COMPTEACH delivered a set of experimental, computational, and language-based results that clarify how teaching and advice shape learning from experience beyond the current state of the art. These include:
(i) A validated teacher-to-pupil experimental framework was developed in a reinforcement-learning setting: teachers learn a probabilistic decision task through trial and error, then write free-text lessons that pupils use before performing the same task. Of note, teaching effectiveness was demonstrated behaviourally: pupils receiving higher-quality lessons performed better than pupils receiving lower-quality lessons or no lesson.
(ii) A new NLP/LLM-based method (LLM-DISC) characterised teaching language by decomposing lessons into four interpretable semantic dimensions—Memorization, Pattern Recognition, Option Ranking, and Randomness—capturing what teachers emphasise when explaining how to succeed.
(iii) The work moved from description to intervention: the extracted semantic “recipe” was used to generate synthetic lessons that reproduced the behavioural benefit of helpful human teaching in an independent sample, showing that specific instructional content can causally improve learning.
(iv) The teacher's perspective was conceptualized as advice-giving during learning, and was mapped as a function of growing expertise and shown to depend on incentives (monetary costs and reputational motives) and individual traits such as social anxiety, which was associated with more conservative information sharing.
(v) Additional work benchmarked human and LLM reasoning using contamination-controlled cognitive tasks, identifying both similarities and qualitative differences across model generations and prompt conditions.
Potential impacts (indicative)
(i) Establishes building blocks for a computational science of teaching, linking properties of explanations to measurable changes in learning.
(ii) Provides a principled basis for future applications in education and training, by identifying teachable components of effective written guidance and showing they can be engineered into instructions.
(iii) Supports development of human-centred tutoring and AI-assisted learning tools, by offering empirically grounded targets for generating or evaluating instructional text.
(iv) Opens pathways for future research in health and wellbeing contexts where learning and social information exchange may be altered, by providing controlled paradigms and measures to study guidance and feedback.
Key needs to ensure further uptake and success, for which the candidate will be applying for an ERC starting grant
(i) Scale-up and replication across larger paired teacher–pupil samples, additional tasks, and diverse populations to strengthen generalisability and enable more fine-grained computational tests.
(ii) Demonstration in more ecological settings (e.g. classroom-like or applied training contexts) while preserving measurement precision.
(iii) Reusable open infrastructure (tasks, analysis pipelines, and shareable anonymised datasets where feasible) to facilitate adoption by other groups.
(iv) Interdisciplinary partnerships with education researchers, practitioners, clinicians, and (where relevant) developers to translate scientific measures into usable interventions.
(v) Resources for translation, including funding for demonstration studies and user-centred evaluation; and, if moving toward deployment, appropriate support for data governance, ethics, and (where relevant) IPR strategy.
