The science behind 
Every design decision is grounded in decades of cognitive science research.
of classroom students outperformed
In 1984, Benjamin Bloom found that students receiving 1-on-1 tutoring performed two standard deviations better than classroom students. He called it the 2 Sigma Problem — the effect is massive, but personal tutoring doesn't scale.
changes the equation
Not by replacing teachers, but by giving every student the adaptive, responsive feedback loop that only 1-on-1 tutoring could provide before. Units is built on this insight.
Bloom, B. S. (1984). The 2 Sigma Problem. Educational Researcher, 13(6), 4-16.
7 evidence-backed principles
Ranked by effect size. Every one is built into Units.
Immediate feedback
d = 0.73After every answer, you get an explanation of WHY — not just whether you were right or wrong.
Every question in Units shows an inline explanation the moment you answer. Wrong answers get targeted explanations for each incorrect option. The AI coach provides follow-up if you're still stuck.
Hattie & Timperley, 2007
Spaced repetition
d = 0.60Review at the right intervals to make memory permanent. Without it, you forget 90% within a week.
Units uses the SM-2 algorithm to schedule reviews at day 1, 3, 7, 14, 30 — adapting based on how well you remember. You never have to decide when to study.
Cepeda et al., 2006; Ebbinghaus, 1885
Worked examples
d = 0.57Before asking you to solve something, we show you how it's done — step by step.
Each new concept begins with an interactive worked example where steps reveal one at a time. You build understanding before you're asked to perform.
Sweller et al., 1998
Mastery-based progression
d = 0.52You advance when you prove you know it — not when the calendar says so.
Every concept requires 3 consecutive correct answers. If you struggle, the system routes you to the prerequisite concept that's causing the gap.
Bloom, 1968; Kulik et al., 1990
Retrieval practice
d = 0.50Testing yourself is 2x more effective than re-reading the same material.
Every concept ends with retrieval questions — MCQ, fill-in-blank, ordering, matching, highlighting. You never earn XP for passive reading.
Roediger & Karpicke, 2006
Interleaving
d = 0.42Mixing different types of problems feels harder but produces deeper learning.
Review sessions deliberately mix concepts from different units. Related concepts are spaced apart to prevent associative interference.
Rohrer & Taylor, 2007
Elaborative interrogation
d = 0.42Asking 'why does this work?' deepens understanding beyond surface-level memorization.
Every correct answer includes a 'Why?' explanation. The AI coach pushes you to explain reasoning, not just answer.
Pressley et al., 1992
Lesson architecture
Every concept follows the same 6-9 minute research-backed flow.
Hook
30s
Visual puzzle that sparks curiosity
Explore
2-3m
Discover the pattern through interaction
Explain
1-2m
Formalize with concise text, one idea per screen
Elaborate
1-2m
'Why does this work?' — articulate reasoning
Retrieve
2-3m
3-5 questions with immediate feedback
Master
3 correct in a row to advance
Optimal session structure
Spaced repetition items due today, interleaved across topics
1-2 new concepts using the flow above
"You reviewed 8 concepts, learned 2 new ones"
The forgetting curve
A platform that teaches but doesn't schedule review is wasting most of its effort.
forgotten within a week
Without review, you forget 50-70% within 24 hours and 90% within a week. This is why most online courses don't produce lasting knowledge — they teach but never revisit.
retrievals to permanence
After 4-5 well-timed retrievals, memory becomes nearly permanent. Units schedules these automatically — day 1, 3, 7, 14, 30, then extending. You never decide when to study.
Ebbinghaus, H. (1885). Cepeda et al. (2006). Spacing effects in learning. Psychological Science, 19(11), 1095-1102.
The learning engine
The most rigorous adaptive learning system in production. These platform-level systems make the learning engine work.
Knowledge graph
Every concept connects to its prerequisites. A course is a graph of interconnected ideas, not a linear sequence. The system pinpoints exactly where a gap exists and routes you there.
Implicit repetition (FIRe)
Practicing advanced topics implicitly reviews prerequisites. The system credits those reps to the prerequisite's SRS schedule — so you spend less time grinding old material.
Halt and remediate
If you're struggling, the system pauses the topic and routes you to unrelated material. You return later with fresh eyes. If the gap persists, it traces back to the exact prerequisite.
Adaptive diagnostics
A placement test finds your knowledge frontier. Correct answers let you skip prerequisites. Students routinely skip 20-40% of content on day one.
Non-interference scheduling
Similar topics are spaced apart to prevent confusion. Review sessions interleave concepts from different units so you learn to distinguish strategies.
Per-student pacing
The system tracks how quickly you master each topic. Easy concepts get fewer reps. Hard ones get more frequent review. Every student moves at their own speed.
10x scaffolding
Each concept has three difficulty levels: worked example, practice problems, and mastery check. 10x more granular than a textbook.
Automaticity
Practice until low-level skills become effortless. When basics are automatic, working memory is freed for higher-level reasoning.
The neuroscience
Meta FAIR's TRIBE v2 model provides direct neural evidence for why multimodal, interactive lessons outperform passive text.
stronger encoding
Multimodal stimuli (visual + audio + text) produce up to 50% stronger neural encoding in brain integration areas vs any single modality alone.
then language
Visual cortex activates first, then propagates to language areas — matching the principle of starting with visual examples before formalizing into theory.
beats passive
Responding, choosing, and manipulating activates brain areas that passive reading does not. Every step in Units requires you to do something.
What this means for lesson design
Our lessons combine visual illustrations, text explanations, and interactive responses in every concept — not as decoration, but because the neural evidence shows combined modalities produce deeper memory traces. The TRIBE v2 model (1B parameters, trained on 1,115 hours of fMRI across 720 subjects) consistently shows that multimodal stimuli produce stronger encoding than any single channel.
d'Ascoli et al. (2026). TRIBE v2: A foundation model of vision, audition, and language for in-silico neuroscience. Meta FAIR. Open weights: facebook/tribev2 on HuggingFace.
Gamification that works
Not all gamification is equal. Some patterns genuinely improve retention. Others create hollow engagement.
What we use
- Mastery maps that show the skill tree filling in
- Streaks with streak freezes to reduce anxiety
- XP for active retrieval only — never for passive reading
- Memory strength indicators per concept
- Personal bests over competitive leaderboards
- Milestone celebrations with sound effects and confetti
What we avoid
- ✕Timed challenges for conceptual learning
- ✕XP for passive behaviors like watching
- ✕Excessive badges that lose all meaning
- ✕Making lessons feel effortlessly smooth
- ✕Competitive leaderboards that discourage
- ✕Punitive mechanics — losing progress, public failure