Research map / Large language models / Retrieval and knowledge
Structured Progressive Knowledge Activation for LLM-Driven Neural Architecture Search
Zhen Liu, Yuhan Liu, Jinjun Wang, Wei Song, Jianyi Liu, Jingwen Fu
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 |
|---|---|---|---|
| use SPARK factor-scoped editing instead of free-form OpenEvolve edits | OpenEvolve with DeepSeek-R1 · CLRS DFS (OOD) | OOD accuracy (%): 32.54 -> 83.74 | Improved |
| use SPARK instead of EvoPrompting, in evaluated candidates | EvoPrompting · CLRS DFS | sample efficiency: 28.1x (1600/57 evaluations) | Improved |
| use SPARK on full 30-task CLRS suite | CLRS reference and EvoPrompting · CLRS 30 tasks | mean OOD accuracy (%): 75.98 -> 83.91 (EvoPrompting 80.89) | Improved |
| add ASR scope selection to RC+SAR (ablation) | SPARK · CLRS DFS (DeepSeek-R1) | OOD accuracy (%): 56.79 -> 83.74 | Improved |
| add RC+SAR refinement to ASR-only (ablation) | SPARK · CLRS DFS (DeepSeek-R1) | OOD accuracy (%): 65.28 -> 83.74 | Improved |
| ablation components repeated with Qwen-Plus editor (ablation) | SPARK · CLRS DFS (Qwen-Plus) | OOD accuracy (%): 64.50 -> 80.50 | Improved |
| run multi-round evolution instead of a single iteration (ablation) | SPARK · CLRS 10-task subset | mean OOD accuracy (%): 72.58 -> 83.92 | Improved |
| constrain edits to one factor scope via ASR (ablation) | OpenEvolve · CLRS DFS search over 100 attempts | cumulative valid rate: 0.25-0.3 -> 0.5-0.6 | Improved |
| constrain edits to one factor scope via ASR (ablation) | OpenEvolve · CLRS DFS search | entanglement rate | Improved |
| use SPARK instead of OpenEvolve on MNIST-1D | OpenEvolve · MNIST-1D | accuracy (%): 70.25 -> 95.99 | Improved |
| use SPARK instead of OpenEvolve on CartPole-v1 | OpenEvolve · CartPole-v1 | mean test return: 377.55 -> 489.00 | Improved |
| use SPARK instead of OpenEvolve on Mice Protein Expression | OpenEvolve · Mice Protein Expression | accuracy (%): 61.78 -> 61.78 | No clear effect |
Closest papers by meaning
- Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search (NeurIPS 2026 · similarity 0.88)
- RZ-NAS: Enhancing LLM-guided Neural Architecture Search via Reflective Zero-Cost Strategy (ICML 2025 · similarity 0.88)
- Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (ICML 2025 · similarity 0.87)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.86)
- An Architecture Search Framework for Inference-Time Techniques (ICML 2025 · similarity 0.86)
- Design Principle Transfer in Neural Architecture Search via Large Language Models (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.
