DynaSchedBench: leaderboard
DynaSchedBench (2026) tests LLMs as step-by-step dispatching agents on calibrated dynamic flexible job shop scheduling instances with random job arrivals, machine breakdowns and uncertain processing times. The score is the mean makespan gap, in percent, to the best schedule observed on each instance across all LLMs and fixed dispatching rules, so lower is better and 0 is the floor.
Metric: Mean relative makespan gap (%): makespan above the best schedule observed for the same instance (all LLMs and 24 fixed priority dispatching rules), averaged over the 70-instance DynaSched-Subset of calibrated dynamic flexible job shop instances; LLM dispatches step by step with the L1 local observation, direct answer, two few-shot examples, temperature 0; 0 means matching the best schedule on every instance and there is no upper limit; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 8B | 1.01 | #1873 |
| 2 | Qwen 3 14B | 1.34 | #1269 |
| 3 | Qwen 3 4B | 1.45 | #2348 |
| 4 | Kimi K2 | 1.46 | #443 |
| 5 | Claude Haiku 4.5 | 1.52 | #583 |
| 6 | Qwen 3 32B | 1.65 | #972 |
| 7 | DeepSeek V3.2 | 1.68 | #419 |
| 8 | Gemini 2.5 Flash Lite | 1.76 | #964 |
| 9 | Grok 4 Fast | 1.88 | #447 |
| 10 | GPT-5 Mini | 1.93 | #364 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dynaschedbench · How It Works · Data refreshed daily, snapshot 2026-11-21.