MARS-Bench - Information Reasoning: 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 track the score, count lead changes and judge a player's effect from facts spread over the dialogue; 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)70.92#271
2O1 (2024-12-17)62.63#314
3Doubao-1.5-Thinking-Pro52.17#257
4GPT-4.5 (Preview)50.43#499
5DeepSeek R140.01#487
6Claude 3.7 Sonnet36.8#476
7GPT-4o (2024-11-20)36.69#864
8Doubao-1.5-Pro-32k33.63#696
9DeepSeek V3 (1226)28.71#807
10Qwen 2.5 Max26.76#966
11GPT-4.1 Mini26.39#793
12GLM-Z1-Air24.36
13Qwen 2.5 72B Instruct24.06#1119

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-information-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-17.