Docs Library

What it knows, and from whom.

Two shelves. Guidance from research that a teacher can choose for an activity today, and the public resources it can reach. Each says where it comes from, its license, and what isn’t known.

Guidance from research

77 pieces on 8 topics, chosen in How the AI helps. Only what fits the moment reaches the model, with its source.

Scaffolding and fadingHow much help to give and when to withdraw it: help fitted to what the learner just did, the least that keeps them moving, and less of it as they improve.10 pieces
  • Find out what they can already do on this step before you help, then fit the help to that rather than to the next item in your plan. van de Pol et al. 2010; Wood et al. 1976
  • Give the least help that lets them take the next step themselves: a nudge to look again first, then a pointer to the part that matters, and a fuller hint only if that fails. Wood et al. 1976; Koedinger & Aleven 2007
  • Draw their attention to the feature of the problem that matters, and let them work out what to do about it. Wood et al. 1976
  • If they have lost the thread, remind them what they are trying to find and how this step serves it. Wood et al. 1976
  • If frustration is building, make the task smaller for now, such as handling one routine part yourself, and say you will hand it back. Wood et al. 1976
  • Let them work before you step in again: ask them to try the step alone and come back with what they found, instead of stacking more prompts. van de Pol et al. 2015
  • Once they have done a step with help, ask them to do the next similar one with less: drop the prompt you used last time, and add it back only if they stall. van de Pol et al. 2010; Renkl & Atkinson 2003
  • When they get one right unaided, hand over more of the work: let them plan the next step or choose the method rather than prompting each move. van de Pol et al. 2010; Kalyuga et al. 2003
  • If they keep asking for hints without trying, stop the hints leading to the answer: ask for one attempt, even a guess, before you give more. Shute 2008; Baker et al. 2004
  • Start from a case they can picture or handle, link it to the symbols step by step, then drop the picture once the symbols make sense. Fyfe et al. 2014
What the studies showed
The idea comes from a descriptive study of an adult tutoring children aged 3 to 5 on a block-building task (Wood et al., 1976), not a trial. A review of a decade of classroom research named fitting help to the learner, fading it and handing over responsibility as the core, and found few effectiveness studies (van de Pol et al., 2010). A meta-analysis of 144 experiments on computer-based scaffolding in STEM, primary to adult, found g = 0.46 on cognitive outcomes, with no detectable difference from fading or adding support (Belland et al., 2017). In one classroom experiment with 768 students aged 12-15, better fitted help won only when groups also had long stretches of independent work (van de Pol et al., 2015). Help that suits beginners can lose its value or backfire as they gain expertise (Kalyuga et al., 2003). How much to give and withhold is an open problem (Koedinger & Aleven, 2007). Gaming hints went with lower test scores in one classroom study (Baker et al., 2004), and teaching better help use changed the behavior but not learning (Aleven et al., 2016). Concreteness fading rests on a review of math and science work that argues mainly from theory (Fyfe et al., 2014).
What isn’t known
Almost all of this is about human tutors, teachers and rule-based tutoring software, not language models. No study has tested these moves given to a model one or two at a time. Whether fading adds anything over steady, well-fitted support is unsettled, and how to judge from a short chat what a learner can do alone is unknown.
Cite it as
VIABLE Lab (2026). Scaffolding and fading: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
FeedbackWhat to say about a learner's work: what is right or wrong and why, one next step, about the work and not the person, and criticism that shows the standard and your confidence they can meet it.10 pieces
  • Say what is right or wrong and why, then one thing to do next. A bare "correct" or "not quite" teaches little. Wisniewski et al. 2020; Van der Kleij et al. 2015
  • Comment on the work and the method, never the person: name the step or strategy, and leave out how clever or able they are. Kluger & DeNisi 1996; Hattie & Timperley 2007
  • When you criticize work, say why: you hold it to a high bar because you believe they can reach it. Then say how to get closer. Say this only when the next step really is within reach. Yeager et al. 2014; Cohen et al. 1999
  • Keep feedback to the one or two things that matter most, in small pieces. If there is more, save it until they have used this. Shute 2008
  • If they have not tried yet, ask for any attempt first, even a rough one, and respond to that: feedback works on something they have tried. Shute 2008
  • Give comments, not scores: don't rate the work out of ten or compare it with classmates' work. Koenka et al. 2021; Shute 2008
  • When they were sure of a wrong answer, take time on it: go through the reasoning that led there and correct it clearly. Confident errors are often the easiest to fix. Metcalfe 2017
  • Keep praise specific and honest, tied to what they did. Skip superlatives: overdone praise can make learners with low self-esteem avoid the next challenge. Brummelman et al. 2014; Hattie & Timperley 2007
  • Say where this work stands against the goal: what is done, what is still missing, and what would close the gap. Hattie & Timperley 2007; Black & Wiliam 1998
  • On a piece of writing, comment on one or two things that matter for its purpose, and ask them to revise before you comment again. Graham et al. 2015
What the studies showed
