List your company

Automated Program Repair on HumanEval Java (164 tasks)

84.5Pass@1 RateBOOSTAPR (+ Rline)
57.4664.4871.578.52May 9, 2026

Evaluation results

MethodLinks
BOOSTAPR (+ Rline)Params=32B2026.0584.5
+ Stage III (PPO, Rseq only)Params=32B2026.0579.4
SWE-RLBackbone=Llama-3-70B, Params=70B2026.0576.2
RLEFBackbone=Llama-3-8B, Params=8B2026.0574.3
SWE-FixerBackbone=Qwen2.5-72B, Params=72B2026.0573.8
+ Stage I (SFT)Params=32B2026.0573.1
Lingma SWE-GPTBackbone=Qwen2.5-72B, Params=72B2026.0572.6
ChatRepairBackbone=GPT-3.5-turbo2026.0572
AgentlessBackbone=GPT-4o2026.0571.3
SWE-GymBackbone=Qwen2.5-Coder-32B, Params=32B2026.0570.7
SWE-agentBackbone=Claude 3.5 Sonnet2026.0568.9
RepairLLaMABackbone=CodeLlama-7B, Params=7B2026.0566.5
AutoCodeRoverBackbone=GPT-4o2026.0565.2
Qwen2.5-Coder-32B (base)Params=32B2026.0564
CodeRLBackbone=CodeT5-large, Params=770M2026.0563
KNODBackbone=CodeT5-base, Params=220M2026.0558.5