Why guided AI tutoring helped maths stick

Introduction

The most important result from the Eedi and LearnLM maths tutoring trial was not that an AI could explain mathematics. Static hints already do that. The more interesting finding was that students appeared to learn more when the system behaved like a tutor: asking questions, diagnosing misunderstandings, and guiding reasoning step by step instead of simply revealing information. In the study, supervised LearnLM tutoring outperformed the platform’s standard pre-written hints on measures of learning transfer, meaning students were better able to solve a different problem afterwards rather than merely complete the immediate task.[Yahoo Finance]finance.yahoo.comYahoo FinanceNew Exploratory Research From Eedi and Google…11 Nov 2025 — The latest trial also measured "knowledge transfer": how tuto…

That distinction matters far beyond one maths platform. If AI is to contribute to the larger vision of educational abundance often discussed within the AI bloom debate, its value will come less from delivering answers cheaply and more from helping millions of people develop durable understanding. The Eedi trial suggests that the mechanism matters. Guided tutoring appears to work differently from static help because it engages the learner’s thinking process rather than treating learning as information delivery.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

What static hints could and could not do

Static hints have been used in educational software for decades. They can remind students of a formula, point out a common mistake, or suggest the next step in a problem. They are often useful because they are consistent, cheap to deliver, and easy to quality-control.

But static hints have an important limitation: they do not know what a particular student is thinking.

A pre-written hint might tell a pupil to reconsider the denominator in a fraction problem. Yet the system cannot easily tell whether the student:

The same hint is shown regardless of which of those situations is true.

This creates a common educational problem. Students can often follow a hint closely enough to finish the current question without actually changing the misconception that caused the error. They learn how to complete that specific task but not necessarily how to solve a related problem later. Educational researchers sometimes describe this as the difference between task completion and genuine knowledge transfer.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

The Eedi study was designed specifically to test this harder question. Rather than only measuring whether students fixed the original problem, researchers looked at whether tutoring improved performance on a subsequent question. That makes the comparison more meaningful because it focuses on understanding rather than short-term answer production.[Yahoo Finance]finance.yahoo.comYahoo FinanceNew Exploratory Research From Eedi and Google…11 Nov 2025 — The latest trial also measured "knowledge transfer": how tuto…

How LearnLM used questions and scaffolding

The supervised LearnLM system approached the problem differently. Instead of delivering a fixed sequence of hints, it generated conversational tutoring responses grounded in learning-science principles and then had those responses reviewed by expert tutors before reaching students.[eedi-labs-67qwnm2lxu4blqi5.webflow.io]eedi-labs-67qwnm2lxu4blqi5.webflow.ioNew Exploratory Research from Eedi and Google…The trial took place in the summer of 2025 across five UK secondary school classrooms, u…

A central idea was scaffolding.

In educational psychology, scaffolding means providing enough support to help a learner progress while still requiring them to do the cognitive work themselves. The support is gradually adjusted according to the learner’s needs rather than being delivered in a standard format.

LearnLM was explicitly trained around principles such as guided questioning, active participation, and encouraging students to explain their reasoning. Google Cloud[2blog.google]blog.googlegoogle learnlm gemini generative aiHow Google's LearnLM generative AI models support…14 May 2024 — Our technical report presents our approach to improving generative AI…Published: May 2024

Instead of saying:

Use this formula.

the tutor might ask:

Those questions force the student to retrieve and organise knowledge rather than merely receive it.

This resembles the Socratic method long used by skilled human tutors. Rather than transmitting answers directly, the tutor guides the learner towards discovering the reasoning themselves. Research on tutoring has repeatedly found that this active engagement often produces stronger learning than passive explanation alone.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…[Tech & Learning]techlearning.comIntegrated with LearnLM—an AI model built specifically for teaching based on learning science—this tool uses the Socratic method and pers…

One striking detail from the Eedi trial was that supervising tutors reported that LearnLM was particularly strong at generating Socratic-style questions. Some tutors even said they picked up useful pedagogical techniques from reviewing the model’s drafts.[arXiv]arxiv.orgSource details in endnotes.

Why transfer to new problems is the harder test

The strongest evidence in favour of guided tutoring was not immediate correctness. It was transfer.

Many educational interventions can improve performance on the exact problem students are currently working on. The challenge is helping them solve a different problem later.

A student who receives the correct method may be able to mimic it once. A student who understands why the method works is more likely to recognise when it should be applied elsewhere.

The Eedi results suggested that supervised LearnLM tutoring improved this transfer measure more than static hints. Reported outcomes showed larger gains on subsequent questions after AI-supported tutoring than after the platform’s standard hint system. Human tutoring also outperformed static hints, but the supervised LearnLM system appeared particularly effective on transfer outcomes.[Yahoo Finance]finance.yahoo.comYahoo FinanceNew Exploratory Research From Eedi and Google…11 Nov 2025 — The latest trial also measured "knowledge transfer": how tuto… 2arXiv

This finding is important because transfer is one of the most difficult goals in education.

