Individualised tutoring with LLMs

In the current era, technology and artificial intelligence are playing a pivotal role in transforming many industries and domains, and education and the learning landscape are no exceptions. We have already seen in previous blog posts how large language models such as GPT-4 and Open Assistant can transform education, and how Virtual Teaching Assistants can be used to support students. In our post, we argued that – as of now – existing large language models might not be the best solutions, as they need to be fed with knowledge about the specific course they are used for and, being much more complex than custom built VTAs, are more likely to hallucinate and provide factually incorrect responses.

While we stand by our position, large language models indeed clearly outperform custom VTAs in some tasks, such as individualised tutoring, and they are already being used on platforms such as duolingo (Duolingo Max) and Khan Academy (Khanmigo). In this blog post we will see how we can unlock the power of individualised tutoring with GPT-like models, and discuss the benefits, challenges, and ethical implications of doing so.

Benefits

Personalized Learning. LLMs empower educators and learners alike by adapting instructional content to suit individual learning styles, pace, and strengths. By analysing the needs of each student, the models can deliver customised learning experiences that optimise comprehension and retention. More specifically, the models can provide

Enhanced Engagement. Traditional teaching methods often struggle to captivate every student's attention. However, by adjusting their style and content, LLMs can (if instructed to do so) focus on maintaining students’ engagement. For instance, by incorporating interactive elements, multimedia resources, and gamification techniques, they can transform the learning process into a more interactive (and thus engaging) experience.

Targeted Intervention. LLMs are capable of identifying and addressing knowledge gaps in real-time, for instance being fed the students’ responses to some exam questions. Thus, they can provide immediate feedback, remedial explanations, and additional resources based on a student's specific areas of struggle. This targeted intervention promotes efficient learning and accelerates academic progress.

Round-the-Clock Support. Students can access assistance and guidance from LLMs at any time, provided that they have an internet connection, breaking the constraints of traditional classroom settings. This availability ensures continuous learning and empowers students to explore subjects at their own pace, instilling a sense of self-directedness. This is true for custom VTAs as well but, in the case of pre-trained LLMs, students’ can access more advanced tutoring and support.

Challenges

Domain Adaptation. LLMs have a vast amount of knowledge across several domains, and can therefore provide (most of the times) helpful tutoring and correct responses. However, it is challenging to limit the knowledge of the LLMs to a specific course and some details which the model cannot have learned from other sources, and are likely to quickly change (e.g. minutes of a lecture, or schedule of an exam session).

LLMs Hallucination. LLMs sometimes hallucinate, they make things up, and as of now there is no way to reliably prevent this. Specific countermeasures should be put in place in an online tutoring LLMs to detect when this occurs and correct (or stop) the interaction.

Bias. Large Language Models learn from vast amounts of data, and if this data is biased, it may perpetuate social inequalities and reinforce discriminatory patterns. It is crucial to implement rigorous mechanisms to identify and mitigate biases, ensuring that the tutoring experience is fair and unbiased for all students.

Lack of Human Connection. LLMs can offer personalised tutoring experiences, and these can be in some cases more engaging than “cold” interactions with human teaching assistants. However they cannot fully replicate the empathy, emotional support, and nuanced understanding that (some) human educators provide, which might be a huge limitation for some students. Therefore, striking a balance between AI-guided tutoring and human interaction, as well as understanding which tutoring approach is preferred by each learner, is essential to optimise the learning outcome of students.

Privacy and Data Security. Individualised tutoring relies on the collection and analysis of vast amounts of student data. Safeguarding this sensitive information is paramount. Stricter regulations and transparent policies are necessary to protect student privacy and prevent the misuse of data gathered during the tutoring process.

It is important to remark how some of these challenges are different from the ones that are encountered with custom built VTAs. Indeed, custom VTAs are built for a specific course and leverage its material, therefore there is no need to adapt them to different domains. Also, being architecturally simpler, they are less capable (meaning that sometimes they cannot provide any answer) but they do not hallucinate, and the output in case of unknown queries is easily controllable. The challenges for Bias, Privacy, and Data Security are also slightly different: indeed, they are still present but more easily controllable as there is more control on the design and training phases of the model, and therefore countermeasures can be put in place earlier.

Ethical Considerations

Access and Equity. Using LLMs for individualised tutoring must be accompanied by efforts to bridge the digital divide. Ensuring equitable access to technology and overcoming socioeconomic disparities will be crucial in preventing further marginalisation of disadvantaged students.

Transparent Algorithms. The inner workings of LLMs can be complex and opaque. Educators, policymakers, and technology developers must prioritise transparency, explainability, and accountability to build trust and ensure that the algorithmic decision-making processes are understandable and fair.

Ethical Guidelines. The integration of LLMs into education necessitates the development and implementation of clear ethical guidelines. These guidelines should address concerns such as data privacy, bias mitigation, informed consent, and the responsible use of AI in the tutoring process (e.g., the Responsible AI Standards from duolingo).

Conclusion

Individualised tutoring with large language models like GPT-4 presents a transformative opportunity for education. It has the potential to revolutionise the way students acquire knowledge and skills, and the benefits can be substantial, as we have seen above. However, the challenges and ethical implications we have listed in this post must be carefully taken into consideration to ensure a responsible, reliable, and equitable implementation, which is possible only by addressing the challenges such as data bias, human connection, privacy, and equity, along with ethical considerations.