How Duolingo Keeps Me Learning French (and What We Can Learn From It)

Last year, I challenged myself to start a 1,000-day Duolingo streak to finally take my French learning seriously. As I write this, I’m on Day 425 and honestly, one of the biggest reasons I’ve stuck with it is simple: Duolingo’s notifications.

If you’ve used the app, you know what I mean. The reminders are quirky, timely, and sometimes downright cheeky (“Your French won’t learn itself, you know ”). But behind that playful tone lies a fascinating machine learning system that keeps millions of people, including me, consistently engaged.

The Multi-Armed Bandit Problem

At the core of Duolingo’s system is something called the multi-armed bandit algorithm.

Think of a person standing in front of multiple slot machines (a.k.a. “one-armed bandits”). Each machine gives a different, unknown reward. The challenge?

• Explore: Try different machines to discover which ones pay best.
• Exploit: Stick with the machine that seems to give the highest reward.

Duolingo applies this exact logic to notifications. Each type of reminder (playful, urgent, encouraging, etc.) is like a different slot machine. The app constantly tests which type is most likely to get you back into a lesson within two hours.

Enter RDSA: The Secret Sauce

But Duolingo doesn’t stop at just testing options. They also use something called the Recovering Difference Softmax Algorithm (RDSA) — a mouthful, but here’s the simple idea:

• Each notification gets a score based on how effective it is for you.
• Notifications that worked well in the past are more likely to be chosen again.
• To keep things fresh, the algorithm penalises recently sent notifications, so you don’t keep seeing the same one.
• Finally, it runs all these scores through a Softmax function, which turns them into probabilities.

In plain English: the algorithm learns over time which nudges work best for each user, while avoiding repetition. It balances trying new strategies with sticking to proven ones.

The Results Speak for Themselves

Duolingo presented this work at the KDD Conference 2020, showing that their approach contributed to a 350% increase in user retention. That’s massive.

For me, this explains why I’ve stayed motivated for over a year. Those little notifications aren’t random. They’re the result of carefully designed, adaptive algorithms that personalise the experience.

Lessons Beyond Duolingo

Duolingo’s system is more than just a clever trick to get people back on the app. It’s a lesson in how data-driven, adaptive tools can support long-term learning. By balancing exploration (trying new nudges) and exploitation (reinforcing effective ones), any platform , whether for languages, fitness, or professional skills , can keep learners engaged without burning them out.

If a streak-based notification system can get me to practice French every single day, imagine what similar systems could do for other areas of learning.

For those curious, you can read Duolingo’s research paper here.