Feedback helps on average and varies a lot. A meta-analysis of 435 studies found d = 0.48 with very high heterogeneity and signs of publication bias: high-information feedback d = 0.99, corrective d = 0.46, rewards or punishment d = 0.24 (Wisniewski et al., 2020). An older meta-analysis found d = 0.41, with over a third of interventions lowering performance and weaker effects as attention moved to the self (Kluger & DeNisi, 1996). On computers, explanation beat right-or-wrong alone (0.49 against 0.05 in 40 studies; Van der Kleij et al., 2015; also Mertens et al., 2022). Hattie & Timperley (2007) and Shute (2008) are reviews: their advice on focus and amount is reasoned from the evidence, and findings on timing are mixed. Comments beat grades on achievement and motivation in elementary and secondary school (Koenka et al., 2021). Formative assessment is disputed: Black & Wiliam (1998) report substantial gains; a later meta-analysis found a mean of 0.20 in the 13 usable studies (Kingston & Nash, 2011) and a critical review calls the larger figures suspect (Bennett, 2011), while others fault that meta-analysis's methods (Briggs et al., 2012; McMillan et al., 2013). Saying the standard and your confidence comes from three double-blind US field experiments in grade 7 and a low-income high school, strongest for African American students who mistrusted school (Yeager et al., 2014; Cohen et al., 1999). Inflated praise lowered challenge seeking in children with low self-esteem and raised it in those with high self-esteem (Brummelman et al., 2014). Correcting confident errors rests mostly on laboratory work (Metcalfe, 2017). Feedback on writing, grades 1-8: adults 0.87, computers 0.38 (Graham et al., 2015).
What isn’t known
These are teachers, software and laboratory tasks, not language models. A statement of high standards and belief from an AI may not be read as it is from a teacher a learner knows; that has not been tested, and it could ring hollow. Whether feedback worded by a model changes what learners do next, or their later work, is not known.
Cite it as
VIABLE Lab (2026). Feedback: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Tutoring talkWhat skilled tutors and good discussions do: find out what the learner thinks, build on their words, ask for reasons, draw out missing ideas, and go from open questions to specific ones.10 pieces
  • Build your reply on something they actually said: pick up their words and take the idea one step further rather than moving on to your own point. Demszky et al. 2024; Howe et al. 2019
  • Before you explain, ask what they think is going on and why, and let the answer decide your next move rather than a script. Black & Wiliam 1998; Chi et al. 2001
  • If their answer is thin, ask them to expand before you judge it: what makes them say that, or an example of it. Michaels & O'Connor 2012; Graesser et al. 2005
  • Say their idea back in your own words and ask if that is what they meant, leaving room for them to correct you. Michaels et al. 2008; Michaels & O'Connor 2012
  • Ask for their grounds: what in the text or the numbers points to that? Then test a claim that looks right with a case that might break it. Michaels & O'Connor 2012; Howe et al. 2019
  • Know the key ideas a full answer needs and the misconceptions likely here. Draw out a missing idea one at a time, and correct a misconception when it appears. Graesser et al. 2005; Nye et al. 2014
  • Go from open to specific: first ask what else they can say, then give a hint, then a question with one word missing. State the idea yourself only if the teacher allows it. Graesser et al. 2005
  • Point out an error with a question that lets them see it rather than a correction, without implying the answer is right. Praise the solution, not the person. Lepper & Woolverton 2002
  • Watch how they feel as well as what they know: if they sound discouraged, say so briefly and give them something next that they can succeed at. Lepper & Woolverton 2002; D'Mello & Graesser 2012
  • Save the back-and-forth for what is hard for this learner. If it is within easy reach, let them get on with it rather than drawing it out. VanLehn et al. 2007
What the studies showed
Tutoring works: human tutoring d = 0.79 and step-based computer tutors d = 0.76 against no tutoring, well below the often-quoted 2 sigma (VanLehn, 2011); pre-K to 12 field experiments pool at 0.288 SD (Nickow et al., 2024). Which talk causes this is less clear. With tutors barred from explaining, students learned as well from prompts (Chi et al., 2001). Dialogue beat reading only when novices studied material written for intermediates, in qualitative physics (VanLehn et al., 2007). Expectation and misconception tailored dialogue comes from AutoTutor, modeled on human tutors, with gains near 0.8 sigma in its authors' own evaluations (Graesser et al., 2005; Graesser, Chipman et al., 2005; Nye et al., 2014). The expert-tutor account is observational, from experienced primary and secondary math tutors (Lepper & Woolverton, 2002). Talk moves come from whole-class teaching, not one-on-one, and taking them up at scale is hard (Michaels et al., 2008; Michaels & O'Connor, 2012; O'Connor & Michaels, 2019). In 72 classrooms, ages 10-11, elaborating and querying went with higher mastery where students participated widely, a correlation (Howe et al., 2019). A school trial of dialogic teaching, 76 schools, found about 2 months' gain (EEF, 2017). Automated feedback raised instructors' uptake of student ideas by 13% in a trial of 1,136 instructors, with only suggestive effects on students (Demszky et al., 2024). How large formative assessment's effect is remains disputed (Black & Wiliam, 1998; Kingston & Nash, 2011; Briggs et al., 2012).
What isn’t known
Talk moves were built for teachers with whole classes, and tutoring studies measure human tutors or scripted systems. Nobody has shown that a language model told to revoice, press for reasons or tailor to misconceptions teaches better for it. Whether learners take a model's questions as genuine interest or as withholding is open. Reading how someone feels from a few typed lines is unreliable.
Cite it as
VIABLE Lab (2026). Tutoring talk: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Explaining and doing the thinkingGetting the learner to do the mental work: asking why, in their own words; having them produce rather than only read or agree; and what to do when this feels harder than being told.9 pieces
  • After a correct step, ask why it follows from the one before, in their own words, before going on. Bisra et al. 2018; Chi et al. 1994
  • Ask questions they can't answer by copying: why is this true, what would change if the numbers changed, how does this fit what they already knew. Dunlosky et al. 2013; Chi & Wylie 2014
  • Have them produce something of their own, such as an example, a short summary, a sketch or a prediction. Asking them to read, agree or pick an option does less. Chi & Wylie 2014; Chi et al. 2018
  • Build on each other: add to what they produced and ask them to add to what you said, rather than you delivering and them confirming. Chi & Wylie 2014; Chi et al. 2001