Schools are not primarily trying to teach students how to answer a single question. They are trying to help learners acquire mental models that can be applied in unfamiliar situations. Mathematics is especially demanding in this regard because understanding often depends on recognising deeper structures beneath surface differences.

A student who truly understands fractions, algebra, or geometry should be able to apply those ideas when the numbers change or when the problem is presented in a new form.

The transfer result therefore offers evidence that guided AI tutoring may be affecting conceptual understanding rather than merely improving short-term performance. The trial was exploratory and relatively small, so it does not settle the question. But it points towards a mechanism that educational researchers care about far more than answer accuracy alone.[socialscienceregistry.org]socialscienceregistry.orgTesting the efficacy of AI tutoring in secondary mathematics13 Apr 2026 — An exploratory randomised controlled trial (RCT) conducted in 2…

Why asking questions can produce deeper learning

Several learning-science ideas help explain why guided tutoring may outperform static hints.

Retrieval practice. Students remember information better when they actively retrieve it rather than passively reread it. Guided questions force retrieval. Static hints often bypass it.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

Error diagnosis. A conversation can uncover the specific misconception causing a mistake. A fixed hint generally cannot adapt to the student’s response.[eedi.com]eedi.comGrounded in Eedi's Diagnostic Engine, it…Read more…

Cognitive engagement. Students learn more when they explain reasoning in their own words. Interactive tutoring creates opportunities for explanation and reflection.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

Adaptive support. A tutor can give more help when a learner is stuck and less when they are progressing. Static hints provide the same support regardless of context.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

Productive struggle. Educational research often finds that some degree of effort improves retention. Giving answers too quickly can reduce learning even if it improves short-term success rates. Guided tutoring can maintain challenge while preventing frustration.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

These mechanisms help explain why many researchers increasingly distinguish between AI systems that answer questions and AI systems designed to teach.

The role of supervision was part of the result

An easy mistake is to interpret the Eedi findings as proof that any advanced chatbot will automatically become an excellent tutor.

That is not what the study showed.

The LearnLM system was constrained, pedagogically tuned, and supervised by expert human tutors. Researchers reported that tutors reviewed model-generated messages and ensured they met educational standards before they reached students. Around three quarters of LearnLM’s drafted messages required only minimal or no edits, but the human oversight remained an important part of the system design.[arXiv]arxiv.orgSource details in endnotes.

This matters because educational AI faces several well-known risks:

The Eedi trial’s success therefore came from a specific architecture: a learning-focused model operating within pedagogical constraints and human supervision. The result was not simply “AI tutoring works”. It was that a carefully designed tutoring system worked better than static hints on a demanding learning measure.[Brookings]brookings.eduwhat the research shows about generative ai in tutoringWhat the research shows about generative AI in tutoring27 Jan 2026 — The study rigorously compared static hints, human tutoring…

Why this matters for the larger AI bloom argument

The broader significance is not really about one maths lesson.

Education has long suffered from a scarcity problem. Personal tutoring is one of the most effective educational interventions ever studied, but providing a skilled tutor for every learner is prohibitively expensive in most systems.

If AI can reliably reproduce some of the mechanisms that make tutoring effective — guided questioning, misconception diagnosis, personalised scaffolding, and patient feedback — then high-quality cognitive support could become far more widely available. That possibility fits directly into the AI bloom idea of making valuable forms of intelligence abundant rather than scarce.

The Eedi and LearnLM trial does not prove that outcome is coming. The study was small, exploratory, and conducted under unusually careful conditions. Large-scale deployment would raise questions about reliability, incentives, access, teacher roles, and long-term effects on learning habits.[socialscienceregistry.org]socialscienceregistry.orgTesting the efficacy of AI tutoring in secondary mathematics13 Apr 2026 — An exploratory randomised controlled trial (RCT) conducted in 2…

But it does offer an important clue about where the value may lie. The strongest early evidence did not come from AI producing answers faster than humans. It came from AI helping students think through problems more effectively than a static support system could. In the context of long-term human flourishing, that distinction may be far more important than raw automation.

Amazon book picks

Further Reading

Books and field guides related to Why guided AI tutoring helped maths stick. Use these as the next step if you want deeper reading beyond the article.

Book

Powerful Teaching

By Pooja K. Agarwal, Patrice M. Bain

Shows how guided questioning and retrieval strengthen learning.

See on Amazon

Book

How Learning Happens

By Paul A. Kirschner, Carl Hendrick

Explains why scaffolding and feedback outperform passive instruction.

See on Amazon

Book

Why Don't Students Like School?

By Daniel T. Willingham

Explains how thinking drives learning and retention.

See on Amazon

Book

Make It Stick

By Peter C. Brown, Henry L. Roediger III et al.

Emphasises retrieval, effortful learning and transfer rather than answer delivery.

