Systematically improving test performance: what works

Summary

Students in Avanti’s Centres of Excellence already show strong academic gains, about 16 to 21 JEE Main percentile points students with the same Class-10 marks who received after-school instruction.(Avanti Fellows 2026). CoEs are residential, 2-year programmes where Avanti’s teachers deliver ~6 hours of tutoring in Physics, Chemistry and Math each day.

Avanti’s prior research on augmenting or replacing in-class instruction with digital content has shown small or negative effects. Classroom videos with workbooks in Haryana’s government schools lowered maths scores by about 0.15 SD (de Barros 2022), and the blended programme in Chennai did not move Class X board scores (Chiplunkar, Dhar and Nagesh 2020). This is corroborated by similar research elsewhere. Across 318 comparisons, distance teaching matched the classroom (Bernard et al. 2004). Exam-preparation videos in Tanzania moved the national exam by no more than 0.13 SD (Seo 2017), and the same lessons on students’ own tablets lowered scores in Pakistan by 0.42 SD (Beg et al. 2022). Given the demonstrated efficacy of the in-class instruction, the biggest area of impact is likely to lie in the time outside of class.

Students in our centres typically spend 3-4 hours each day studying by themselves, typically with Avanti’s modules (workbooks), NCERT books and reference books. Typical study patterns include solving problems in an unsupervised manner, reading solutions and discussions with peers.

Avanti runs a mentorship programme in which each Grade 11 and 12 student meets a teacher-mentor one-to-one three times a year, six times over the two years (Student Mentorship SOP, 2026-27). Each 35-40 minute conversation is holistic: academic progress is checked inside it, from the test report or the LMS. The centres do not hold a separate academic mentorship session after each test, which they found was not operationally feasible. We have also piloted AI-generated academic mentorship reports (an anonymised sample). They are used only in the online programme, not in the centres, and focus on which chapters a student should work on, without remedying conceptual mistakes, conceptual gaps or a lack of knowledge of specific problem-solving steps. What the centres do have after every major test is a subject-level doubt-solving session the next day, and the test report, which recommends the chapters each student should prioritise. That window is crowded: teachers also use it to finish the syllabus, run chapter tests and hold mentorship sessions.

This review focuses on low-friction, AI-assisted interventions that fit within the flow of existing classrooms and allow improvements over an already high base, and above all on the time students spend studying on their own. Human tutoring in similar large-scale contexts has been found to be effective, raising scores on standardised tests by about 0.16 to 0.21 SD in large programmes, when it is delivered in person, at school and during school hours, with one or two students per tutor and a set curriculum for the tutors (Kraft, Schueler and Falken 2024; see Appendix B). That is a third to a half of the 0.42 SD found across all tutoring trials, and the drop is driven mainly by scale: effects decline as programmes grow, a pattern common across education research. The authors point to smaller studies serving narrower groups, tutor ratios and hours loosening as programmes grow, later students benefiting less, and the quality of delivery falling.

The closest pooled evidence is on test preparation in general (Hao et al. 2026, 28 studies, about 16,700 students). Test preparation raised scores by 0.26 SD on average. Contact hours and duration made no difference. The gains came from teaching specific test-taking skills, such as guessing, eliminating wrong options and managing time (0.35 SD, against 0.05 without them), and from working through workbooks (0.40 against 0.15). Timed practice tests and sample items added little on their own. On college admission tests, the closest thing to JEE, the effect was smaller, 0.14 SD, and not statistically significant.

This is the case for AI: to standardise what makes tutoring and test preparation work, the same steps in every session for every student, at a scale where human tutoring loses its effect, with the teacher checking its work. Avanti’s Centres of Excellence provide a unique context where we can use AI to augment teacher-led tutoring as well as run guided self-study sessions for students. The evidence we produce here will likely be the largest-scale and most focused evidence on our specific use case.

The presence of a teacher-in-the-loop will also allow us to focus on designing learning and practice sessions that combine best practices from the literature, as studies have shown that unguided self-study dramatically underperforms focused tutoring. Working ideas out alone does worse than being taught, by about 0.38 SD (Alfieri et al. 2011), and quizzing yourself on multi-step problems adds only about 0.10 SD (Ruitenburg et al. 2026). An ideal tutor would:

• Start with a conversation about why the student answered a question wrong. It uncovers the misconception behind the answer, and naming and refuting a misconception beats a standard explanation (Schroeder and Kucera 2022).
• Teach to that misconception, then generate similar problems to test whether it has cleared. A retry on a new question carries about half of the classic tutoring effect (von Hippel 2024).
• For students struggling with the concepts, use solved examples, then faded steps, then practice tests again. Solved examples help students weak in a chapter by about 0.20 SD (Barbieri et al. 2023).