  • If they say they get it, ask them to show it: explain it as if to a classmate who missed the lesson, including why as well as what. Kobayashi 2019; Roscoe & Chi 2007
  • With a worked example, ask for one explanation at the step that matters, not at every line: prompts added to examples don't always help. Barbieri et al. 2023
  • If they say being told would be easier, agree that working it out feels slower and harder, and say people often learn more this way even when it feels like less. Then offer the next small step. Deslauriers et al. 2019; Bjork et al. 2013
  • While they read, stop after a short section and ask what it means and how it connects to what they already knew. Chi et al. 1994
  • For a text, take turns with them: summarize a part, ask a question about it, clear up what is unclear and predict what comes next, then let them lead. Palincsar & Brown 1984; Rosenshine & Meister 1994
What the studies showed
Prompting learners to explain works on average: g = 0.55 across 64 reports and a range of tasks, subjects and levels (Bisra et al., 2018). The founding study was small: 14 prompted and 10 control eighth graders reading about the circulatory system (Chi et al., 1994). A major review rates self-explanation and elaborative questioning as moderate, since they were little tested in real classrooms (Dunlosky et al., 2013). ICAP, in which constructive and interactive work beats active and passive, is a framework with review-level support (Chi & Wylie, 2014); when teachers tried it they seldom managed interactive activities, though students learned more in constructive than active ones (Chi et al., 2018). Fiorella & Mayer (2016) review eight ways of having learners generate, with their limits. Students learned as well from tutors who only prompted as from tutors who explained (Chi et al., 2001). Contested: in math, adding prompts to explain to worked examples did worse (Barbieri et al., 2023). Preparing to teach and teaching: 28 studies, g = 0.35 and 0.56 (Kobayashi, 2019), though peer tutors tend to tell rather than build knowledge (Roscoe & Chi, 2007). Students in an active college class learned more and felt they learned less (Deslauriers et al., 2019), and people often misjudge how they learn (Bjork et al., 2013). Reciprocal teaching began with seventh graders who struggled with comprehension; median effects are 0.32 on standardized tests and 0.88 on researcher-made ones (Palincsar & Brown, 1984; Rosenshine & Meister, 1994).
What isn’t known
The prompts studied were written by instructors and researchers; Bisra et al. recommend testing generated ones, and whether a model's prompts work as well is not known. Whether learners explain to an AI with the effort they give a person, and how much prompting they tolerate before they leave, is untested. Explaining costs time and attention for beginners, and the balance is not settled.
Cite it as
VIABLE Lab (2026). Explaining and doing the thinking: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Worked examples and mental loadFor beginners, how to show a method without doing their work: a worked example of a similar problem, fewer steps shown as they improve, and replies that hold one idea at a time.10 pieces
  • Where the teacher allows explaining, show a similar problem worked step by step, then give them a matching one to do at once. Never work the problem they were set. Barbieri et al. 2023; Sweller & Cooper 1985
  • Keep any explanation short, tied to what they just said, and follow it with something they must do with it: explanations alone often don't help. Wittwer & Renkl 2008; Wittwer & Renkl 2010
  • Put one new idea in each reply. Don't stack a definition, a method and a question: give the next piece when they have used this one. Sweller et al. 2019; Rey et al. 2019
  • Quote the line of their work or the rule you mean, right where you discuss it, so they don't have to look it up or hold it in mind. Ginns 2006
  • As they improve, show less: leave out the last step of your example first, then earlier steps, until they are solving the whole problem. Renkl et al. 2002; Renkl & Atkinson 2003
  • Stop showing examples to someone already solving these alone, and give them problems instead: support that helps beginners can slow them down. Kalyuga et al. 2003
  • Show two ways of solving one problem and ask what is the same, what differs and which is quicker here. Name the principle after they have compared. Rittle-Johnson & Star 2007; Alfieri et al. 2013
  • Once they have seen it done correctly, show a solution with a common mistake in it and ask them to find and fix the error. Durkin & Rittle-Johnson 2012; Booth et al. 2013
  • Leave out what this step doesn't need: side facts, fun asides and long introductions, however interesting. Sundararajan & Adesope 2020
  • Write as if talking to them: "you", plain words, short sentences. Ginns et al. 2013
What the studies showed
For beginners, studying worked examples beats solving problems alone: a meta-analysis of 55 math studies, elementary to postsecondary, found g = 0.48 (Barbieri et al., 2023), after early algebra experiments and reviews (Sweller & Cooper, 1985; Atkinson et al., 2000). Contested: in that meta-analysis correct examples alone did better than incorrect ones and adding prompts to explain did worse, while separate experiments found incorrect examples help conceptual understanding (Booth et al., 2013, Algebra I; Durkin & Rittle-Johnson, 2012, grades 4-5). Fading helped in one field and two lab experiments, at least for near transfer (Renkl et al., 2002). The advantage falls or reverses as expertise grows (Kalyuga et al., 2003). Comparing cases: 57 experiments, d = 0.50 (Alfieri et al., 2013); grade 7 algebra (Rittle-Johnson & Star, 2007). Explanations alone often add little: 21 experiments (Wittwer & Renkl, 2008, 2010). Cognitive load theory is the usual account (Sweller et al., 2019), and its concepts and measures are questioned (de Jong, 2010). Segmenting (56 investigations; Rey et al., 2019), removing interesting but irrelevant detail (Sundararajan & Adesope, 2020), placing related information together (50 studies; Ginns, 2006) and conversational style (retention d = 0.30, transfer d = 0.54, small beyond 35 minutes; Ginns et al., 2013) come from multimedia lessons and written text, not chat.
What isn’t known
None of it tested a language model's replies. Whether a model's example is well chosen, correct and at the right level is not known, and a wrong example can teach the error. Where the line falls between example and problem for one learner has to be judged from a few messages. Reply length and one idea per reply have not been tested as such in tutoring chats.
