MARS-Bench - Instruction Following: leaderboard

Each model, in an ultra multi-turn chat over the text play-by-play of 30 games (15 English NBA, 10 Chinese NBA, 5 English NHL), must follow per-turn and inferred output-format instructions; the score is the category score. A DeepSeek-V3-0324 judge awards checklist points built from the play-by-play.

Metric: Score (0-100). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro (Preview 05-06)65.08#271
2GPT-4.5 (Preview)55.52#499
3O1 (2024-12-17)53.09#314
4DeepSeek R153.04#487
5Doubao-1.5-Thinking-Pro51.99#257
6Doubao-1.5-Pro-32k42.8#696
7GPT-4.1 Mini40.23#793
8Qwen 2.5 Max39.77#966
9GPT-4o (2024-11-20)39.28#864
10Qwen 2.5 72B Instruct38.38#1119
11DeepSeek V3 (1226)37.23#807
12GLM-Z1-Air35.75
13Claude 3.7 Sonnet34.13#476

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=mars-bench-instruction-following · How It Works · Data refreshed daily, snapshot 2026-11-17.