Singapore Math: CPA Progression, Bar Modelling, and the MOE Framework

Findings

1. The CPA progression and its theoretical base

CPA (Concrete–Pictorial–Abstract) traces to psychologist Jerome Bruner’s theory of three modes of representing knowledge: enactive (action/objects), iconic (images), and symbolic (language/notation) [1]. Singapore’s Ministry of Education (MOE) adopted this as an explicit pedagogical sequence: students first manipulate physical objects (counters, base-ten blocks), then move to pictorial representations (drawings, number lines, bar models), and only then work with numbers and symbols alone [1][2]. CPA is advocated by MOE, embedded directly in Singapore textbooks, and taught in pre-service teacher training [1]. It is described in the mastery literature as more than a linear three-step recipe — a “sophisticated, interlaced learning system” connected to Piaget, Vygotsky, and Dienes as well as Bruner, aiming at relational understanding (why, not just how) [2].

2. Bar modelling (model drawing)

The bar model method (also “model method” or “model drawing”) was developed by Dr. Kho Tek Hong as part of Singapore’s Primary Mathematics Project and formally entered the curriculum in 1983 [3]. It represents quantities as rectangular bars whose relative lengths encode part-whole or comparison relationships, with unknowns marked by a question mark — effectively pre-algebraic reasoning made visual [3]. Three canonical bar types are used in classroom practice:

• Part-whole model: parts combine into a total; used for addition/subtraction problems where either a part or the whole is unknown [4].
• Comparison model: two (or more) bars of different lengths represent a “how many more/fewer” or scaling relationship [4].
• Equal-groups model: a bar divided into equal segments, used for multiplication, division, and equal-sharing problems [4].

A widely-taught step sequence for constructing a bar model: (1) read the whole problem once for gist; (2) circle/underline key information; (3) identify the variables (who/what is involved); (4) choose the appropriate bar type; (5) draw bars with rough proportionality; (6) label known quantities and mark the unknown with ”?”; (7) re-read to check the model matches the problem; (8) perform and check the calculation [3][4]. In UK KS1/KS2 adaptations, the progression itself follows CPA: concrete objects → representational counters → pictorial drawings → labeled abstract bars by Year 2, extending to multi-step, fraction, ratio, proportion and pre-algebra problems by Year 6 [4].

3. Number bonds

Number bonds express the part-part-whole relationship (e.g., 7 = 3 + 4) and are foundational from Singapore’s Primary 1 (and even pre-school/kindergarten) onward, taught via the same CPA sequence — physical objects, then bond diagrams, then abstract equations [5]. They are the numeric precursor that makes later bar-model work with addition/subtraction fluent: a student who has internalized bonds to 10 and 20 can read a bar model’s missing segment almost automatically [5]. The term itself dates to the 1920s in English-language pedagogy but was formalized into Singapore’s national curriculum in the early 1970s [5].

4. The MOE Mathematics Framework (“Pentagon”) and spiral syllabus

Singapore’s Mathematics Curriculum Framework, introduced in the 1990s, places mathematical problem solving at its center, supported by five interrelated components arranged in a pentagon diagram: Concepts (numerical, algebraic, geometric, statistical, probabilistic, analytical — interdependent, not siloed), Skills (procedures and their conditions of use, e.g., numerical calculation, algebraic manipulation, spatial visualization, use of tools), Processes (reasoning, communication, connections, application/modelling, and heuristics such as drawing a diagram or working backwards), Attitudes (motivation, confidence, appreciation of mathematics), and Metacognition (monitoring and regulating one’s own problem-solving, “getting unstuck,” reviewing solutions) [7]. This framework is stated explicitly across MOE syllabus documents and is reproduced in NIE teacher-training materials [7][8].

Around this framework, the syllabus itself is a spiral curriculum spanning roughly 12 years (primary through pre-university): each content strand (Numbers & Algebra; Geometry & Measurement; Statistics & Probability, etc.) is revisited at every grade level with increasing depth rather than taught once and left behind [6]. This is the structural link to CPA: a topic met concretely in an early grade is often re-met pictorially in a later grade and abstractly still later, rather than each topic getting a single CPA pass [6][7].

5. NIE’s research role

The National Institute of Education is Singapore’s sole teacher-education institution and, funded via MOE’s Office of Education Research, runs the country’s mathematics-education research program, including studies of CPA implementation, the Pentagon framework’s use by “competent teachers,” and classroom enactment more broadly [7][8]. NIE materials are also where much of the Pentagon framework’s canonical description (concepts/skills/processes/attitudes/metacognition, with problem solving at the center) is documented and disseminated to teacher trainees [7].