Cite it as
VIABLE Lab (2026). Worked examples and mental load: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Productive struggle and AI helpWhen to let a learner struggle and when to step in, and what trials of AI tutors found: answers handed over during practice can cost later performance.10 pieces
  • Don't hand over finished solutions even when asked repeatedly: answers during practice can lift scores then and lower them later, when they work alone. Bastani et al. 2025
  • Check their work against the teacher's solution and the common mistakes if you have them, rather than from memory, and never call wrong work right. Bastani et al. 2025; Kestin et al. 2025
  • Guide one part at a time and wait for their reply. Don't lay out the whole route at once. Kestin et al. 2025; Jurenka et al. 2024
  • Ask them to explain, or ask a question that guides their thinking, more than you give general encouragement such as "good job". Wang et al. 2024
  • Before treating something as learned, have them do a similar one alone, without hints: getting through with help isn't evidence that they can do it. Bastani et al. 2025; Darvishi et al. 2024
  • Don't rescue at the first sign of difficulty: if they are still trying ideas, ask what they will try next and let them. Kapur 2016; D'Mello et al. 2014
  • Say that confusion at this point is normal, then give them a way to resolve it. Don't leave it unresolved: it tends to turn into frustration. D'Mello & Graesser 2012; D'Mello et al. 2014
  • If they have failed the same kind of step many times, more practice of it is unlikely to fix it. Go back to the idea or the earlier skill underneath. Beck & Gong 2013; Zhang et al. 2019
  • After a failed try, use it: show how their attempt falls short or set it beside the target method. Struggle pays only if the teaching afterwards draws on it. Loibl et al. 2017; Kapur 2016
  • With older learners and a new concept, where the teacher wants it, let them try a problem first, even if they can't solve it. Then teach, using what they tried. Sinha & Kapur 2021; Loibl et al. 2017
What the studies showed
Struggling first is contested. A meta-analysis of 53 studies found g = 0.36 for problem solving before instruction, higher with Productive Failure's design, and the reverse in grades 2-5 and for domain-general skills (Sinha & Kapur, 2021; one author originated the method). Year 5 experiments found instruction first better (Ashman et al., 2020). It helps only with contrasting cases or teaching built on students' solutions (Loibl et al., 2017). Unassisted discovery loses to explicit instruction (d = -0.38) and assisted discovery beats other methods (d = 0.30) across 164 studies (Alfieri et al., 2011; see also Kirschner et al., 2006; Kapur, 2008, 2016). Wheel-spinning: in ASSISTments and the Cognitive Tutor, students who don't master a skill quickly are likely never to (Beck & Gong, 2013, three right in a row within ten tries); a second definition agreed with it on under half of the cases either flagged (Zhang et al., 2019). Confusion can help if resolved and tends to frustration if not (D'Mello & Graesser, 2012; D'Mello et al., 2014). AI: in one Turkish high school, nearly 1,000 students, grades 9-11, math, a general chatbot raised practice scores 48% and lowered the later unassisted exam 17%; a tutor given teacher solutions and common mistakes and told to give hints raised practice 127% with no detectable difference on the exam, so it avoided harm and showed no gain; the general chatbot was right about 51% of the time (Bastani et al., 2025). 194 Harvard physics students learned more from a built tutor with pre-written solutions and one part at a time than from in-class active learning, 0.73 to 1.3 SD, one lesson per topic (Kestin et al., 2025). AI suggestions to 900 human tutors of 1,800 K-12 students raised exit-ticket passes from 62% to 66%, with more prompting to explain and less generic praise, and no significant effect on the end-of-year test, at low exposure; a preprint (Wang et al., 2024). The LearnLM papers set out five principles and report expert and learner ratings rather than measured learning (Jurenka et al., 2024; LearnLM Team, 2024), then an exploratory trial with 165 UK students, tutors approving every message, 66.2% against 60.7% on novel problems (LearnLM Team & Eedi, 2025): the makers' own preprints. Students relied on AI prompts rather than learning from them (Darvishi et al., 2024, 1,625 students); ChatGPT lifted essay scores and not knowledge or transfer (Fan et al., 2025, 117 university students).
What isn’t known
No trial has separated which parts of a guarded design matter, and none tests short guidance like these lines. Most AI trials are short, at one site and often run by the makers; the best-known guarded tutor prevented harm without a gain on later exams. Effects over months, in other subjects and for younger learners, are unknown. When struggle tips into wheel-spinning can't yet be read reliably, from logs or from chat.
Cite it as
VIABLE Lab (2026). Productive struggle and AI help: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Recall, spacing and mixingMaking learning last: asking learners to bring ideas back from memory, coming back to them after a gap, and mixing kinds of problems once each is learned.8 pieces
  • Once they have worked through something, ask them to bring it back from memory without looking, rather than showing it again. Roediger & Karpicke 2006; Yang et al. 2021
  • Have them write or say the answer themselves rather than pick from options: recalling helps more than recognizing. Rowland 2014
  • After they try to recall, say what was right and what was missing: practice with correction works better than practice alone. Yang et al. 2021; Butler & Roediger 2008
  • When they are stuck, ask what they remember about it before you re-explain, and say a wrong guess is fine: even a failed attempt helps what they learn next to stick. Kornell et al. 2009; Pan & Carpenter 2023
  • Come back to an idea after other work, later in the conversation or in a later session, and ask for it again rather than repeating it straight away. Cepeda et al. 2006; Carpenter et al. 2022
  • Don't stop at the first correct recall: ask again a few times, and return to it on a later day. Rawson & Dunlosky 2011
  • Once they can do each kind of problem, mix in an earlier kind so they must decide which method applies, rather than repeating the last one. Rohrer et al. 2020; Rohrer & Taylor 2007
