Summary: Researchers tracked 90 adult learners attempting to decipher Swahili object placement instructions given by a robot. The findings demonstrate a distinct temporal trade-off: highly detailed, personalized error feedback slowed down immediate recovery on subsequent attempts due to cognitive load, yet drove superior overall task performance over time.
The results highlight that effective robotic tutors must dynamically balance personalization with real-time cognitive and affective assessment, knowing when brief guidance is superior to detailed instruction.
Key Facts
- Swahili Spatial Puzzle Task: To test real-time feedback processing, 90 adult participants had to place household objects (bottle, cup, book) in correct room locations based on instructions delivered by a humanoid robot in Swahili, a language unknown to the learners.
- Three Feedback Paradigms Tested: Researchers evaluated non-adaptive fixed feedback, performance/enjoyment-adapted feedback, and fully personalized feedback referencing the learner’s specific error history and past steps.
- The Immediate vs. Overall Performance Trade-off: While personalized feedback yielded the highest overall task mastery over time, it temporarily reduced the effectiveness of task-focused hints on the user’s very next response due to acute cognitive overload.
- Cognitive Ability Moderation: High cognitive ability learners derived less benefit from immediate task-focused hints, as they independently processed error corrections; extra feedback occasionally served as a distraction.
- Boredom and Affect Redirection: Participants reporting higher levels of situational boredom experienced greater performance gains from task-focused hints, which served to re-engage their attention.
Source: TUB
A learner stands in a room with a bottle, a cup, and a book. The task is to figure out, step by step, where each object belongs in the room. Under the table? Or perhaps on the chair? A humanoid robot gives the instructions – but in Swahili, a language the learner does not know. To solve the puzzle, the learner must gradually decipher what the robot is saying.
After each placement attempt, the robot responds. When the learner makes a mistake, it sometimes simply says that the placement was wrong. Sometimes it gives an additional hint: one that helps the learner think through the task, or one that asks them to reflect on their strategy.
This was the setup of the new study “Real-time cognitive-affective dynamics of failure feedback in a technology-based learning task” from the Cluster of Excellence Science of Intelligence (SCIoI) in Berlin, published in Communications Psychology. First author Helene Ackermann, together with Anna L. Lange, Hanna Dumont, Verena V. Hafner, and Rebecca Lazarides, investigated how 90 adult learners responded to automated feedback in a robot-supported learning task.
The study comes at a time when humanoid robots and AI tutors are increasingly discussed as future assistants in classrooms, workplaces, and everyday life. But while public debate often focuses on what robots may soon be able to do, the new findings point to a wider question: when does robotic support actually help a human learner?
The answer is actually more nuanced: Robot-delivered feedback helped learners recover from mistakes, but more personalized feedback was not automatically more helpful in the next moment.
“Our study shows that automated feedback can support learning after mistakes,” says Helene. “But it also shows that help has to fit the moment. Right after an error, more information is not always better.”
The timing of help
The researchers compared three feedback settings. In one, the robot gave additional feedback after every mistake without considering the learner’s current needs. In another, the amount of feedback was adapted to the learner’s most recent performance and self-reported enjoyment. In a third, the feedback was also personalized to the learner’s specific errors and previous steps in the task.
At first, this sounds like the smartest version: a robot that remembers what the learner has done before and responds more specifically, for example by reminding the learner that they have tried that exact position for the object already. But the findings, in fact, were more complex.
Personalized feedback made task-focused hints less effective for the learner’s very next response. One possible reason is cognitive load. The personalized messages contained more specific information and were therefore longer and harder to process. Directly after a mistake, this extra detail may have been too much to process before the learner’s next attempt.
At the same time, personalized feedback was linked to better overall performance across the whole task. This suggests a trade-off: detailed feedback can slow learners down in the moment, while still helping them build understanding over time.
“The personalized feedback was not seen as just good or bad,” says Anna Lange. “It rather seems to depend on the time scale. In the moment after a mistake, more specific feedback could make focusing on the next step harder. But across the task, it was still connected to better performance.”
