26× | Improved |
| use LEGO instead of Inductor max-autotune on ResNet18 | LEGO · ResNet18, H100 | speedup vs Eager: 3.87× -> 3.96× | No clear effect |
| use structure-aware FusionTree with context isolation instead of monolithic baseline | LEGO · Mamba2-1.3B | token consumption: 7.02× reduction | Improved |
| parallel critical-path exploration instead of serial monolithic search | LEGO · 5 workloads | exploration time: 1.94× to 2.47× faster | Improved |
Closest papers by meaning
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.89)
- StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning (ICML 2026 · similarity 0.87)
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training (NeurIPS 2026 · similarity 0.87)
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.86)
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.86)
- Practical Offloading for Fine-Tuning LLM on Commodity GPU via Learned Sparse Projectors (AAAI 2025 · similarity 0.85)
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.
