Do teenagers learn from rewards and punishments like adults?
Adolescents' learning was best explained by a simple reward-tracking algorithm, while adults also learned from outcomes they didn't choose and from context, which helped them avoid punishments.
Source
The Computational Development of Reinforcement Learning during Adolescence
Study at a glance
- Design
- Human experiment — Within-subject 2x2 probabilistic learning task (reward vs punishment, partial vs complete feedback) with computational model comparison between age groups
- N
- N=38 · 38 participants after IQ matching (18 adolescents aged 12-17, 20 adults aged 18-32), from 50 recruited
- Population
- Adolescents and adults recruited in London
- Outcome
- Model posterior probabilities, correct-choice rate, reaction times and post-learning cue preferences
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
Adolescents and adults chose between pairs of symbols that led to winning or losing points with 75% or 25% probability. Some pairs offered rewards and others threatened losses, and some pairs showed only the chosen outcome while others also showed the outcome of the option not chosen. The authors fitted three nested reinforcement-learning models (basic Q-learning, plus counterfactual learning, plus value contextualisation) to each person's choices and compared them, then checked predictions against behaviour and a later test.
What they found
Model comparison favoured basic Q-learning in adolescents and the full three-module model in adults, with a strong group-by-model interaction. Behaviourally, adolescents improved less at avoiding punishment and did not benefit from seeing the unchosen outcome, while adults did. Both groups learned equally well in the simplest reward-with-partial-feedback context, arguing against a general motivation or attention deficit.
The limits
What it doesn't show
The final sample was small, 38 people, after excluding participants to match non-verbal IQ, and the design is cross-sectional, so it cannot track change within individuals. The age cut-off at 18 is arbitrary and the adolescent range is wide, which the authors acknowledge. Some behavioural effects (such as the feedback-by-group interaction on choice rate) were not statistically significant, and links to brain development are inferred from other studies rather than measured.
Key terms
- Q-learning
- A basic reinforcement-learning algorithm that updates the value of the chosen action from the reward received.
- Counterfactual learning
- Updating the value of the option you did not choose, using its shown outcome.
- Value contextualisation
- Judging outcomes relative to the average value of the situation, so avoiding a loss can feel rewarding.
- Model comparison
- Choosing among candidate models by how well they fit data after penalising complexity.
- Inverse temperature
- A parameter controlling how deterministically choices follow learned values.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What distinguished Complete from Partial feedback contexts?
Common questions
Does this mean teenagers can't learn from punishment?
No; they still learned, but less efficiently than adults in punishment contexts, consistent with not re-framing 'no loss' as a good outcome.
Why fit computational models instead of just comparing accuracy?
Accuracy can't distinguish different learning strategies; model comparison asks which algorithm best explains each person's trial-by-trial choices.
Could lower motivation explain the adolescents' results?
Unlikely, since both groups performed equally in the simple reward context with partial feedback.
More on Reinforcement learning