Paper Overview
Research Area: ML Authors: Ethan Lin, Akanksha Sarkar, Ziang Liu, Kristin Branson, Jennifer J. Sun Release Date: 2026-11-12 arXiv: 2611.17323
Summary
Transferring machine learning systems to new domains faces two major bottlenecks: acquiring labeled data and performing manual pipeline engineering. Recent MLE agents show promise in automating pipeline design, but research has almost exclusively been conducted in fully supervised settings and remains untested in realistic scenarios where labeled data is scarce and unlabeled data is abundant. This work proposes Agentic Semi-Supervised Learning (AgentSSL), which frames SSL as a program search problem and studies how agents explore this space under weak supervision.
Key Findings
- Even with weak feedback signals and equal compute budgets, agents can synthesize code that combines and adapts existing SSL techniques, beating the strongest human-engineered baselines in 8 of 12 settings, with improvements of up to +21.8%.
- Program selection is a critical bottleneck: if programs are chosen according to the agent's own feedback metrics, performance can drop by up to 47%.
- To systematically study this bottleneck, the authors release ProgramSelectBench, a large-scale benchmark containing 16,977 agent-generated and human-engineered SSL programs, and evaluate unsupervised and active program selection methods.
- AgentSSL was also evaluated on a real-world ecology detection task, demonstrating both the opportunities and challenges of transferring agent-discovered strategies beyond controlled benchmarks.
Original Abstract (Excerpt)
> Applying machine learning systems to new domains introduces bottlenecks in both labeled data acquisition and manual pipeline engineering. While recent MLE agents show promise in automating pipeline design, they have been studied almost exclusively in fully supervised settings, leaving them untested in realistic regimes where labeled data is scarce but unlabeled data is abundant. 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,...
Paper link: https://arxiv.org/abs/2611.17323
*Auto-collected on 2026-11-14*