Withdrawal groups, setting or mixed classes? What the grouping evidence says

Most of the figures here come from Evidence for Learning’s Teaching & Learning Toolkit, the Australian version of the Education Endowment Foundation’s toolkit. Each carries its evidence rating, because the ratings matter as much as the numbers. The toolkit pages were last updated in 2021, and figures can change when it does. Much of the research on grouping is from England or the US. The only Australian primary study we saw is the NSW streaming study below, which we know only as Evidence for Learning reports it.

Do withdrawal groups work?

It depends what happens in them. Structured small-group tuition adds about four months of progress on average, and about three in maths. Evidence for Learning rates the evidence as moderate, from 62 studies. It works best in groups of up to about six, when it is targeted at specific needs found by assessment, and when it is linked to classroom content.¹

Teaching assistants delivering structured, time-limited interventions add about four months, and about three in primary maths (65 studies, rated moderate). Teaching assistants used as general classroom support show no consistent benefit. Where effects were negative, their time had likely replaced teacher time rather than added to it, and lower achievers can end up with less contact with the teacher.²

So the negative findings are about unstructured support that replaces the teacher. The Education Endowment Foundation’s guidance adds a warning about withdrawal that makes a child miss new class content. It says to fix core classroom teaching first so fewer children need extra help, then use structured intervention guided by assessment, link it to whole-class teaching, and pay careful attention to what a child might miss when they’re taken out. If the intervention isn’t better than the class teaching they miss, the child can end up worse off than if they had stayed in the lesson.³ That is guidance, not a measured effect.

Australia’s AERO describes small-group support as supplementing whole-class teaching: “it does not replace it”. It should have clear entry and exit criteria, and be faded when goals are met.⁴ That guidance is written for Years 7 to 9, not primary, and it doesn’t grade the evidence.

Does setting or streaming raise attainment?

On average, the toolkit says no. Setting and streaming add no months of progress, with a small negative effect for lower achievers and a small positive one for higher achievers. The pattern looks similar in maths, reading and science, and in primary and secondary. The evidence is rated “very limited” (58 studies, updated July 2021), and much of it is old and about US-style tracking.^5,6 An NSW study of streaming in primary schools found “a zero to negative effect on student outcomes”, as reported by Evidence for Learning, which we haven’t checked against the study itself.

Students from disadvantaged backgrounds are more likely to be placed in lower sets. The toolkit describes some evidence, mostly from UK studies, that they are more likely to be misallocated there,⁵ and the Education Endowment Foundation’s guidance says a large number of meta-analyses consistently suggest setting widens the difference between lower and higher sets.³ Whether setting leaves disadvantaged students worse off is now contested, as the next section shows.

Is mixed-attainment teaching better?

The evidence doesn’t say so. There is no randomised trial showing that mixed-attainment teaching raises attainment. A 2018 pilot in English secondary schools found schools “largely reluctant” to be randomised and gave no impact findings. A larger trial of improved setting, also in English secondary schools, found “no evidence that it improves attainment or self-confidence”, but had high attrition.⁷

The newest and largest study is the Education Endowment Foundation’s Student Grouping Study.⁶ Its limits matter, so here they are first:

• It covered English secondary schools, Years 7 and 8 maths, from September 2022 to July 2024. It says nothing about primary schools or Australia.
• It was not randomised. It compared 28 schools using mixed-attainment classes with 69 matched schools that set by attainment.
• Its security rating is 2 of 5 padlocks, low to moderate, partly because about 29% of students dropped out of the analysis and the two groups of schools may differ in ways it couldn’t measure.
• It compared two whole-school ways of forming classes. It tells us nothing about grouping within a class or about withdrawal.

