Research

When does reinforcement learning succeed? New research that sheds light on DLPC®

Nadav Cohen's group proved that the structure of training data, not just its volume, decides whether reinforcement learning succeeds beyond what it trained on.

AI Research Spotlight graphic reading "When does RL succeed? It starts with the right data."

Imubit's CTO Prof. Nadav Cohen and his group at Tel Aviv University develop mathematical theory that explains when reinforcement learning (RL) succeeds. Their latest paper will appear next month in NeurIPS, the most prestigious venue in machine learning, and its findings shed light on what makes reinforcement learning work in Imubit's DLPC®.

The paper answers a question customers often ask: how will a model perform when a unit runs in conditions different from its history? It depends on the training data. With the right mix of examples, RL successfully learns and performs well beyond the conditions it trained on. When the data includes enough simpler examples, the model learns the underlying structure and carries it to harder problems it has never seen. Train it only on the hardest cases, and learning becomes intractable, even with exponential training time.

They confirmed the theory on synthetic tasks and on real models. The lesson is that the structure of the training data, not just its volume, decides whether reinforcement learning succeeds.

DLPC® is the only commercial solution applying closed-loop reinforcement learning at scale in a production environment, and it has long been backed by theory. Results like this one sharpen our understanding of when RL succeeds, and when it provably cannot. It is why we map and model a unit, making sure the data is right, before we ever close the loop.

Read the paper: "Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data" (Ran-Milo, Alexander, Mendel, Cohen), NeurIPS 2026

Where to start

Find out what your gap to plan is costing you.

  • One to two hours for your team to prepare and upload the data package

  • Results within a few working days of approval

  • No workshops

Send your data as it is. Gaps, outages and noisy tags are normal, and our engineers screen and clean them in the platform. If anything limits a use case, we tell you.