AgentSSL: Can MLE Agents Leverage Unlabeled Data?
Agents beat the strongest human-engineered baselines in 8 of 12 settings, but choosing their best code remained a major bottleneck.
AgentSSL treats semi-supervised learning as a search for programs that use scarce labeled data alongside abundant unlabeled data. With matched compute and weak feedback, agents combined and adapted established techniques, posting gains of up to 21.8%. Selecting programs using the agent’s feedback metric, however, cut performance by as much as 47%. The researchers introduced ProgramSelectBench, containing 16,977 agent-generated and human-engineered programs, to study that selection problem.
ArXiv · AI/CL/LG's note
AgentSSL treats semi-supervised learning as a search for programs that use scarce labeled data alongside abundant unlabeled data. With matched compute and weak feedback, agents combined and adapted established techniques, posting gains of up to 21.8%. Selecting programs using the agent’s feedback metric, however, cut performance by as much as 47%. The researchers introduced ProgramSelectBench, containing 16,977 agent-generated and human-engineered programs, to study that selection problem.
ArXiv · AI/CL/LG's note
score 5