With those limits, what it found was this. Students in the mixed-attainment schools made about one month less progress in maths on average. That difference was not statistically significant, so it is a best estimate, not a demonstrated harm. Students eligible for free school meals (a measure of disadvantage in England) and students with lower prior attainment made similar progress in both kinds of school. The authors say the free school meals result “questions” earlier evidence that setting disadvantages students from disadvantaged backgrounds. They don’t say that about lower prior attainers. Students with higher prior attainment made about two months less progress in mixed-attainment schools. That is a single subgroup result, and it wasn’t security-rated.

The report also found negative effects on maths self-confidence in the mixed-attainment schools, but only the effect for lower prior attainers (moderate) was statistically significant. The overall and free school meals effects, both small, were not. Teaching was broadly similar across the two kinds of school, but mixed-attainment lessons looked more like bottom-set lessons, especially in pace, and only the setting schools appeared to be stretching higher attainers. So mixing classes did not by itself make teaching differentiated.

Put together, the evidence doesn’t settle setting versus mixed classes. The older reviews lean slightly against setting for lower achievers. The newest study leans slightly the other way overall, at low to moderate security, and it found less progress for higher attainers in mixed classes, a subgroup result with no security rating.

What about grouping inside the class?

Grouping students by attainment within the class for particular activities adds about two months overall, about three in primary and about four in maths. The evidence is rated “very limited” (22 studies, updated August 2021). Groups need regular review. The page also tells schools to consider the effect on students with low prior achievement, and says grouping may have longer-term negative effects on the attitudes and engagement of lower achievers.⁸

A review of 100 years of meta-analyses found within-class grouping gave small benefits, and grouping into separate classes gave negligible effects.⁹ It is built on older meta-analyses, and unlike the toolkit it found no difference by achievement level. Reviews disagree on lower achievers, and the 2026 study found similar progress for lower prior attainers whether schools set or mixed.

A meta-analysis of 21 studies of differentiation in primary schools found small to moderate positive effects when differentiation was done with computer systems or as part of a broader school program, and a small negative effect of ability grouping for lower prior attainers.¹⁰

Does adaptive software help?

A little, and not for everyone. A meta-analysis of intelligent tutoring systems in K to 12 maths, covering 26 reports, found no harm and perhaps a small positive effect, with smaller effects for lower achievers than for students generally.¹¹ A broader review of 50 tutoring-system evaluations found larger gains, but much larger on locally developed tests than on standardised tests.¹² Read the two together.

Individualised instruction, including digital, adds about four months on average and about three in maths, with evidence rated limited. It works best supplementing class teaching, not replacing it, because it can leave teachers less time for teaching interactions.¹³

What does that mean for a numeracy block?

The evidence doesn’t back a firm rule on how to form classes. It points toward a few things, none of them settled:

• However classes are formed, teaching inside them should adjust to what each child knows. That means flexible task grouping, assessment-led targeting and stretch for higher attainers.
• Strengthen core class teaching first, so fewer children need extra help.
• Where a child does need it, use structured, short, time-limited small-group help, linked to what the class is learning and aimed at a need an assessment found.
• Check what a child misses when they’re withdrawn, and whether the help is better than the lesson they’re leaving.
• Keep any groups flexible and review them regularly.
• Treat adaptive software as a supplement to teaching, not a replacement for it, and expect small effects.

How Workings fits

Workings is adaptive software, so the adaptive-software evidence above is the evidence closest to it. It shows small, mixed effects on standardised tests, and smaller effects for lower achievers. We haven’t yet measured Workings’ effect on learning. We don’t publish efficacy percentages or “grade levels of growth”, and we plan an independent, pre-registered evaluation.

Here is what it does. Workings doesn’t replace your maths program: it’s practice and diagnosis, not a scope and sequence, and it sits alongside your teaching. In twenty minutes, every child works at their own frontier, with no grouping and no withdrawal. The class grid shows every child’s working stage and mastery by strand, names the prerequisite a group is missing, and sends stall alerts with a suggested teaching move. That is information a teacher can weigh alongside their own assessment when deciding who needs a structured small group. See what teachers and leaders see.