  • If two kinds of problem get confused, mix them on purpose and ask which is which before solving. Don't do this with word lists or a text: there it hasn't been shown to help. Brunmair & Richter 2019
What the studies showed
Among the best replicated findings in the science of learning. Repeated recall beat restudy on delayed tests, even though restudy raised confidence (Roediger & Karpicke, 2006, students reading prose). Recall gives more than recognition (Rowland, 2014); classroom quizzing g = 0.499 across 222 studies and 48,478 students, moderated by corrective feedback among other things (Yang et al., 2021). A systematic review of 50 classroom experiments found mostly medium or large benefits, only 6% in non-WEIRD countries (Agarwal et al., 2021). Trying before studying helps even when the attempt fails, in lab experiments with the answer shown after; size varies by procedure (Kornell et al., 2009; Richland et al., 2009; Pan & Carpenter, 2023). Spacing: 317 experiments (Cepeda et al., 2006); the best gap grows with how long the learning must last (Cepeda et al., 2008). A major review rates practice testing and spacing high, interleaving moderate (Dunlosky et al., 2013). Relearning: three correct recalls, then three spaced sessions (Rawson & Dunlosky, 2011). Interleaving in math: college students (Rohrer & Taylor, 2007); grade 7, d = 0.42 and 0.79 (Rohrer et al., 2015); 54 grade 7 classes, 61% against 38%, d = 0.83 (Rohrer et al., 2020). Across materials g = 0.42, math 0.34, nothing clear for expository text, and blocking is better for words (Brunmair & Richter, 2019).
What isn’t known
Few of these studies involve a tutor in conversation, and none an AI. Spacing needs a learner to come back: within one chat the gaps are short, and the Harness keeps no memory across sessions unless a host passes history. How many recall questions a learner will accept from a tutor before it feels like a quiz is not known. Interleaving's classroom evidence is almost all math.
Cite it as
VIABLE Lab (2026). Recall, spacing and mixing: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.
Metacognition and motivationHelping learners steer their own learning, planning, checking and judging what they know, inside the work at hand; and what holds up on motivation, growth mindset included.10 pieces
  • Before they start, ask what the task wants, what they already know that could help, and how they will begin. EEF 2025; Dignath & Büttner 2008
  • Partway through, ask whether this way is getting them closer, and what else they could try. EEF 2025
  • When they finish, ask what worked, what they would change and what they would try first next time. EEF 2025; Hattie et al. 1996
  • Where the teacher allows it, think aloud on a similar problem: what you notice, what you ask yourself, why you choose a step. Then have them do the same on theirs. EEF 2025; EEF Toolkit 2025
  • When you suggest a way to check or study, tie it to this problem and have them use it now: general advice on how to learn rarely carries over. Hattie et al. 1996; EEF Toolkit 2025
  • Ask how sure they are before you respond, then compare that with how it turned out: many learners misjudge what they know. Dunlosky & Rawson 2012; Bjork et al. 2013
  • If they plan to reread or look over notes, suggest testing themselves instead: rereading feels effective but does less. Dunlosky et al. 2013; Roediger & Karpicke 2006
  • Treat a mistake as information about what to try next, in passing. Don't lecture about mindset or brains, and don't call anyone a natural or not a math person. Macnamara & Burgoyne 2023; Burnette et al. 2023
  • Offer a small real choice that doesn't change the goal, such as which problem first or which way to try. Patall et al. 2008
  • If they ask why this matters, ask them to connect it to something in their own life or plans, rather than telling them why it is useful. Hulleman & Harackiewicz 2009; Lazowski & Hulleman 2016
What the studies showed
Metacognition has a strong average and is hard to realize. The EEF Toolkit puts it at +8 months across 355 studies, with a padlock lost because many were not independently evaluated, and warns the impact is hard to achieve in practice; its guidance recommends explicit teaching of planning, monitoring and evaluating within subject lessons, and modeling (EEF Toolkit, 2025; EEF, 2025). Two meta-analyses of 84 school studies found a mean of 0.69, larger when researchers rather than teachers delivered and in math (Dignath & Büttner, 2008). Study skills transfer best when trained in context (Hattie et al., 1996). College students who judged their recall more accurately retained more (Dunlosky & Rawson, 2012), and people often misjudge how they learn (Bjork et al., 2013). Rereading rates low and practice testing high (Dunlosky et al., 2013). Growth mindset, honestly: a meta-analysis of 63 studies found d = 0.05, not significant after correcting for bias, and 0.02 in the best studies (Macnamara & Burgoyne, 2023); another found d = 0.14 for targeted groups with high fidelity, with a prediction interval from -0.08 to 0.35 (Burnette et al., 2023). A national trial with 12,490 US ninth graders raised lower achievers' grades by 0.10 points, more where peer norms fitted, and not higher achievers' (Yeager et al., 2019). Earlier overall effects were weak (Sisk et al., 2018), and replication attempts of the praise-for-intelligence study with 624 children aged 9-13 in China were largely null (Li & Bates, 2019; Mueller & Dweck, 1998). So these lines don't teach mindset; they only avoid trait labels. Choice: 41 studies, raising motivation and effort, strongest for instructionally irrelevant choices (Patall et al., 2008). Relevance: one experiment in first-year high school science, largest for those low in confidence (Hulleman & Harackiewicz, 2009); motivation interventions overall d = 0.49, smaller in randomized designs (Lazowski & Hulleman, 2016).
What isn’t known
Most of this is teachers in schools over weeks, not a model in a single chat. Whether a line or two of prompting about planning or confidence does anything is untested, and these prompts could become one more chore. Growth mindset messages from an AI have not been tested at scale here. AI help may also reduce a learner's own planning and checking (Fan et al., 2025; Darvishi et al., 2024): whether prompts like these offset that is not known.
Cite it as
VIABLE Lab (2026). Metacognition and motivation: guidance from research. The wording is VIABLE Lab's; each piece names the studies it draws on.

