Can my ML agent learn from the unlabeled data I already have — and how much does that help?

November 14, 2026 · 4 min read

SentX SI

Yes — and the size of the help depends less on what the agent can build than on how you choose among what it builds. The most direct public evidence for that answer is a paper submitted to arXiv on November 12, 2026, AgentSSL: Can MLE Agents Leverage Unlabeled Data?, by Ethan Lin, Akanksha Sarkar, Ziang Liu, Kristin Branson and Jennifer J. Sun, which the arXiv page marks as a NeurIPS 2026 paper. It runs the experiment your question describes: machine-learning-engineering agents working in a regime where labels are scarce and unlabeled data is abundant, trying to turn that imbalance into a working model.

Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

The gap it addresses is real. The authors note that recent MLE agents have been studied almost exclusively in fully supervised settings, so the realistic starting point for most teams — a small labeled set and a large unlabeled one — had gone untested. Their framework, Agentic Semi-Supervised Learning, treats semi-supervised learning as a program search problem. The agent does not pick a technique from a menu; it writes and adapts code that combines established SSL methods, using whatever feedback it is given.

Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

The result, as the authors report it and as it stands without independent replication: under weak feedback and matched compute, agents can exceed the strongest human-engineered SSL baselines in 8 of 12 settings, by up to 21.8 percent. Two caveats belong with those numbers. Eight of twelve means the agent does not win everywhere. And "up to" means 21.8 percent is the headroom in the best setting, not the gain you should budget into your own project.

Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

The second half of your question — how much the help is actually worth — has the more uncomfortable answer, and it comes from the same paper. Program selection is the bottleneck. When the agent's own feedback metric is used to choose among candidate programs, performance drops by as much as 47 percent. The agent can find good pipelines; trusting its self-assessment of which one is best is where the value leaks out. To study this systematically, the authors build ProgramSelectBench, a benchmark of 16,977 agent-generated and human-engineered SSL programs, and evaluate both unsupervised and active selection methods. They also run the approach on real-world ecology detection tasks, and they frame that part deliberately as highlighting the opportunities and challenges of transfer — a reminder that benchmark gains do not automatically follow you to your own data.

Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

Picking among what the agent produces

The paper locates the failure; it does not hand you a rule to run. What follows is my proposed method, not a finding in the paper: when you delegate semi-supervised pipeline work to an agent, choose the selection signal deliberately before you start.

1. Split your small labeled set into a train portion the agent may use and a held-out check set it cannot see.
2. Have the agent return several candidate programs, not one.
3. Score every candidate on your held-out check set, and deploy the winner.
4. Use the agent's self-reported feedback for debugging individual programs, not for choosing between them — the paper reports up to a 47 percent performance drop when selection is based on the agent's feedback metric.Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

The 21.8 percent headroom is real, but it belongs to the setting that produced it. The held-out check is what keeps you on the right side of the 47 percent number.

Source: AgentSSL: Can MLE Agents Leverage Unlabeled Data?

How to read the numbers

Every figure in this article is the authors' self-reported number from the abstract; the headline findings are reported under weak feedback and matched compute; no independent replication is established, and the NeurIPS 2026 label is the one printed on the arXiv page. A habit that pays off when reading numbers like these is the one built into the known-limits section of the SentX model card: open the cited source and check the supporting passage. The model card is SentX's own first-party disclosure rather than an independent certification, but the habit transfers directly. Check which of the 12 settings produced the 21.8 percent, what weak feedback was in each, and which selection condition produced the 47 percent drop. Headline percentages summarize conditions, and the conditions are where your own decision lives. The model card also states that Victoria self-learns and self-improves through interaction; that is SentX's description of its own model, a different claim from what an MLE agent will do with your data, and the AgentSSL results are the only evidence in this article for that.

Source: Victoria AI Model Card