idw – Informationsdienst Wissenschaft
Nachrichten, Termine, Experten
Nachrichten, Termine, Experten
idw - Informationsdienst
Wissenschaft
AI assistants have rapidly become part of everyday life – whether through our interactions with large language models like ChatGPT or in medical applications that help interpret complex datasets. Yet, despite their widespread use, AI systems still make surprisingly simple mistakes that persist even after extensive training. They also lack the ability to think logically or to truly ‘understand’ nested input. For Michael Hahn, Professor of Computational Linguistics at Saarland University, the problem lies in the architecture itself. According to Hahn, the fundamental design of large language models needs to change.
But before researchers can make progress on these issues, they first need to gain a deeper theoretical understanding of current structures. Michael Hahn will now be able to pursue this goal with the €1.4 million in funding he has been awarded from the German Research Foundation’s Emmy Noether Programme.
Today’s large language models (LLMs) are built on what is known as transformer architecture – a framework inspired by the human ability to focus attention on relevant information while ignoring less important details. Mathematically, this means prioritizing data that appears most relevant to the question at hand. ‘In doing so, these neural networks mimic another human trait, namely associative thinking, which is our ability to link ideas and memories,’ explains Professor Hahn. The AI searches vast datasets for patterns and connections, learning through continuous training. In this process of ongoing refinement, the system is exposed to new data and given corrective feedback enabling it to retrieve the right information at the right time for increasingly precise results.
But this approach has its flaws. ‘Serious errors can occur when the AI forms incorrect associations. These mistakes are compounded by the fact that current neural networks typically operate with a fixed number of layers in which the mathematical operations are carried out – thus limiting the network’s flexibility,’ explains Michael Hahn. Hahn and his team have already demonstrated mathematically that such networks make systematic errors – errors that cannot be eliminated by more training on even larger data sets or by using better prompts, i.e. more precise instructions to the AI.
According to Hahn, today’s large language models are hitting performance ceilings due to three main shortcomings. ‘Today’s LLMs are poor at handling changing conditions. They fail to update when a situation has changed.’ Hahn’s team tested this with a simple scenario in which several people pass two different books around a group. The AI’s job was to determine who holds which book at the end. The more times the books were passed around, the less accurate the AI became. In medical applications, this weakness can have potentially serious implications. ‘Medical AI systems generate connections between different types of data, such as diagnoses, medications and test results. If the AI does not assign the chronological sequence correctly and misinterprets the sequence of symptoms, diagnoses, test results and medication, there are potentially dangerous consequences for patients,’ says Hahn.
The second shortcoming is that today’s large language models lack logical reasoning. ‘AIs are not yet capable of thinking logically. Looking at the field of medicine again, if an AI is tasked with selecting the right medication for a specific clinical condition from a large database, it must be able to infer which symptoms correspond to that condition. Similarly, when assisting with a diagnosis, the AI must understand the rules doctors use to exclude certain diseases, which means it needs to understand how doctors rule out specific conditions if particular symptoms are absent. But this involves the systematic application of logical rules – something that is currently beyond the reach of today’s neural networks,’ says Hahn.
The third area where Hahn believes AI output is unreliable arises from their inability to process complex, nested inputs. ‘Large language models often fail to process intricate, layered information in a meaningful way. This becomes evident in legal contexts. Determining the liability of a person or company alleged to have harmed some other party requires an understanding of both the underlying legal principles involved and the chronology of the alleged events. Such reasoning chains, challenging even for humans, remain beyond the capabilities of the neural networks available today,’ explains Professor Hahn.
In his Emmy Noether research project, Michael Hahn will initially focus on the theoretical foundations of the transformer architecture. The aim is to gain a better understanding of the mathematical principles that underpin how neural networks arrive at their results. He will also explore how many layers these networks need in order to act more ‘intelligently’. In the next phase, Hahn plans to investigate hybrid systems or even design entirely new architectures that exhibit more predictable capabilities and that are both more reliable and more powerful than current large language models.
The Emmy Noether Programme, which is funded by the German Research Foundation (DFG), supports outstanding early-career researchers who have completed their doctorate within the past four years, have international experience and have completed a postdoctoral phase. With €1.4 million in funding, Hahn will now establish an Emmy Noether Research Group at Saarland University, working with five doctoral researchers on the project ‘Understanding and Overcoming Architectural Limitations in Neural Language Models’. This is already the third Emmy Noether Group to be approved for computer science research in Saarbrücken in 2025. The other two groups were recently launched at the Max Planck Institute for Computer Science (see press release dated 29 October 2025: https://idw-online.de/de/news860561).This is a remarkable result given that nationwide only three Emmy Noether groups focused on computer science research were funded last year (see Gepris database: https://gepris.dfg.de/gepris/OCTOPUS?language=en).
Prof. Michael Hahn, Language, Computation and Cognition Lab
Tel. +49 681 302-4343
Email: mhahn@lst.uni-saarland.de
https://www.dfg.de/en/research-funding/funding-opportunities/programmes/individu... - Emmy Noether Programme (German Research Foundation – DFG)
https://www.uni-saarland.de/en/department/lst.html - Department of Language Science and Technology at Saarland University
https://www.mhahn.info - Professor Michael Hahn’s personal website
Computer linguistics professor Michael Hahn wants to fundamentally improve large language models suc ...
Source: Thorsten Mohr
Copyright: Universität des Saarlandes
Criteria of this press release:
Business and commerce, Journalists, Scientists and scholars
Information technology, Language / literature, Social studies
transregional, national
Contests / awards, Research projects
English
You can combine search terms with and, or and/or not, e.g. Philo not logy.
You can use brackets to separate combinations from each other, e.g. (Philo not logy) or (Psycho and logy).
Coherent groups of words will be located as complete phrases if you put them into quotation marks, e.g. “Federal Republic of Germany”.
You can also use the advanced search without entering search terms. It will then follow the criteria you have selected (e.g. country or subject area).
If you have not selected any criteria in a given category, the entire category will be searched (e.g. all subject areas or all countries).