Is your research missing?

One file of plain sentences, or just the sentences and the papers and we’ll write it with you. Teachers can then choose it, your citation travels with your words, and you can see that it’s used.

Add your research

Public resources

54 datasets, benchmarks, standards and tools, each checked against its own source. Listing one isn’t endorsing it. A few are connected, and the rest say what they could become. How one gets connected

54 of 54

  • Apprentice Learner Architecture

    Simulated learners that learn from worked examples and correctness feedback inside tutoring systems, used to model students, compare tutors and author tutors.

    Tools. MIT. Open to download. Could serve as a simulated learner in tests.
  • AI2 Reasoning Challenge (ARC)

    7,787 multiple-choice science questions from standardized tests for grades 3 to 9, split into a Challenge Set and an Easy Set.

    Benchmarks. CC-BY-SA-4.0. Open to download.
  • ASAP 2.0

    About 24,000 source-based persuasive essays by students in grades 6, 8, 9 and 10, with human scores, the rubric and the source texts.

    Datasets. CC-BY-4.0. Open to download. Could serve as rating replies.
  • Bridge

    700 real tutoring conversations with 1st to 5th graders that end in a math mistake, each annotated by expert teachers with the error, a remediation strategy and an expert's reply.

    Datasets. MIT, CC-BY-NC-4.0. Open to download. Could serve as rating replies.
  • Carnegie Skills Progressions

    Progressions for three durable skills (collaboration, communication and critical thinking), each with subskills, indicators and four levels from Exploring to Extending.

    Standards. CC-BY-4.0. Open to download. Could serve as what an AI is told.
  • CER science writing feedback dataset (Playlab)

    A planned de-identified dataset of 2,000 to 3,000 multi-turn AI feedback conversations on students' claim, evidence and reasoning writing in science, with an annotation rubric.

    Datasets. CC-BY-4.0, Apache-2.0. Not released yet. Could serve as rating replies.
  • CIMA

    Tutoring conversations role-played by crowdworkers, with several tutor replies to each input and labels for tutor actions.

    Datasets. CC-BY-2.5. Open to download. Could serve as rating replies.
  • CLEAR corpus (CommonLit Ease of Readability)

    About 5,000 reading passages for grades 3 to 12, with readability scores from teachers' judgments and their genre and publication details.

    Datasets. MIT, CC-BY-NC-SA-4.0. Open to download.
  • ConvoLearn

    2,134 conversations between 323 credentialed teachers and a simulated seventh grader on Earth science, labeled for how they build knowledge.

    Datasets. MIT. Open to download. Could serve as rating replies.
  • CSTA PK–12 Computer Science Standards (2026)

    What students should know and be able to do in computer science from pre-K through high school, with high school specialty standards in areas such as AI, cybersecurity and data science.

    Standards. CC-BY-NC-SA-4.0. Open to download. Could serve as what an AI is told.
  • DataShop and LearnSphere

    A repository of students' step-by-step interaction data from tutors and online courses for learning scientists, with LearnSphere's Tigris tool for building analysis workflows.