See on Amazon

Browse more on Amazon:Powerful TeachingHow Learning HappensWhy Don't Students Like School?

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

Know & Learn Your Alphabet Educational Poster Large / Wall Chart - ABC's & Maths

SearcheBay.co.uk: math learning poster

Browse similar oneBay.co.uk

Example eBay listing

12x12 Multiplication Game Table Kids Math Learning Posters, Fidget Popper

SearcheBay.co.uk: math learning poster

Browse similar oneBay.co.uk

Example eBay listing

Division Educational Math Poster Teaching Classroom Learning Posters For Kids

SearcheBay.co.uk: math learning poster

Browse similar oneBay.co.uk

Example eBay listing

Subtraction Educational Math Poster Teaching Classroom Learning Posters For Kids

SearcheBay.co.uk: math learning poster

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: finance.yahoo.com
Link:https://finance.yahoo.com/news/exploratory-research-eedi-google-deepmind-090000225.html

2. Source: arxiv.org
Link:https://arxiv.org/abs/2512.23633

3. Source: brookings.edu
Title: what the research shows about generative ai in tutoring
Link:https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/

4. Source: cloud.google.com
Link:https://cloud.google.com/solutions/learnlm

5. Source: arxiv.org
Link:https://arxiv.org/abs/2002.12552

6. Source: socialscienceregistry.org
Link:https://www.socialscienceregistry.org/trials/18079

7. Source: eedi-labs-67qwnm2lxu4blqi5.webflow.io
Link:https://eedi-labs-67qwnm2lxu4blqi5.webflow.io/news/new-exploratory-research-from-eedi-and-google-deepmind-reveals-human-in-the-loop-ai-tutoring-outperforms-human-only-support

8. Source: blog.google
Title: google learnlm gemini generative ai
Link:https://blog.google/products-and-platforms/products/education/google-learnlm-gemini-generative-ai/

Published: May 2024

9. Source: arxiv.org
Link:https://arxiv.org/abs/2305.13272

10. Source: eedi.com
Link:https://www.eedi.com/news/eedi-labs-selected-for-accelerates-national-program-to-scale-evidence-based-ai-math-tutoring-in-us-middle-schools

11. Source: arxiv.org
Link:https://arxiv.org/html/2512.23633v1

12. Source: eedi.com
Link:https://www.eedi.com/news/just-launched—our-second-ai-tutor-rct

13. Source: eedi.com
Link:https://www.eedi.com/news/new-uk-study-finds-students-using-eedi-gain-2-4-months-of-additional-maths-progress

14. Source: techlearning.com
Link:https://www.techlearning.com/how-to/geminis-guided-learning-mode-from-google-ai-what-educators-need-to-know

15. Source: techlearning.com
Title: googles new ai tutor learnlm is trained on learning science and it shows
Link:https://www.techlearning.com/how-to/googles-new-ai-tutor-learnlm-is-trained-on-learning-science-and-it-shows

16. Source: api.emergentmind.com
Link:https://api.emergentmind.com/topics/learnlm

Additional References

17. Source: linkedin.com
Link:https://www.linkedin.com/posts/simon-woodhead_edtech-aied-edm-activity-7393960569541652481-kR0c

18. Source: finanznachrichten.de
Link:https://www.finanznachrichten.de/nachrichten-2025-11/66943468-new-exploratory-research-from-eedi-and-google-deepmind-reveals-human-in-the-loop-ai-tutoring-outperforms-human-only-support-004.htm

19. Source: facebook.com
Link:https://www.facebook.com/marius.comper/photos/students-tutored-by-googles-learnlm-a-generative-ai-model-fine-tuned-for-pedagog/10164198279814621/

20. Source: pressreleasehub.pa.media
Link:https://pressreleasehub.pa.media/article/new-exploratory-research-from-eedi-and-google-deepmind-reveals-human-in-the-loop-ai-tutoring-outperforms-human-only-support-59458.html

21. Source: linkedin.com
Link:https://www.linkedin.com/posts/benkornell_breakthrough-aiedu-aiforeducation-activity-7394047991826870272-mP3s

22. Source: youtube.com
Link:https://www.youtube.com/watch?v=o712grs7ONo

23. Source: annmichaelsen.com
Link:https://annmichaelsen.com/2024/12/11/revolutionizing-education-googles-learnlm-and-the-ai-powered-future-of-learning/

24. Source: atalupadhyay.wordpress.com
Title: googles learn lm the future of ai powered education
Link:https://atalupadhyay.wordpress.com/2025/08/10/googles-learn-lm-the-future-of-ai-powered-education/

25. Source: researchgate.net
Title: 387351061 LearnLM Improving Gemini for Learning
Link:https://www.researchgate.net/publication/387351061_LearnLM_Improving_Gemini_for_Learning

26. Source: meratutor.ai
Title: learnlm and gemini ai prompting guide
Link:https://www.meratutor.ai/blog/learnlm-and-gemini-ai-prompting-guide/