Addressing the limitations of Large Language Models

Large language models (LLM) are a very powerful tool that can be used in a wide range of different applications and they are currently the main driving force of progress in the field of artificial intelligence (AI) for several reasons – e.g. because they help AI systems incorporate a wide range of general knowledge about the world, they are able to follow natural-language instructions, and perform many tasks in few-shot mode, i.e. based on a very small number of examples, thanks to their in-context learning capabilities. They are also able to integrate other modalities (e.g. image and audio).

Despite this unprecedented progress, LLMs also suffer from several significant limitations that currently prevent their wider and safe use in many domains. These restrictions include e.g. the tendency to generate answers that have no support in the training corpus or in the input context (hallucinations), limited ability to perform multi-step reasoning and planning (and apply critical reasoning during training), but also difficulties associated with the integration of other data modalities such as a limited ability to recognize fine-grained visual concepts, etc. LLMs are also much less sample efficient than humans when acquiring new knowledge and skills, which is a significant challenge in some cases – especially for low-resource languages.

The aim of this research is to examine such limitations and – after focusing on one or two of them – propose strategies to mitigate them. Such strategies may include e.g.:

• Developing the ability to perform reasoning e.g. by building upon the boostrapping reasoning paradigm, adjusting the training paradigm, training on less traditional tasks (e.g. from the reinforcement learning domain), etc.;
• New, more effective self-correction mechanisms and self-evaluation pipelines;
• Improvement of multimodal properties of models, e.g. the ability to recognize fine visual concepts;
• Reducing the rate of hallucinations e.g. by designing new training and fine-tuning techniques, new kind of LLM pipelines, etc.;
• Mechanisms for reasoning during the training process, supporting the ability to better contextualize the content (e.g. understanding that the text is meant ironically, that it is of lower quality, contains false information, etc.);
• An active training paradigm where models reason and distill during training to acquire new knowledge and skills with improved sample-efficiency;
• …

The application domain might be e.g. support for fact-checking and disinformation combatting, where many of these shortcomings are absolutely critical – but there, of course, is a range of other options.

Relevant publications:

• Srba, I., Pecher, B., Tomlein, M., Moro, R., Stefancova, E., Simko, J. and Bielikova, M., 2022, July. Monant medical misinformation dataset: Mapping articles to fact-checked claims. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 2949-2959). https://dl.acm.org/doi/10.1145/3477495.3531726
• Pikuliak, M., Srba, I., Moro, R., Hromadka, T., Smolen, T., Melisek, M., Vykopal, I., Simko, J., Podrouzek, J. and Bielikova, M., 2023. Multilingual Previously Fact-Checked Claim Retrieval. https://arxiv.org/abs/2305.07991

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.