Robots need to read the situation, not just the error
The study also found that the same feedback did not help everyone equally. Learners with higher cognitive ability benefited less from task-focused feedback, possibly because they were already able to work through the relevant steps on their own. For them, additional hints may have added little or even distracted them.
Another finding was more surprising: learners who reported feeling more bored benefited more from task-focused feedback. In this case, the robot’s hint may have helped redirect their attention and bring them back into the task.
Together, the results show how much effective robotic support depends on the amount and quality of information provided, and the learner’s cognitive and emotional state in the moment.
“Educational technologies are often discussed as if personalization were the final goal,” says Rebecca Lazarides. “Our findings show that the real challenge is more dynamic: systems need to combine personalization with a sensitive assessment of the learners’ situative cognitive and emotional states, and adjust the timing of personalized support.”
What robot tutors still need to learn
As humanoid robots become more visible in public debates about the future of education and assistance, the study reminds us: intelligent support is not the same as more support.
The findings do not suggest that robots should replace teachers, or that personalized feedback should be avoided. In fact, personalized feedback was the most effective strategy for overall performance, even though it may have increased situational cognitive load. The challenge is to balance the benefits of personalization with the risk of overwhelming learners and to design human-robot interactions with attention to learners’ situative cognitive and emotional states.
For SCIoI, the study contributes to a broader understanding of intelligent interaction. A robot may become a helpful partner in solving complex learning tasks, but the quality of its support depends on whether it can respond to the cognitive and affective dynamics of the human in front of it.
The most intelligent tutor may not be the one that always knows what to say, but rather the one that knows when to say less.
Key Questions Answered:
A: Personalized feedback contained specific details about the learner’s past attempts and error history. Right after making a mistake, processing this extra length and detail increased immediate cognitive load, making it harder for the learner to apply the hint on their very next attempt.
A: It helped over time. Despite creating a temporary processing bottleneck directly after an error, personalized feedback led to the highest overall task success across the entire experiment, helping learners build deeper structural understanding.
A: The study found that learners who felt bored benefited significantly more from task-focused hints. The robot’s automated feedback acted as an attentional anchor, redirecting bored participants back into active problem-solving.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by our staff.
About this robotics and neurotech research news
Author: Maria Ott
Source: TUB
Contact: Maria Ott – TUB
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Real-time cognitive-affective dynamics of failure feedback in a technology-based learning task” by Helene Ackermann, Anna L. Lange, Hanna Dumont, Verena V. Hafner & Rebecca Lazarides. Communications Psychology
DOI:10.1038/s44271-026-00487-8
Abstract
Real-time cognitive-affective dynamics of failure feedback in a technology-based learning task
As technology-based learning environments increasingly employ automated feedback, understanding how learners process feedback in real time is essential. This study examined how automated cognitive and metacognitive failure feedback delivered by a humanoid robot affected performance and how effects were moderated by feedback characteristics and learner characteristics.
Ninety adults (18-59 years, Mage = 29.53, 61 female, 27 male, 2 diverse) completed a learning task in three conditions: (1) fixed guidance condition with fixed-frequency and content-generic feedback, (2) basic-adaptive condition with frequency-adaptive but content-generic feedback, or (3) personalized-adaptive condition with frequency-adaptive and content-personalized feedback adjusting content to learners specific errors and prior steps.
A three-level generalized path model (trials nested within time blocks within learners) was estimated to investigate effects of failure feedback on immediate task performance and cross-level moderation effects. Results showed that cognitive and metacognitive failure feedback increased the likelihood of a correct subsequent response across conditions.
Relative to fixed guidance (condition 1), the implemented form of frequency-adaptive feedback (condition 2) did not show statistically significant moderation to these effects. Content-personalized feedback (condition 3) reduced effectiveness of cognitive failure feedback on immediate performance but improved overall performance as compared to content-generic feedback (condition 2).
Across conditions, learners with higher cognitive ability benefited less, while those reporting higher momentary on-task boredom benefited more from cognitive feedback.
These findings highlight that the effectiveness of automated failure feedback depends on both its design and learners’ situational cognitive and emotional states, illustrating how a situational, temporally sensitive approach can help open the “black box” of feedback effectiveness.