6. Evidence on adoption outcomes: US and UK

US (What Works Clearinghouse): WWC reviewed “Singapore Math” curricula (Primary Mathematics for grades 1–6, and by extension Math in Focus / New Elementary Mathematics) in 2009 and updated the review in December 2015. Its conclusion: no studies met WWC group-design evidence standards, so WWC assigned no effectiveness rating and explicitly stated it could draw no research-based conclusions, positive or negative, about Singapore Math’s impact on K–8 achievement [9]. The absence of a rating is not evidence of ineffectiveness, but it means the strongest independent US evidence body found nothing rigorous enough to certify.

Publisher evidence base: HMH’s evidence document for Math in Focus (the US edition of Singapore’s Marshall Cavendish curriculum) leans on TIMSS ranking data (Singapore #1, US #12–15 depending on grade/year across 1995–2015) and smaller studies — Ginsberg et al. (2005) on textbook design; Hiebert & Grouws (2007) on conceptual-first instruction; Leong, Ho & Cheng (2015) and Salingay & Tan (2018) on positive CPA effects on performance/retention — but reports no US-specific effect sizes or RCT results [12].

UK (EEF): The Education Endowment Foundation funded an independent RCT of “Mathematics Mastery” — a UK program built on Singapore’s mastery approach (systematic mathematical language, frequent concrete/pictorial representation, high expectations), not a direct import — running 2011–2014 across 83 primary schools (4,176 pupils) and 44 secondary schools (5,938 pupils), with a larger later primary evaluation covering roughly 90 schools and 5,108 pupils [10]. Result: primary pupils made on average about two months’ additional progress, statistically significant, security rating 3/5 on EEF’s padlock scale; the secondary trial and a combined meta-analysis found a smaller effect (roughly one month) [10]. This is the most credible causal estimate available for a Singapore-derived approach — moderate, not transformative.

Commentary (Fordham Institute): A National Math Panel-adjacent commentary treats Singapore Math as “worth examining” but warns against directly mapping cross-country curriculum data onto US practice without controlled study, echoing WWC’s caution [13].

7. Touchscreen manipulation of bar models

Direct empirical work on touch-based bar-model interaction is thin, but a directly relevant strand exists in embodied-cognition research on touchscreen math manipulatives. Duijzer et al. (2017, Frontiers in Psychology) studied 9–11-year-olds using a tablet app (“Mathematical Imagery Trainer for Proportion”) where students dragged two vertical bars to maintain a target ratio (e.g., 1:2) [11]. Key findings: students spontaneously formed “attentional anchors” — consistent gaze/action patterns on mathematically relevant points (bar tops, midpoints) — and nearly all progressed from incorrect/additive strategies to correct multiplicative reasoning purely through guided touch manipulation, with symbolic scaffolds (blank screen → grid → numbered grid) introduced progressively as students’ informal strategies matured [11]. Existing digital bar-model tools in the wild (Visnos’s bar-model tool, MathsBot’s bar modelling manipulative) let users create bars on a number line, resize, cut, join, and clone them in “manual” or “automatic” modes, confirming this is an established manipulative pattern in UK maths-mastery classrooms even without large-scale efficacy studies specific to the digital version [14].

