AI & softwareResearch findings
AgentSSL: A Method for MLE Agents to Leverage Unlabeled Data
Original published 2026.11.13 03:59 KSTAdded to Blackdeck 2026.11.14 12:50 KST
What has changed?
AgentSSL proposes a method for MLE agents to combine and adapt established SSL techniques using unlabeled data.
Key points
- AgentSSL introduces a framework where MLE agents combine and adapt SSL techniques using unlabeled data.
- In 8 out of 12 settings, agent-generated code outperformed the strongest human-engineered baselines by up to +21.8%.
- Program selection based on agent feedback leads to a significant performance drop of up to 47%.
Connect it to your work
The approach frames semi-supervised learning as a program search problem, enabling agents to explore and synthesize effective strategies in data-scarce settings
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.