SCHEDBench (Solver Code): leaderboard
SCHEDBench (2026) verbalizes canonical job-shop, project-scheduling, nurse-rostering and timetabling instances as natural-language problems and checks whether large language models produce constraint-feasible schedules. Feasibility is the share of instances whose output satisfies all hard constraints under an exact OR-Tools verifier, in percent, higher is better, on the setting where the model writes solver code, executed in a sandbox, rather than a schedule directly.
Metric: Feasibility rate (%; share of SCHEDBench natural-language scheduling instances whose model output satisfies all hard constraints, verified by OR-Tools CP-SAT; six canonical job-shop, project-scheduling, nurse-rostering and timetabling families; greedy decoding, zero-shot, no external solver; the model writes solver code executed in a sandbox rather than a schedule directly; full surface variation). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (2026-03-05) | 39.6 | #84 |
| 2 | Qwen 3.5 397B A17B | 27.5 | #242 |
| 3 | GPT-5.4 Mini (2026-03-17) | 27 | #285 |
| 4 | Gemini 3.1 Flash Lite | 20.6 | #363 |
| 5 | Qwen 3.5 122B A10B | 19.4 | #329 |
| 6 | Qwen 3.5 27B | 14.6 | #358 |
| 7 | Claude Haiku 4.5 | 6 | #568 |
| 8 | Llama 4 Maverick | 1.3 | #1105 |
| 9 | Llama 3.3 70B | 0.6 | #1511 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=schedbench-solver-code · How It Works · Data refreshed daily, snapshot 2026-11-12.