Multimodality, reasoning models and embodied AI — a maturity guide

At the start-of-term curriculum planning meeting, the IT team projected three words from a vendor deck: multimodal, reasoning, embodied AI. The principal turned to the English panel head: "Should we roll out all three this year?" She answered concretely: "Oral feedback we can trial next week; step-by-step mathematics can use reasoning mode when needed; a robot teaching assistant — we have not even chosen which room would hold it."

Exhibition decks like putting capabilities with wildly different maturity on one slide. This closing lesson of the LLM Classroom follows automated evaluation and leaderboards, separates the three trend types, and gives you judgement lines for procurement and pilots that keep working.

---

Three maturity levels

Multimodality — seeing images, hearing speech, reading handwriting — already works in Hong Kong classrooms today, not only in demo reels. Reasoning models — internal working before an answer — can be bought now but should be enabled by task, because they are slower and dearer. Embodied AI — AI inside robots and physical devices — remains largely at research and demonstration stage and should not drive procurement.

When all three sit in one "AI transformation plan", the usual outcome is: what should be trialled is not (oral practice still relies on teachers correcting every line), what should be economised is not (circular polishing runs on the dearest reasoning tier), and what should wait is prepaid (budget reserved for a "future robot assistant").

The most useful question about a trend is not "will it happen" but "can I buy it today, and does my school actually have this problem".

Multimodality: already a classroom tool

Multimodality means the model handles more than text — speech, images and handwritten input. These scenarios work in Hong Kong schools today:

Speech

Pupils read English or Putonghua passages aloud; the system transcribes and gives pronunciation feedback; teachers dictate lesson plans. Cantonese recognition now reaches "usable with teacher review on critical sounds" for primary pupils reading aloud, unclear articulation and classroom noise — English and General Studies teachers report saving roughly twenty minutes per week on line-by-line correction at Secondary 1, provided pupils know practice counts are logged for teacher spot-checks.

Images and handwriting

A teacher photographs a worksheet page and asks for variant questions of the same type; photographs whiteboard notes for a handout outline; scans handwritten answers into the marking flow. State the limits plainly: complex charts, mathematical notation and messy handwriting still fail, and the model sounds just as confident when wrong. Wherever marks are affected, teacher review cannot be dropped.

Reasoning models: enable by task

A reasoning model produces internal working before stating a conclusion — like a pupil who stops shouting the first answer and works on rough paper first. That brings three concrete changes.

Clear gains on multi-step problems — mathematics word problems, logical chains and debugging show fewer skipped steps. Slower and dearer — thinking longer uses more computation; for "tidy up this English sentence" it is pure waste. Adjustable effort — current models generally allow settings from quick answer to deep deliberation.

One misunderstanding: the reasoning shown to you is not necessarily what was computed internally — it is helpful text, usually consistent with the conclusion, but not proof. Teachers still check the conclusion itself. Practical rule: composition feedback, circular drafts and vocabulary exercises go to something fast and cheap; step-by-step mathematics and complex administrative calculations escalate to a reasoning model. For cost detail see context caching and cost control.

Embodied AI: still demonstration stage

Embodied AI puts AI inside something with a body — a robot, a robotic arm — so it perceives, plans and acts. Current state: several laboratories have demonstrated everyday manipulation from natural language instructions, but the distance to stable, safe, affordable deployment in a school is large, and nobody can honestly give a definite timeline.

For schools today, two meanings matter: as STEM teaching material — perception, planning and control are solid curriculum content; as information literacy practice — having pupils compare a demonstration video with a purchasable product. Writing it into a three-year procurement plan is not defensible now.

Four procurement judgement lines

One secondary IT team responding to a vendor deck skipped trend analysis and answered four questions item by item — more useful than any technology forecast.

Can I buy it today and afford it? Not "has anyone demoed it" but "are schools using it, what does it cost, who maintains it". Cantonese oral feedback: yes, already in the platform, rolling out at Secondary 1 this year. Embodied robot assistant: no, not in the plan.

What happens when it fails? An oral practice error makes a pupil laugh; a marking error affects university applications. Higher stakes mean more human review. Reasoning mathematics explanations: enabled for Mathematics and Physics only, with teachers spot-checking conclusions.

Does my school actually have this problem? English oral practice is short of staff time — yes. Need a robot patrolling classrooms — no. Capability without a pain point is a burden.

Does the contract let me take all data and leave? Betting on any vendor roadmap carries risk; a contract with one-click export to JSON, CSV and original files does not. Edor.ai commits only to live capability; unreleased features do not enter contracts; multimodal speech and image input are in use; reasoning models allocate by task difficulty; there is no embodied AI product.

Where people go wrong

Enabling the strongest reasoning mode for every task. Using a reasoning model to polish a circular is like sending a letter by freight truck. Self-check: does this task require multi-step reasoning? If it is only "smooth this sentence", use a fast-answer tier.

Turning exhibition demos into next year's procurement list. Self-check: for each demo item, can you name a school, a class and months of use? If there is only a video and no contract case, classify it as "discussion material" not "procurement item".

Ignoring multimodal overconfidence. Image misreads still sound certain. Self-check: could this affect marks or parent communication? If yes, output must pass teacher review before release.

Try it once

What to do: Spend ten minutes filling a "three-trend pilot sheet" for your panel — only items you will genuinely run this school year.

Step 1 (4 minutes): List three real pain points — for example, "not enough time to correct Secondary 1 oral work line by line", "variant question setting takes too long", "Secondary 3 pupils cannot follow step-by-step word-problem explanations".

Step 2 (4 minutes): Match each pain point to a capability: multimodal speech, multimodal image, reasoning explanation, or "not applicable (state why)".

Step 3 (2 minutes): For each item, write one sentence on "what happens if it fails" and "who reviews".

What to hand in: A nine-cell sheet (3 pain points × 3 columns: capability match, failure impact, reviewer).

How you know it worked: At least one item points to multimodality you can trial today (speech or image); reasoning appears only where multi-step reasoning is genuinely needed; nothing lists embodied AI as this year's procurement. If all nine cells say "not applicable", the pain points are too abstract — return to step 1 and rewrite as "which week, which class, which piece of work".

---

From lesson 1 to here, you have covered how a model works inside, the three training stages, the ecosystem and how models are obtained, application building, local deployment and fine-tuning, and safety, evaluation and ethics. You should now be able to explain to a colleague why AI gets things wrong, judge whether a vendor's safety claims are specific, read the limits of a leaderboard, and decide what data must never be pasted in.

Suggested next steps: Return to the LLM Classroom syllabus to revise, or continue with the Prompt Engineering Classroom — eighteen lessons set in Hong Kong classrooms, beginning with what makes a good prompt, covering planning, tiered worksheets, question setting, marking feedback, SEN adaptation and panel template libraries. The judgement taught here plus the technique taught there is the complete set.