Research map / Large language models / Agents, code and math
OrchestrationBench: LLM-Driven Agentic Planning and Tool Use in Multi-Domain Scenarios
Aelim Ahn, Sooyeon Lee, Hyosun Wang, Chiwan Park, Daeryong Kim, Jihyeon Roh, Kichang Yang, Wonjun Jang, Hwang Woosung, Min Seok Kim, Jihoon kang
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 Claude-sonnet-4 on English instead of Korean data | claude-sonnet-4 · OrchestrationBench | call/reject accuracy: 0.7586->0.8682 | Improved |
| Larger or stronger planner (Gemini-2.5-pro instead of GPT-4.1-nano) | main LLM planning · OrchestrationBench English | planning score (1-GED): 0.462 -> 0.850 | Improved |
| Use GPT-5 with low reasoning effort for planning instead of GPT-4.1 | gpt-5 · OrchestrationBench English | planning score: 0.744 -> 0.583 | Hurt |
| Use dense Qwen3-235B-A22B instead of MoE variants for planning | Qwen3 · OrchestrationBench English | planning score: 0.768 | Mixed |
| Compare planning vs function calling spread across models | LLM orchestration · OrchestrationBench | score variation: 0.057 to 0.850 | Mixed |
| Correlate workflow score with function calling score | models · OrchestrationBench English | correlation: 0.9215 | Improved |
| Correlate workflow score with call/reject accuracy | models · OrchestrationBench Korean | correlation: 0.4480 | No clear effect |
| Use Claude-sonnet-4 for function calling instead of Gemini-2.5-pro | claude-sonnet-4 · OrchestrationBench English | function calling score: 0.836 -> 0.885 | Improved |
| Use Gemini-2.5-pro for Korean call/reject instead of Qwen3-30B-A3B | Gemini-2.5-pro · OrchestrationBench Korean | call/reject accuracy: 0.449 -> 0.813 | Improved |
Closest papers by meaning
- T1: A Tool-Oriented Conversational Dataset for Multi-Turn Agentic Planning (NeurIPS 2025 · similarity 0.91)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.90)
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.90)
- SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.89)
- NaviAgent: Graph‑Driven Bilevel Planning for Scalable Tool Orchestration (ICML 2026 · similarity 0.89)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)
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.
