Empower growth and innovation with the latest AI Dev insights

Child follows AI's geometry steps and gets points deducted—can switching to a bigger model in 2026 recover them?

Nov 15, 2026 Read: 16

Start with the conclusion: when a child follows the steps explained by AI for homework and the teacher deducts points, in 2026 the more common reason is not that the model cannot solve the problem, but that the explanation pipeline has only done “solving the problem” right and has not done “explaining by grade level, by textbook edition, and at the step granularity the teacher accepts.” The usual priority is to fix the explanation template and step validation first, and then decide whether to change the model—the rework cost of these two moves differs by around a factor of two, based on experience.

1. The deduction often lies not in the answer, but in the layer of “who the explanation is for”

The model outputs a general solution, while the student submits homework for a specific textbook and a specific grade level. For the same auxiliary line, different textbook editions may differ in writing order, notation conventions, and whether a certain theorem is allowed. The model has no obligation to align with these; if the business layer does not align, the situation where the teacher ticks the answer but crosses out the process will keep recurring.

To judge whether the capability layer needs to be changed, there is a near-zero-cost method: ask the same question twice. The first time, add no constraints; the second time, specify the grade level, textbook edition, and “no fewer than N steps, with a reason for each step.” If the solution direction is the same both times and only the level of detail differs, the model is sufficient and the problem is in the business layer; only if the solution direction changes between the two attempts is it worth comparing the image recognition and reasoning stability of different model families.

  • Grade-level scope: equations should not appear in primary school, and first-year middle school students have not yet learned fractional equations; the model can easily finish the problem with a “more advanced but not yet learned” method.
  • Textbook and solution preference: for the same geometry problem, whether similar triangles or the Pythagorean theorem is used, teacher acceptance differs.
  • Step granularity: a correct answer with skipped steps in the process is the main source of lost process points.
  • Missing validation: without rule-based validation and without teacher spot checks, skipped steps will keep repeating.

2. Four-layer alignment: capability layer, business layer, channel, and data and risk control

The reason for four layers is that they solve different problems: the capability layer determines “whether the problem can be solved,” the business layer determines “whether the explanation sounds like a teacher for this subject,” the channel determines “whether the student can see the question stem clearly,” and data and risk control determine “whether wrong answers will recur and whether content will go beyond the syllabus.” When these are mixed in one layer, a common symptom is repeatedly changing prompts while a batch of questions remains wrong.

  1. Capability layer: multimodal image recognition plus reasoning; the focus is not on leaderboards, but on recognition stability with handwriting, photo tilt, and mixed layouts of figures and formulas; the same problem photographed under different lighting may produce different recognition results.
  2. Business layer: write the four-part set of “grade level—textbook edition—knowledge point tags—step granularity” into the explanation template, and for geometry problems require output of “how the auxiliary line was conceived + the basis + the reason for each step.”
  3. Channel: mini programs, H5, apps, and learning devices all exist; the key is to keep a “question stem confirmation” step so students can correct misrecognized text and conditions.
  4. Data and risk control: a wrong-answer feedback loop into the database, a beyond-syllabus word list and knowledge point whitelist, and weekly spot checks with manual review—none of the three can be missing.

The acceptance line can be very specific: take 20–30 real questions from the target grade level through the full pipeline, and only when the proportion of steps accepted by the subject teacher reaches an experience range of 70% to 80% is the template detailed enough; below that number, fix the template first and do not rush to change the model.

3. Delivery in practice: a four-week first version, with the cost paid in one extra click

In a photo-based question-and-answer mini program project for middle school students, the constraints were typical: parents’ phone photos generally had mediocre lighting, the children spanned grades 7 to 9, the first-version budget was limited, and the timeline was compressed to around four weeks. The approach was not to cover all subjects at first, but only math, fixing the output structure for geometry problems as “given—find—auxiliary line—basis for each step,” and when recognition confidence was low, asking users to confirm the question stem before explaining. The cost was one extra click, and the explanation completion rate for new users in the first version would drop by a portion—typical range 10% to 20%, varying by user group and question type; but if this step is skipped, a common result is concentrated parent complaints about skipped steps two weeks after launch, followed by going back to change prompts and output structure, with more rework and scheduling pressure than building the structure from the start.