The evidence also points to the fact that having these AI-assisted self-study sessions supervised by teachers would likely increase adherence as well as outcomes.

• A teacher in the room keeps students practising. In Uttar Pradesh’s residential government schools, an adult who protected weekly practice time and supervised use raised Khan Academy practice from 7 to 47 minutes a week, and maths scores by 0.33 to 0.44 SD (Oreopoulos, Keyes-Krysakowski and Agarwal 2026). Tutoring held during the school day does better than tutoring held after it (Nickow, Oreopoulos and Quan 2020); at scale, tutoring outside the school day is associated with effects about 0.19 SD smaller (Kraft, Schueler and Falken 2024).
• A teacher can step in where the software gets it wrong. In a UK trial, expert tutors read every message an AI tutor drafted and could change any of them. They sent 76% unchanged or nearly so, and the supervised AI did at least as well as a human tutor (LearnLM Team, Google and Eedi 2025). Without that supervision, a chatbot that handed students the answers made them do worse on an exam taken without it (Bastani et al. 2025).

a) What leads to conceptual learning

• Learn the basics in a lesson first. A clear explanation of the chapter’s ideas comes first for everyone. Leaving students to work the ideas out for themselves, with no explanation afterwards, does worse than teaching them directly, by about 0.38 SD.

  Evidence: high. A large meta-analysis, with a larger gap for adolescents (Alfieri et al. 2011). Section 3.1.

• Name the wrong idea behind an answer and show why it fails. That works better than a standard explanation, by about 0.28 SD after a correction for publication bias. But check first, with a quick “how sure were you?”: one wrong answer from one student is often a guess, not a wrong idea. And see the working: from the option alone you can only guess where a student went wrong.

  Evidence: moderate for refutation, mostly from university reading studies (Schroeder and Kucera 2022); several studies show one answer is a noisy signal (Lasry et al. 2011; Kaltakci Gurel, Eryilmaz and McDermott 2015). Peer Instruction found the same: nearly half of correct answers came only after discussion (Crouch and Mazur 2001). Section 3.8.

b) How students develop problem-solving skills

• Weak in the chapter: study solved examples before practice. For a new kind of problem, studying full solutions teaches as much as solving problems, in about half the time (g = 0.48).

  Evidence: moderate. 55 studies in maths agree, with smaller gains in real classrooms (Barbieri et al. 2023; McLaren et al. 2016). Section 3.2.

• Then complete solutions with steps missing, one more step removed each time. A natural bridge from examples to practice that costs nothing to build, but its own effect is not established, so treat it as worth trying rather than proven.

  Evidence: low. Small trials only, all pointing the same way (Renkl, Atkinson and Maier 2000; Salden et al. 2010). Section 3.3.

• Strong in the chapter: go straight to problems, with the solution there if stuck. Solved examples make strong students learn less, by about 0.43 SD, and giving one to a strong student costs less than withholding it from a weak one, so when unsure, show it.

  Evidence: moderate. One meta-analysis (Tetzlaff et al. 2025). Section 3.7.

c) What helps students retain what they learn

• Practise problems spread over weeks, with question types mixed. Quizzing yourself works for formulas and facts, but not for multi-step problems you cannot yet do (about 0.10 SD, with a range that crosses zero), and JEE is mostly those.

  Evidence: moderate for spreading practice out, which adds about 0.2 SD in classrooms. Mixing question types helps on topic tests but has not yet shown up on real exams (Ruitenburg et al. 2026; van der Haar et al. 2023). Section 3.4.

• Judge learning days later, unaided, on new questions. Practice scores with a tutor can rise sharply while exam scores do not move.

  Evidence: moderate. Two large trials, and a consistent gap between researcher-written and independent tests (Bastani et al. 2025; LearnLM Team, Google and Eedi 2026). Section 3.10.

d) What helps students learn from the questions they get wrong

• See the correct answer and the working first, then work out where your own working went wrong. Being told only “wrong” does almost nothing. Most of the gain, about 0.25 SD, comes from seeing the right answer. This needs the student’s own working, so the rough sheet has to be kept and brought to the review.

  Evidence: moderate. Two meta-analyses of feedback types agree (Mertens, Finn and Lindner 2022; Van der Kleij et al. 2015). Keeping the rough sheet is our inference. Section 3.5.

• Then try a new question of the same kind, without help. In the classic tutoring studies, about half the gain came from this retry.