Design implications for Math Challenge

1. Build a first-class Bar Model widget, not a generic drag-shape tool: bars snap to a shared vertical/horizontal scale, support part-whole, comparison, and equal-groups layouts as distinct modes, and always render a ”?” marker on the unknown segment — matching the three canonical model types from the literature [4].
2. Touch interaction should be direct-manipulation, not menu-driven: let the child drag a bar’s edge to resize it, drag a divider to split a bar into named parts, and drag one bar to align it against another for comparison — mirroring the tablet-based proportional-reasoning study where dragging bar length (not typing numbers) produced the additive→multiplicative shift [11].
3. Auto-snap bar length to rough proportionality but not exact pixel-perfect scale for younger bands: the pedagogy explicitly treats bars as “roughly proportional,” not to-scale rulers, so the widget should reward reasonable proportion, not punish for exact-pixel mismatches [3][4].
4. Map CPA explicitly onto difficulty tiers, not just onto age: e.g., Tier 1 (K–G1) is Concrete-only (draggable counters/objects, no bars yet); Tier 2 (G1–G2) introduces Pictorial bar models with heavy scaffolding (pre-drawn bar outlines, labels supplied); Tier 3 (G2–G4) requires the student to construct the bar model themselves from a word problem; Tier 4 (G4+) drops the bar-model scaffold as optional and lets students choose between bar model and pure abstract/algebraic notation — consistent with how CPA is a progression within a topic and revisited spirally at each grade, not a one-time pass [1][6].
5. Number-bond widget as a distinct, earlier-introduced primitive from the bar model: a simple “two circles feeding one circle” (or bar-split) UI for parts-and-whole, used from the youngest grade band, that later visually morphs into the part-whole bar model — giving continuity between the two most load-bearing Singapore visual tools [5].
6. Represent the MOE Pentagon in the scoring/feedback model, not just in content design: the AI tutor’s post-challenge feedback should be able to speak to more than correctness — e.g., flag when a student picked an inefficient process/heuristic (Processes), praise persistence after a wrong attempt (Attitudes), and prompt a “did that answer make sense?” check (Metacognition) — giving Math Challenge’s feedback loop the same five dimensions MOE uses to define mathematical competence, not just right/wrong [7].
7. Build the content graph as a spiral, not a strict tree: the same underlying concept (e.g., “fractions as parts of a whole”) should have multiple entries across grade bands at increasing depth, and the platform’s spaced-repetition/points engine should be aware that revisiting an “old” topic at a harder tier is expected curriculum design, not remediation [6].
8. Do not oversell efficacy claims in marketing or parent-facing copy: because WWC found no qualifying studies for Singapore Math specifically, avoid claims like “proven to raise scores by X%”; instead cite the more defensible EEF Mathematics Mastery RCT figure (~2 months’ additional progress in primary, moderate security) if an evidence claim is needed at all, and label it clearly as evidence for a related UK mastery program, not a direct study of Math Challenge itself [9][10].
9. Bar-model authoring tool for question-writers/AI generator: since bar models are central to word-problem pedagogy, the internal problem-generation pipeline (including any AI-assisted item generation) should be able to emit a structured bar-model spec (bar type, segments, labels, unknown position) alongside the word-problem text, not just plain text — so the same problem can render consistently as text + interactive bar model across locales (EN/ES/FR/PT/DE) [3][4].
10. Accessibility/motor consideration for young children (ages ~4-7) on touch: given the fine-motor demands of precise dragging (Duijzer et al.’s participants were 9-11), younger bands likely need larger touch targets, snap-to-grid segment boundaries, and possibly a tap-to-select + tap-to-place alternative to continuous dragging, rather than assuming drag precision scales down to kindergarten hands [11].
11. Localize CPA-consistent iconography per language/theme, not just text: because the “Concrete” stage in Singapore classrooms uses locally familiar objects (counters, blocks), Math Challenge’s grade-band UI themes should swap concrete-stage imagery per locale/theme rather than reusing one universal object set, keeping the pedagogical structure (enactive→iconic→symbolic) constant while the surface representation varies [1].
12. Track model-type mastery as its own skill, separate from arithmetic mastery: a student can compute correctly but fail to select/construct the right bar-model type (or vice versa); the anti-cheating/points/AI-tutor system should be able to distinguish “wrong operation choice” (a modelling/process failure) from “arithmetic slip” (a skills failure), which the Pentagon framework already treats as separate dimensions (Concepts/Skills vs. Processes) [7].

Open questions for the project owner

1. Should Math Challenge license or closely emulate an existing digital bar-model tool’s interaction pattern (e.g., Visnos-style cut/join/clone bars, MathsBot-style manipulative) as a starting UX reference, or design the widget from scratch to fit our own grade-band theming?
2. Given WWC’s “no qualifying studies” finding, is the project comfortable citing only the UK EEF Mathematics Mastery RCT (a related-but-different program) as evidence in parent/teacher-facing marketing, or should we commission/track our own internal efficacy data instead of citing external Singapore Math studies at all?
3. Do we want the spiral curriculum graph (topic revisited across grade bands at increasing depth) authored by human curriculum designers up front, or generated/inferred by an AI content pipeline validated against a human-authored skill taxonomy?
4. For the youngest band (~ages 4-6), should the “Concrete” CPA stage be simulated on-screen (virtual counters) or does the product intend any physical/AR component, given research access here was limited to touchscreen-only studies (Duijzer et al. tested ages 9-11)?
5. How much of the Pentagon framework’s “Attitudes” and “Metacognition” dimensions should be surfaced in visible UI (e.g., a confidence check-in, a “did this make sense?” prompt) versus handled invisibly inside the AI tutor’s feedback text?