ResearchBenchmarksCode Generation on HumanEval (avg@4)Code Generation on HumanEval (avg@4)Follow85.5Pass@4 (HumanEval)SHEAR64.90870.25475.680.946Apr 25, 2026Evaluation resultsMethodMethodLinksPass@4 (HumanEval)SHEARBackbone=Qwen2.5-14B-BaseBackbone=Qwen2.5-14B-Base2026.0485.5GRPOBackbone=Qwen2.5-14B-BaseBackbone=Qwen2.5-14B-Base2026.0485SHEARBackbone=Qwen2.5-Coder-7BBackbone=Qwen2.5-Coder-7B2026.0476.1GRPOBackbone=Qwen2.5-Coder-7BBackbone=Qwen2.5-Coder-7B2026.0475.8SHEARBackbone=Llama3.1-8B-I...Backbone=Llama3.1-8B-Instruct2026.0470GRPOBackbone=Llama3.1-8B-I...Backbone=Llama3.1-8B-Instruct2026.0465.7