  Evidence: low to moderate. A re-analysis of those studies, not a trial of the retry alone (von Hippel 2024). Section 3.6.

• Hold back the answer until the student has tried. A chatbot that handed out answers made students do worse on an exam taken without it.

  Evidence: low. Seven school trials of structured tutors, with a wide spread, and one large trial showing the harm (Bastani et al. 2025). Section 3.9.

Who gives the help. A teacher can run every step above. An AI tutor helps most with d), naming the wrong idea and setting the retry, and does harm where it hands out answers.

How big. On the coached topics, each of these moves a student in the middle of the class up by roughly 4 to 15 percentiles. On the whole JEE paper, where those topics are a small share of the marks, the combined effect is about 2 to 6 percentiles, and could be zero.

Evidence: low. These are planning judgements, and no study here was run with Indian students preparing for JEE or NEET. Section 3.11.

AI problems to solve

There are three, listed here in the order we would build them.

1. Why the student went wrong, and the nudge. This is the core problem. From the wrong option, pooled across every school that sat the paper, and from the student’s own working, the AI proposes the likely wrong idea and a question that tests it. The teacher confirms it before anyone names it to the student. It comes first because it is the step only AI can do at scale: on AIAT-04, 16 of 75 questions at EMRS Bhopal had 10 or more wrong answers with half or more on one option, too many for a teacher to diagnose by hand each week. It is also the riskiest, because one wrong answer is often a guess. So the AI’s output is a candidate, and we measure how often teachers agree with it before it goes further.
2. New questions for the retry, and solutions. The worked solution and its steps-missing version can be adapted from the solutions the question bank already holds, so that part is light editing. The real gap is new questions of the same kind for the unaided retry and the weekly check. These have to be generated, and each needs its answer and arithmetic checked before a student sees it. This comes second because the retry carries about half the classic tutoring effect and needs fresh questions to work.
3. Which problems to practise. Teacher and student agree three to five questions from the ranked report. This is mostly data, not AI: rank by how many students missed a question, the chapter’s priority (High, Medium, Low) and whether the student is weak or strong in that chapter. Build it as rules first; AI adds little beyond drafting the shortlist.

Instructional design flows to test

There are four, cheapest first, each adding one thing to the flow before it.

• A. Reports only. The “why” goes into the test report students already read after each major test, with no change to any session. This tests whether better information alone changes what students practise.
• B. Teacher-approved diagnosis. The AI drafts the likely wrong idea behind each missed question, and the teacher approves, edits or rejects it before it reaches the student’s test report. This tests how often the AI’s diagnosis matches the teacher’s, which is the gate for everything after it.
• C. Co-teaching session. This runs inside the doubt-solving session teachers already hold the day after each major test. The teacher takes a small group through its most-missed questions, reads each student’s working with Claude, then runs the solved examples and the unaided retry. Comparing groups of 3 and 6 shows how much teacher time each student needs, which matters because that window is already crowded.
• D. Supervised practice on the tablets. Students practise on their own devices with the teacher in the room and the AI holding back answers. This runs only once B shows the AI’s diagnoses can be trusted.

All four are judged on the weekly unaided check on new questions, not on practice scores, which can rise with a tutor while exam scores do not move. A and B need no new product, only a change to the test report: the “why” added to it, and for B a step where the teacher approves it first. C is the pilot at EMRS Bhopal, and D waits on B’s agreement rate. One open question: should B and C run in different classes, so that their effects can be told apart?

The evidence in full and further reading

• Techniques that raise test performance after a test, the systematic review and meta-analysis behind this note: methods, the estimate behind each technique, the tables of studies and the references. Section and appendix numbers in this note refer to it.
• The AI coach, step by step, on three questions: the coach’s steps worked through on three AIAT-04 questions from EMRS Bhopal.
• Dataset: Motion in Two Dimension (zip, 3 MB): 63 questions from Avanti’s question bank, with text, options, key, worked solution and figures, and 34,234 answers from 1,914 CoE students in 2026-27 under random study IDs. Its README explains every file.
• Dataset: EMRS Bhopal Grade 12 (zip, 83 MB): every test answer from one CoE class of 70 students, August 2025 to October 2026, under random study IDs, with each student’s test scores and the full text, key, solution and figures of 1,945 questions. Its README explains every file.
• The literature reviews, one per strand: AI tutors, learning science, adaptive platforms, human and AI together, the India context, step-based tutoring, lesson design, AI authoring, and video, live teaching and self-study in richer and in poorer countries.
• The reading library: 26 papers read but not cited, each in plain language with a download link.
• Sources: every source opened, with its archived copy.