MARS: Enabling Autoregressive Models Multi-Token Generation
Open paper
Benchmark
Model
Rank
Results
code-generation-on-humaneval
MARS-7B (B=4)
#53
Pass@1: 81.7
code-generation-on-humaneval
MARS (Qwen2.5-0.5B-Instruct)
#112
Pass@1: 40.2
multi-task-language-understanding-on-mmlu-pro
MARS-0.5B (B=4)
#137
Accuracy: 12.4
Rank counts only results with a code link.