Marks-weighted analysis of 436 questions across 4,675 marks from 17 IB Math AI HL exam sittings — topic weightage, P1/P2/P3 splits, the Paper 3 modelling breakdown, and Photon Academy's data-driven predictions for May 2026. Applications & Interpretation is IB's stats-heavy, GDC-on-every-paper route — and the data proves it.
Data-Driven May 2026 exam forecast — built from 4,675 real marks See PredictionsPercentage of total exam marks allocated to each topic across all 17 sittings, with marks distributed among tags to avoid double-counting. Statistics and probability sit right at the top — the AI HL signature.
Both papers are worth 110 marks and both allow a GDC, but topics still split. Eigenvalues/Markov and Graph Theory are Paper 2 engines; Functions and Complex Numbers are front-loaded onto Paper 1.
More than a quarter of every AI HL exam is data and inference — distributions, hypothesis tests, regression and bivariate analysis. No other IB Math route is this statistics-heavy, and it is the single biggest reason AI students who neglect stats underperform.
Number of questions tagged to each topic per session, 2021 to 2025. Calculus (integration and differentiation) is clearly climbing.
Average marks per question — higher means the topic almost always appears as a big extended-response question rather than a short one.
Five years of data, six findings worth memorising before May 2026.
Binomial, Poisson and normal distributions are the single largest topic by marks, and they turn up on both Paper 1 and Paper 2. Master the GDC distribution menus first.
Probability distributions, hypothesis testing, regression and bivariate stats together account for more than a quarter of the exam. AI HL rewards statistical fluency over algebraic tricks.
The biggest single contributor to Paper 3. χ² tests, t-tests and goodness-of-fit are how the modelling paper asks you to justify whether a model actually fits the data.
Eigenvalues/Markov are 7.7% of Paper 2 vs just 2.7% of Paper 1, and Graph Theory is 7.2% of Paper 2. Transition matrices and networks are calculator-paper staples.
Functions and modelling/regression are front-loaded onto Paper 1's shorter items, where you set up and interpret a model by hand before the calculator does the heavy lifting on Paper 2.
Integration questions rose from 2 (2021) to 8 (2025) and differentiation from 4 to 8. IB is testing more optimisation, area/volume and rates of change embedded in real-world contexts.
Photon Academy's data-driven projection, built only from the marks weighting, paper splits and year trends above. This is a forecast of the exam's shape — not leaked content — and it maps exactly to how our tutors prioritise revision.
HL only. Roughly 55 marks in one hour of extended, real-world problem-solving. Unlike AA's pure-math investigations, AI HL Paper 3 is applied modelling — a single scenario that layers several topics on top of each other.
The largest single block on Paper 3. Expect a χ² or t-test near the end to decide whether the model you built actually explains the data. Interpreting the p-value in context earns the final marks.
Differential Equations (100) plus Coupled DEs (60.5) form the dynamic heart of Paper 3 — logistic growth, predator-prey systems and Euler's method for models with no neat closed form.
Adjacency and transition matrices, shortest paths and spanning trees get embedded inside the scenario. Eigenvalues/Markov (56.5) often ride alongside for long-run steady states.
The recurring shape: fit or interpret a regression, extend it into a differential-equation or Markov model, then run a hypothesis test to validate it. Recognising the arc is half the battle.
How well candidates actually performed on each major AI HL topic, session by session, straight from the examiners' own commentary. Green means the cohort handled it well; red means it was a recurring weak spot.
| Topic | M22 T1 | M22 T2 | N22 | M23 T1 | M23 T2 | N23 | M24 T1 | M24 T2 | N24 | M25 T1 | M25 T2 | M25 T3 | N25 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Eigenvalues & Markov Chains | Mixed | Mixed | Good | Poor | Mixed | Good | Poor | Mixed | Mixed | Good | Poor | Poor | Poor |
| Hypothesis Testing | Poor | Poor | Mixed | Poor | Poor | -- | Mixed | Poor | Mixed | Poor | Mixed | Poor | Poor |
| Complex Numbers | Poor | Poor | Poor | -- | Mixed | Poor | Mixed | Poor | Good | Poor | Poor | Poor | Poor |
| Probability Distributions | Poor | -- | Poor | Poor | Good | Poor | Poor | Mixed | Mixed | Poor | Mixed | Poor | Good |
| Vectors | Mixed | Poor | Poor | Mixed | -- | Poor | -- | Mixed | Poor | Poor | Poor | Poor | Poor |
| Graph Theory | -- | Mixed | Poor | Poor | Mixed | Mixed | Poor | Mixed | Mixed | -- | Good | Poor | Poor |
| Differentiation | Poor | Poor | Poor | -- | Poor | Poor | -- | Poor | Poor | -- | Poor | Mixed | Poor |
| Loans & Annuities | -- | Poor | Poor | Mixed | Good | Poor | Mixed | Good | Mixed | -- | Mixed | Mixed | -- |
| Trigonometry | -- | Mixed | Mixed | Poor | Poor | Good | -- | Good | Mixed | Good | -- | Mixed | Good |
| Integration | -- | -- | Poor | Mixed | Poor | Poor | Mixed | Mixed | Mixed | Poor | Poor | -- | Poor |
| Differential Equations | Poor | -- | -- | -- | Poor | Poor | Poor | Mixed | -- | Poor | Poor | Poor | Mixed |
| Modelling & Regression | Poor | Poor | Good | Poor | -- | Good | Poor | Poor | -- | -- | Mixed | Poor | -- |
| Sequences & Series | Mixed | Poor | -- | -- | -- | Poor | Good | Poor | Mixed | Poor | -- | -- | Mixed |
| Coupled Differential Equations | Poor | -- | Poor | -- | Good | -- | -- | Poor | Poor | -- | -- | Poor | -- |
| Voronoi Diagrams | Mixed | -- | -- | -- | -- | -- | Poor | -- | Good | Mixed | Poor | -- | -- |
Synthesised from 13 official IB examiner subject reports spanning M22 to N25 (May 2022 through November 2025, including the three May 2025 time zones). Ratings reflect the examiners' written commentary on each session, not raw score data.
