Research map / Large language models / Agents, code and math
RAAS: LLM Agentic System Architecture Search with GRPO
Jiayi Yang, Guancheng Wan, Man Zhang, Mang Ye
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 |
|---|---|---|---|
| replace absolute-score updates with CAO and MTAS in agentic supernet search | MaAS (gpt-4o-mini) · MATH | accuracy: 52.08 -> 60.87 | Improved |
| replace absolute-score updates with CAO and MTAS | MaAS (gpt-4o-mini) · HumanEval | pass@1: 92.23 -> 96.31 | Improved |
| replace absolute-score updates with CAO and MTAS | MaAS (qwen-2.5-72b) · MATH | accuracy: 51.49 -> 60.14 | Improved |
| apply RAAS to GAIA multi-step tool use | MaAS (gpt-4o-mini) · GAIA | accuracy: 18.06 -> 20.84 | Improved |
| increase cohort size N (K=5) (ablation) | RAAS · MATH | accuracy: 54.12 -> 58.26 -> 60.87 -> 61.24 | Improved |
| increase trials K to 5 at N=5 (ablation) | RAAS · MATH | accuracy: +4.45 | Improved |
| increase trials K beyond 5 (ablation) | RAAS · MATH | accuracy | No clear effect |
| increase K when cohort size N is small (ablation) | RAAS · MATH | accuracy | No clear effect |
| add CAO (contextual baseline) on top of MaAS (ablation) | MaAS · benchmark suite | accuracy | Improved |
| add MTAS (multi-trial synthesis) on top of CAO (ablation) | RAAS · benchmark suite | accuracy | Improved |
| use RAAS instead of MaAS (cost per query on MATH) | MaAS · MATH | cost per query: cost reduced by 6%, $0.31 per query | Improved |
Closest papers by meaning
- Multi-agent Architecture Search via Agentic Supernet (ICML 2025 · similarity 0.92)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
- HACO: Hedged Agent Computing for Reliable LLM Systems (NeurIPS 2026 · similarity 0.88)
- Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.87)
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.87)
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.