In similar projects, Xiyue Company also schedules in this order: define the output structure first, then discuss the model. Common lessons from doing it the other way include: only tuning prompts without changing the output structure—prompts can change tone but cannot change a hard constraint such as “must write the basis for the auxiliary line”; directly showing the original answer from a question bank—the answer is too brief and lacks process, so parents cannot understand it; feeding one template to both primary and middle school—beyond-syllabus complaints will concentrate in the primary school segment; and without spot checks, explanation quality stops on the launch day.

4. Direct output vs. question bank + templates: experience ranges for cost and timeline

The two routes are not opposites; they are more like two stages. If the first version needs to be fast, let a multimodal model output directly and get image recognition and the pipeline running first; to make step stability solid, it is usually necessary to add a layer of question bank and knowledge point tags. The following are experience ranges and will vary with subject, question volume, and team size.

  • First-version timeline: direct output commonly takes 1–2 weeks to run through the closed loop; question bank + templates + RAG commonly takes 4–8 weeks (experience range).
  • Cost per Q and A session: image understanding costs more than plain text; a single session commonly falls in the range of a few cents to one or two tenths of a yuan (experience range).
  • Upfront investment: money for the question-bank route is mainly spent on question cleaning, knowledge point tagging, and step structuring, not on the model.
  • Skipped steps and beyond-syllabus content: direct output relies more on prompts and spot checks as a fallback; the knowledge point whitelist route is more controllable, but maintenance cost persists.
  • Who it suits: for scattered question formats and wide subject coverage, use direct output first; for only one or two grade levels with relatively fixed questions, the question-bank route is more cost-effective.

A practical order is: first use direct output to get image recognition and interaction running, while turning questions returned by teacher spot checks into structured samples, then gradually switch to “retrieving similar questions + templated explanation.” When switching, only the explanation backend changes, and the product does not need to be rebuilt.

5. Where it applies and where it does not

It fits: science and math explanation for primary and middle school, homework tutoring where process points are graded, regions with relatively fixed textbook editions, and scenarios where parents are present but cannot explain the steps clearly. In these needs, the value of step structure is higher than the warmth of the wording.

It is not very suitable or not necessary for: tools that only require an answer and do not grade the process; competition problems and open inquiry problems, where the standard steps are themselves not unique; and formal in-school exam grading, which involves scoring credibility and complete records, must be handled separately according to exam system specifications, and should not be directly replaced by a consumer-grade Q and A pipeline.

A boundary judgment that can be quoted on its own: If the business only requires an answer and does not require process points, piling on explanation templates is over-design; conversely, as long as homework process is graded, step granularity must enter the business layer rather than being left to the model to improvise.

Common Questions

Can photo-based homework help skip the question bank and use a large model directly?

It can run. A common approach is to use direct output first, turn questions returned by teacher spot checks into structured samples, and then gradually add a question bank and knowledge point tags; piling up a question bank from the start costs more and shows results slowly.

How many steps are enough for explaining a geometry problem?

It depends on the grade level. Middle school geometry usually requires clearly writing how the auxiliary line was conceived and the basis for each step; better one extra step than a skipped step—lost process points mainly come from skipped steps and missing bases.

What if the photo taken by a parent is blurry and recognition is wrong?

Make question stem confirmation a required step so students can correct the text and figure conditions. This saves more cost than explaining mistakes afterward and takes effect faster than changing the model.

How can I judge whether an explanation product is good?

Take your own 20–30 real questions from the target grade level and run them through; see whether the steps are accepted by the subject teacher. Then ask whether there are spot checks and a wrong-answer feedback mechanism—these two points say more than a demo.

Does photo-based homework help require private deployment?

Public-facing mini programs mostly start with APIs paid by usage; private deployment is considered only for scenarios involving school data or where external network access is not allowed, and hardware and operations costs will rise noticeably.


One thing you can do right away: pick 20 real questions from the target grade level, run each under two conditions—without constraints and with constraints—and see whether the solution direction is stable, then decide whether to fix the template or change the model. This judgment applies to science and math Q and A and homework tutoring in primary and middle school; if the business does not grade process points, or the questions themselves have no standard steps, this step is not necessary.

Interested in this topic?
10-year tech team — reference proposal within 24 hours
Obtain Proposal
Are you ready?
Then reach out to us!
+86-13370032918
Discover more services, feel free to contact us anytime.
Please fill in your requirements
What services would you like us to provide for you?
Your Budget
ct.
Our WeChat
Professional technical solutions
Phone
+86-13370032918 (Manager Jin)
The phone is busy or unavailable; feel free to add me on WeChat.
E-mail
349077570@qq.com
Submitted successfully
Thank you for your trust. We will contact you soon!
Recommended projects for you