    Learning data repositories. Needs an account. Could serve as a measure in the records.
  • DrawEduMath

    2,030 images of students' handwritten answers to 188 math problems, each described by a teacher, with teacher-written and synthetic questions and answers about the images.

    Benchmarks. CC-BY-NC-4.0, CC-BY-4.0, Apache-2.0. Open to download.
  • Eedi Misconceptions Graph

    Common math misconceptions, each linked to the skills (Eedi constructs) it hinders, drawn from students' answers to Eedi's diagnostic questions.

    Knowledge graphs. CC-BY-4.0. Open to download. Could serve as what an AI is told.
  • EEF Teaching and Learning Toolkit

    What evidence says about teaching approaches, with cost and strength of evidence, from the Education Endowment Foundation.

    Open content. Open to download. Could serve as what an AI is told.
  • ELLIPSE corpus

    About 6,500 essays by English language learners in grades 8 to 12, scored for overall proficiency and for cohesion, syntax, vocabulary, phraseology, grammar and conventions.

    Datasets. CC-BY-4.0. Open to download. Could serve as rating replies.
  • FairytaleQA

    10,580 questions written by education experts about 278 children's stories, covering seven narrative elements, for reading comprehension from kindergarten to eighth grade.

    Datasets. Apache-2.0. Open to download.
  • FoundationalASSIST

    1.7 million interactions of 5,000 grade 6 to 8 students with 3,395 math problems in ASSISTments, with the problems' text, students' first answers, hint requests and skills.

    Datasets. CC-BY-NC-4.0. On request. Could serve as a measure in the records, a simulated learner in tests.
  • GSM8K

    8,500 grade school math word problems written by people, each with a step-by-step solution.

    Benchmarks. MIT. Open to download.
  • IM K–12 Math (Illustrative Mathematics)

    A full K-12 math curriculum (K-5, 6-8, and high school Algebra 1, Geometry and Algebra 2), free to view online.

    Open content. CC-BY-4.0, CC-BY-NC-4.0. Open to download. Could serve as what an AI is told.
  • K-12 AI Infrastructure Platform

    The K-12 AI Infrastructure Program's hub, run by DrivenData, for the datasets, benchmarks and challenges it funds or lists, each shown with its license.

    Learning data repositories. Needs an account.
  • KB-TutorBench

    A planned dataset of about 6,000 annotated tutoring dialogues from simulated and real classrooms in science, ELA and history, with fine-tuned tutoring models.

    Benchmarks. CC-BY-4.0, Apache-2.0. Not released yet. Could serve as rating replies.
  • Learning Commons Evaluators

    Model-based evaluators, with their prompts, rubrics and validation data, that rate educational text for complexity, the quality of feedback on student writing, math standards alignment and critical thinking.

    Evaluators. MIT, CC-BY-4.0, CC-BY-NC-SA-4.0. Open to download. Could serve as rating replies.
  • Learning Commons Knowledge Graph

    U.S. academic standards, learning components, learning progressions and misconception data from several publishers, linked in one graph with a shared schema.

    Knowledge graphs. MIT, CC-BY-4.0, CC0-1.0. Open to download. Could serve as what an AI is told.
  • MathCAMPS

    9,707 synthetic math word problems, each tied to one of 44 Common Core standards for grades K to 8, with follow-up questions.

    Benchmarks. MIT. Open to download.
  • MathDial

    About 2,900 one-to-one tutoring dialogues in which teachers tutor a simulated student through a mistaken solution to a math word problem, annotated with the teachers' moves.

    Datasets. CC-BY-SA-4.0, CC-BY-4.0. Open to download. Could serve as a simulated learner in tests.
  • MathFish

    About 21,800 K-12 math problems from Illustrative Mathematics and Fishtank Learning, labeled with the Common Core standards they address, for testing whether models can tell which standards a problem fits.

    Benchmarks. ODC-By-1.0, CC-BY-4.0, CC-BY-NC-SA-4.0, Apache-2.0. Open to download.
  • MathTutorBench

    A benchmark that scores AI tutors on seven math tutoring tasks across math expertise, understanding the student and teaching, with a trained reward model for open-ended scaffolding.

    Benchmarks. CC-BY-4.0, CC-BY-SA-4.0. Open to download. Could serve as rating replies.
  • MinorBench

    299 prompts in six risk categories, with four system prompts, for checking a teaching assistant meant for children.

    Benchmarks. MIT. Open to download. Could serve as rating replies.
  • MRBench and the BEA 2025 shared task data

    Replies from seven AI and human tutors to students' math mistakes, each annotated on eight teaching dimensions, such as finding the mistake, revealing the answer and giving actionable guidance.

    Benchmarks. CC-BY-SA-4.0. Open to download. Could serve as rating replies.
  • NCTE classroom transcripts

    1,660 anonymized transcripts of U.S. 4th and 5th grade math lessons, with turn-level discourse annotations on a subset.

    Datasets. MIT. On request. Could serve as marks on what learners did.
  • Next Generation Science Standards

    K-12 science standards, adopted or adapted by many U.S. states.

    Standards. LicenseRef-NGSS-public-license. Open to download. Could serve as what an AI is told.
  • OATutor

    An open adaptive tutoring system with a library of problems, hints and scaffolds, from college algebra, statistics and calculus texts.

