AgentCE-Bench - Course Scheduling: leaderboard

AgentCE-Bench (2026) tests tool-using agents on grid planning: the agent fills hidden slots of a weekly schedule by querying static JSON data, with decoy candidates that break global constraints. Task score in percent on the course scheduling domain; higher is better.

Metric: Task score (%) on AgentCE-Bench course domain (54 instances): the agent fills the hidden slots of a 5x7 planning grid with tool calls on static JSON (attribute filter queries, slot and global constraint checkers), each hidden slot having 25 candidates including decoys that violate only the global constraints; every cell is a whole count of instances over the 54 per domain (6 hidden-slot counts by 9 decoy budgets); at most 600 steps; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 397B A17B94.4#264
2Qwen 3.5 122B A10B81.5#346
3Qwen 3.5 27B79.6#375
4Qwen 3.5 35B A3B68.5#492
5MiniMax-M2.159.3#575
6MiniMax-M2.557.4#600
7Qwen 3.5 9B51.9#828
8GLM-4.7 FP844.4#335
9Qwen 3.5 4B38.9#1241
10MiniMax-M238.9#689
11AgentCE-Bench MiroThinker-mini 30B (checkpoint unspecified)14.8
12Qwen 3.5 2B3.7#5307
13Qwen 3.5 0.8B1.9#8232

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=agentce-bench-course-scheduling · How It Works · Data refreshed daily, snapshot 2026-11-21.