Benchmark
MARS-Bench
What is MARS-Bench?
Multi-turn dialogue benchmark built from sports commentary; evaluates ultra multi-turn, cross-turn, and interactive multi-turn capabilities.
- Evaluates
- reasoning
- Openness
- unknown
- Catalog source
- Model reports
- Reported scores
- 3
MARS-Bench links
MARS-Bench results and reported scores
Model reports
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Seed2.0 Lite | ByteDance | 80.5 | 2026-02-14 | evidence |
| Seed2.0 Mini | ByteDance | 62.4 | 2026-02-14 | evidence |
| Seed2.0 Pro | ByteDance | 85.6 | 2026-02-14 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite MARS-Bench?
- Seed2.0 (Pro / Lite / Mini) ByteDance, 2026-02-14