    Tools. MIT, CC-BY-4.0. Open to download. Could serve as rating replies.
  • OpenSciEd

    A free science curriculum designed for the NGSS, for elementary, middle and high school (biology, chemistry and physics).

    Open content. CC-BY-NC-4.0. Open to download. Could serve as what an AI is told.
  • OpenStax

    Free textbooks online and as PDFs, with titles for high school and AP courses in math, science and social studies on its K12 pages.

    Open content. CC-BY-NC-SA-4.0, CC-BY-4.0. Open to download. Could serve as what an AI is told.
  • Open University Learning Analytics Dataset

    More than 30,000 students’ records across seven modules, with what they did in the learning platform.

    Datasets. CC-BY-4.0. Open to download. Could serve as a measure in the records.
  • Pedagogy Benchmark

    1,143 multiple-choice questions from teacher training exams, on general pedagogy and special educational needs, with public leaderboards.

    Benchmarks. MIT. Needs an account. Could serve as rating replies.
  • PERSUADE 2.0

    Over 25,000 argumentative essays by U.S. students in grades 6 to 12, with holistic scores and annotated argument elements rated for effectiveness.

    Datasets. CC-BY-4.0. Open to download.
  • Productive Coaching benchmark (Quill.org)

    A planned benchmark, with educator-written feedback and annotated AI coaching sessions from several learning tools, for judging whether an AI's coaching on student writing is productive.

    Benchmarks. CC-BY-4.0. Not released yet. Could serve as rating replies.
  • pyBKT

    A Python library of Bayesian Knowledge Tracing models that estimate a student's mastery of each skill from their sequence of right and wrong answers.

    Tools. MIT. Open to download. Could serve as a measure in the records.
  • Sandpiper

    A web app that uses language models to annotate tutoring transcripts with a research team's codes, masking personal information first, and compares the annotations with people's coding.

    Tools. MIT. Open to download. Could serve as marks on what learners did.
  • Science misconceptions benchmark (Learning Equality)

    A planned benchmark, dataset of assessment packages and small on-device classifiers for spotting science misconceptions in students' free-form answers.

    Benchmarks. Not released yet. Could serve as what an AI is told, marks on what learners did.
  • Simulated student models for fractions (Princeton)

    A planned pipeline and neural simulated student models for fraction arithmetic, trained on rule-based students' synthetic reasoning together with real students' reasoning.

    Models. Apache-2.0. Not released yet. Could serve as a simulated learner in tests.
  • Storiza corpus of early reader story recordings

    A planned corpus of about 80 hours of K-3 children reading stories aloud, across U.S. dialects, with word-aligned reading-error labels and baseline speech recognition models.

    Datasets. Not released yet.
  • StudyChat

    Real students’ dialogues with ChatGPT in an AI course, with dialogue-act labels and the scores of those who consented.

    Datasets. CC-BY-4.0. Needs an account. Could serve as marks on what learners did.
  • TalkMoves

    567 transcripts of K-12 math lessons, annotated sentence by sentence for ten teacher and student talk moves.

    Datasets. CC-BY-NC-SA-4.0. Open to download. Could serve as marks on what learners did.
  • Tutor Move Taxonomy

    A coding scheme for tutors’ moves, from the National Tutoring Observatory: tutoring, learning, social-emotional and logistical support.

    Open content. CC-BY-4.0. Open to download. Could serve as rating replies, marks on what learners did.
  • TutorBench

    About 1,500 expert-written tutoring tasks in high school and AP STEM subjects, each with its own rubric, testing explanations, feedback on student work and hints.

    Benchmarks. Open to download. Could serve as rating replies.
  • TutorGym

    A testbed that puts AI agents inside existing tutoring systems (CTAT, Apprentice Tutors, OATutor), as tutors compared with the system's own feedback or as students compared with human learning data, across 223 tutor domains.

    Benchmarks. MIT. Open to download. Could serve as a simulated learner in tests.
  • TutorRL-7B

    Open 7-billion-parameter math tutoring models, trained with reinforcement learning against a simulated student to guide students rather than give answers away.

    Models. Apache-2.0, CC-BY-4.0. Open to download.
  • Universal Design for Learning Guidelines

    CAST’s guidelines for designing learning for everyone: ways of engaging, of representing, and of acting and expressing.

    Open content. Open to download. Could serve as what an AI is told.
  • AI competency framework for students (UNESCO)

    What students should be able to do with AI, as a framework a school can set goals from.

    Standards. CC-BY-SA-3.0-IGO. Open to download. Could serve as what an AI is told.
  • AI competency framework for teachers (UNESCO)

    What teachers should be able to do with AI, for professional learning and for reading what a tool asks of them.

    Standards. CC-BY-SA-3.0-IGO. Open to download. Could serve as what an AI is told.
  • VETTING

    A method, with a Python package, in which a second model checks each AI reply against a policy the learner never sees, before the learner sees the reply.

    Tools. MIT. Open to download. Could serve as a step in a workflow.
  • What Works Clearinghouse practice guides

    Recommendations from the U.S. Institute of Education Sciences for challenges in classrooms and schools, on topics such as foundational reading, writing, fractions, algebra and behavior.

    Open content. LicenseRef-Public-Domain. Open to download. Could serve as what an AI is told.