AgentSSL: Can MLE Agents Leverage Unlabeled Data?
arXiv cs.AI · 13 Nov 2026, 04:59 AEST
AgentSSL introduces a new approach to semi-supervised learning by framing it as a program search problem. The study reveals that agents can outperform human-engineered baselines in many scenarios, but face challenges in program selection.
Key facts
- Applying machine learning systems to new domains introduces bottlenecks in both labeled data acquisition and manual pipeline engineering.
- In this work, we introduce Agentic Semi-Supervised Learning (AgentSSL), framing SSL as a program search problem and studying how agents navigate this space under limited supervision.
- We find that even under weak feedback and matched compute, agents can synthesize code that combines and adapts established SSL techniques, exceeding the strongest human-engineered baselines in 8 of 12 settings by up to +21.8%.
- However, our results identify program selection as a critical bottleneck in AgentSSL, where performance drops by as much as 47% when selecting based on the agent's feedback metric.
- To systematically study this bottleneck, we introduce ProgramSelectBench, a large-scale benchmark with a combined total of 16,977 agent-generated and human-engineered SSL programs.
Read aloud
AgentSSL introduces a new approach to semi-supervised learning by framing it as a program search problem. The study reveals that agents can outperform human-engineered baselines in many scenarios, but face challenges in program selection. Applying machine learning systems to new domains introduces bottlenecks in both labeled data acquisition and manual pipeline engineering. In this work, we introduce Agentic Semi-Supervised Learning AgentSSL, framing SSL as a program search problem and studying how agents navigate this space under limited supervision. We find that even under weak feedback and matched compute, agents can synthesize code that combines and adapts established SSL techniques, exceeding the strongest human-engineered baselines in eight of twelve settings by up to +twenty one point eight percent. However, our results identify program selection as a critical bottleneck in AgentSSL, where performance drops by as much as forty seven percent when selecting based on the agent's feedback metric. To systematically study this bottleneck, we introduce ProgramSelectBench, a large-scale benchmark with a combined total of sixteen thousand nine hundred seventy seven agent-generated and human-engineered SSL programs.
Sources in this cluster
- AgentSSL: Can MLE Agents Leverage Unlabeled Data?
arXiv cs.AI · 13 Nov 2026, 04:59 AEST