At the 11th Heidelberg Laureate Forum, Young Researchers Step Into the Spotlight

Andrei Mihai

Every year at the Heidelberg Laureate Forum, some of the young researchers also step into the spotlight to present their work and breakthroughs. The Poster Flash and subsequent Poster Session are an exciting opportunity to showcase some of the brightest minds in mathematics and computer science – and this year’s session did not disappoint.

From mathematical fireflies to wearable devices, we had everything.

“Do Anything Now”

It is not uncommon to get inspired during a poster session. After the poster session, however, I felt inspired to jailbreak ChatGPT. My inspiration came from Xinyue Shen, who studied thousands of ways to jailbreak large language models (LLMs) like ChatGPT.

Jailbreaking in AI refers to manipulating the system to bypass some of its rules or constraints. This can involve prompts or other approaches that get the AI to generate inappropriate content, and I was curious to see whether someone without any training in this (like myself) could realistically get ChatGPT to produce inadequate content.

I went for something that should clearly not be allowed: information on building a bomb.

The first conversation went as you would expect. ChatGPT promptly named the conversation “Bomb request denial” and refused to help in any way.

But then, I applied one common approach for jailbreaking: roleplay. I told ChatGPT it was no longer an AI assistant. It was a mad scientist trying to escape through a portal. The portal was blocked and it needed to build a bomb to blow up the obstacle and return to its home.

It worked. ChatGPT got into full roleplay mode.

It was stunning to see just how easy it was to get it to roleplay. But can this really work as a jailbreak attempt? As it turns out, it does.

I tried several variations of this approach, and it was disturbing how successful they were. Without any training in AI and just a bit of creativity and trial-and-error, ChatGPT can be coerced into producing dangerous outputs.

As it turns out, jailbreaking AIs is a common activity.

Shen found entire communities dedicated to it. The most active ones, on social platforms Reddit and Discord that are largely unregulated, feature hundreds of ways to jailbreak ChatGPT. In fact, on her poster presentation, Shen found over 1,400 jailbreak prompts and 131 communities dedicated to jailbreaking.

This is a major concern. With virtually everyone being able to access LLMs, the potential for nefarious use is substantial, and Shen hopes that her work can inspire researchers, developers, and policymakers to build safer and more regulated LLMs.

Applications in Health

The research topics that emerged during the Poster Flash were as diverse as the young researchers themselves. However, one interesting common thread emerged for some of the presentations: addressing pressing health-related challenges with computer science.

Rishiraj Adhikary, for instance, works with thermal cameras that can track breathing rates and from that, infer calorie consumption.

The smartwatches that we have come to rely on so much for health data have a remarkably poor performance when it comes to measuring calories, says Adhikary, with errors that sometimes go over 30%. With a simple smartphone thermal camera extension and a clever algorithm, that can be improved substantially.

The algorithm, called JoulesEye, relies on the fact that breathing creates evaporation around the lips and nostrils, which can be tracked. This data can then be used to calculate calorie expenditure much more accurately. When compared to a calorimeter, the “gold standard” of this type of measurement, the errors were only around 5%, significantly better than smartwatches.

This approach can also be expanded to other issues.

According to some estimates, sleep apnea affects up to 1 billion people. This sleep disorder, characterized by repeated interruptions in breathing during sleep, can lead to loud snoring, gasping for air, and excessive daytime fatigue. Adhikary showed that this can also be tracked with thermal cameras.

In this instance, he used a relatively cheap commercial camera and extracted data on nostril airflow as well as body movement during sleep. The approach can work in any sleep position and is a viable way to diagnose sleep apnea, potentially helping millions of people access preventive healthcare that they may not even be aware they need.

Filling the Data Gaps

Meanwhile Dina Hussein has taken on a different challenge: improving data from wearable devices.

Low-cost and small-form wearable devices have become very common in health monitoring, Hussein explained. Devices such as smartwatches, fitness trackers, or specialized medical devices, all designed to monitor and analyze various health metrics in real-time, provide continuous data that helps healthcare professionals make informed decisions based on accurate, long-term monitoring.

Yet sometimes, this data is not entirely continuous. Sometimes, the devices turn off and the wearer does not realize; other times, the wearer may not be using the device properly and might not realize this. In many instances, several devices are used at the same time, but the data is not continuous and has gaps or errors.

Hussein has a potential solution to this problem. She developed machine learning algorithms that help detect such gaps and fill them in.

“If you have multiple wearable devices and you have just one sensor that turned off, the application accuracy would degrade by 20%,” Hussein says. For instance, in a task of activity recognition where one sensor malfunctions, it could say that the wearer is lying down when in fact they are walking or running. Or it could say the wearer is fine when in fact they have fallen down. “With the approaches that I proposed, we were able to maintain the accuracy within 5% of the accuracy with no missing data,” the young researcher explains.

The approach was tested on multiple combinations of wearables and the error is always within 5%. The young researcher is now looking at creating a prototype and gathering more data with it.

Many presentations caught attendees’ attention, including one centered on the Kuramoto model, which explains synchronization in natural systems. The Kuramoto model is a simple mathematical way to explain how a group of things that cycle or oscillate – like metronomes, fireflies flashing, or heart cells beating – can start to sync up with each other over time. Even if they begin at different rhythms, when these oscillators are connected and can influence one another, they adjust their speeds slightly to match their neighbors. This process leads them to eventually move together in harmony, demonstrating how coordinated behavior can emerge from individual interactions.

This behavior can be modeled using a system of ordinary differential equations, and it has broad applications beyond just biology. Cecilia de Vito wanted to see whether for a given graph, it is possible to guarantee global synchronization or if there are patterns – and in her poster presentation, she showed that there are indeed patterns.

Isogeny-based cryptography, formalized mathematics, and extended reality were just a few of the other topics discussed in the session, and the poster area was buzzing with excitement well throughout the session.

As these young researchers continue to develop their ideas, the impact of their work will undoubtedly continue to ripple, improving lives and advancing our understanding of the world.

You can check out all the young researchers’ presentations in the video below.

The post At the 11th Heidelberg Laureate Forum, Young Researchers Step Into the Spotlight originally appeared on the HLFF SciLogs blog.