The ten mistakes examiners flag most often across the 2022–2025 AI HL reports. Fixing these is the fastest way to stop leaking marks you have already earned the method for.
The single most-cited fault in every session: rounding intermediate values instead of storing the full GDC display, then giving finals to 2sf, 4dp or calculator precision rather than the required three significant figures. Carry unrounded values and only round at the very end.
Hypotheses written as inequalities or in words, null and alternative reversed, the population parameter or distribution never stated, and conclusions like "reject H₀" left without context. A non-significant result never "proves" the null — always compare the p-value (not r) to the significance level.
Indefinite integrals repeatedly lose the constant of integration, the dx is dropped from the integral, and applied answers arrive without units. These are free marks examiners cannot award once they are gone.
The GDC is left in degrees when radians are required (or vice-versa after a reset), quietly corrupting cosine-rule and trigonometric values. Check the mode before every trig or geometry question.
Missing steps and unreadable layout mean method and follow-through marks cannot be credited when the final answer is wrong. Show every calculator entry and the order of your reasoning — especially in graph-theory algorithms where edge-selection order carries marks.
Reading [AB] or a⃗ as a magnitude, writing row vectors with commas like coordinates, and confusing position with direction vectors (or magnitude with a scalar absolute value). Keep vectors, points and lengths clearly distinct.
Treating without-replacement problems as with-replacement, confusing the pdf with the cdf, modelling a continuous variable as discrete, and adding standard deviations instead of variances when combining random variables. State the distribution and its parameters every time.
Candidates substitute or start from the given answer instead of deriving towards it, which earns no marks. Build the result independently and only quote the given value to finish.
Answers left in GDC form such as 2.51E-7, or GDC syntax copied verbatim, instead of correct mathematical notation. Translate every calculator output into proper form before writing it down.
Over-working "state" and "write down" items, missing "hence", and ignoring the answer form the question demands (exact/rational values, a specific line equation, or logged data). Read what form is required before you start.
Official IB grade boundaries (overall, % of total marks). Source: official IB subject reports, M22–N25 session.
| Session | G1 | G2 | G3 | G4 | G5 | G6 | G7 |
|---|---|---|---|---|---|---|---|
| N25 | 0 | 14 | 29 | 40 | 52 | 64 | 76 |
| M25 TZ3 | 0 | 13 | 26 | 38 | 52 | 65 | 77 |
| M25 TZ2 | 0 | 12 | 24 | 33 | 45 | 56 | 66 |
| M25 TZ1 | 0 | 11 | 23 | 32 | 44 | 55 | 66 |
| N24 | 0 | 15 | 28 | 39 | 52 | 65 | 77 |
| M24 TZ2 | 0 | 13 | 25 | 35 | 48 | 62 | 74 |
| M24 TZ1 | 0 | 13 | 26 | 37 | 51 | 63 | 74 |
| N23 | 0 | 12 | 24 | 35 | 50 | 62 | 77 |
| M23 TZ2 | 0 | 12 | 23 | 32 | 44 | 57 | 68 |
| M23 TZ1 | 0 | 11 | 24 | 34 | 46 | 57 | 69 |
| N22 | 0 | 8 | 16 | 24 | 39 | 52 | 64 |
| M22 TZ2 | 0 | 10 | 18 | 26 | 37 | 49 | 62 |
| M22 TZ1 | 0 | 8 | 16 | 26 | 37 | 50 | 63 |
Source: official IB subject reports (grade-boundaries page), M22 TZ1–N25. Each value is the minimum mark for that grade, out of 100.
AI HL Grade 7 boundary ranged from 62% (May 2022 TZ2) to 77% (Nov 2023, Nov 2024 & May 2025 TZ3). Average across all sessions: 70%.
AI HL Grade 5 boundary ranged from 37% to 52%, averaging 46% of total marks — a mark you reach by banking method and accuracy marks across every paper.
Early post-pandemic sessions sat in the low 60s; recent sessions have pushed the Grade 7 boundary into the mid-to-high 70s. Target the 70% average as a working baseline for a 7.
We build full predicted paper sets straight from the analysis on this page. The AA HL and AA SL sets are already live; the AI HL P1 / P2 / P3 sets are being written now to mirror the real topic weighting, paper splits and Paper 3 modelling structure shown above.