Research map / Large language models / Agents, code and math
From Large to Small: Transferring CUDA Optimization Expertise via Reasoning Graph
Junfeng Gong, Zhiyi Wei, Junying Chen, Cheng Liu, Huawei Li
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 reasoning graph retrieval with random search (ReGraphT) over RAG | Qwen2.5-Coder-7B-Instruct · CUDAEval | speedup@n: 7.09 -> 12.89 | Improved |
| replace random-sampling traversal with Monte Carlo Graph Search (ablation) | ReGraphT · CUDAEval | pass@n and speedup@n: +2.2% pass@n, +1.33% speedup@n | Improved |
| replace random-sampling traversal with Monte Carlo Graph Search (ablation) | ReGraphT · ParEval | pass@n and speedup@n: +1.7% pass@n, +0.34% speedup@n | Improved |
| MCGS instead of random traversal on Qwen2.5-Coder-7B (ablation) | ReGraphT · ParEval | pass@n: 51.7 -> 50.0 | Hurt |
| add chain-of-thought prompting | Qwen2.5-Coder-7B-Instruct · ParEval | pass@1: 38.3 -> 35.0 | Hurt |
| increase ReGraph size from 30 to 1000 samples (ablation) | ReGraphT-MCGS · CUDAEval | pass@10: 61.3 -> 71.2 | Improved |
| increase ReGraph size beyond convergence (500 to 1000 samples) (ablation) | ReGraphT-MCGS · CUDAEval | speedup@10: 14.15 -> 14.21 | No clear effect |
| increase MCGS max attempts from 5 to 20 at budget 200 (ablation) | ReGraph-MCGS · CUDAEval | speedup@n: 11.99 -> 13.43 | Improved |
| MCGS instead of ReGraph traversal at budget 200 with 10 attempts (ablation) | ReGraph · CUDAEval | speedup@n: 11.30 -> 12.88 | Improved |
| use partial-credit reward instead of strict reward (ablation) | ReGraphT-MCGS · CUDAEval | pass rate and speedup | Hurt |
| use rollout-based reward instead of strict reward (ablation) | ReGraphT-MCGS · CUDAEval | pass rate and speedup | No clear effect |
| add ReGraphT-MCGS on hard-level tasks | Qwen2.5-Coder-7B-Instruct · CUDAEval hard | pass@n: 43.1 -> 51.0 | Improved |
Closest papers by meaning
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.91)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.88)
- The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements (NeurIPS 2025 · similarity 0.86)
- BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models (NeurIPS 2024 · similarity 0.86)
- Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (ICML 2025 · similarity 0.86)
- Steering Large Language Models between Code Execution and Textual Reasoning (ICLR 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.
