AI Homework Grading Workflow

Why this exists

Even after homework load was reduced, schools still assign work to keep teaching quality, and subject teachers still mark everything. The point of marking is not finishing the stack. It is letting teachers see the knowledge each student’s answers expose. For teachers, this is mechanical, repetitive, and still cannot be skipped.

If an AI workflow can mark homework and gather problems and error-cause analysis into a report teachers can inspect, teachers do not have to choose between “seeing learning state” and “less mechanical labor”: insight into mastery stays, and the burden can come down.

First define a reliable question object

For that report to be trusted, what decides the result first is not the marking prompt. It is which question the model actually saw. A real paper is not one neat image per item: a large item may contain several sub-questions; stem and answer may cross a page; a diagram may sit far from the stem; formulas, tables, and handwriting may be split into different regions by layout models. If the upstream cut on question boundaries is wrong, later OCR, reference answers, and knowledge points keep reasoning on the wrong object, and the report looks complete while answering the wrong question.

I split what a paper has to restore into three relations:

• Question boundary: large-item numbers do the split; sub-questions, working, and student answers stay inside the same item; first and last blocks on adjacent pages have to be judged as one cross-page item or not.
• Figure–text relation: stems, tables, formulas, diagrams, and handwritten answers have to be composed in reading order, without hanging a neighboring item’s figure on the wrong question.
• Result evidence: each item stores not only a judgment, but the original crop, stem, student answer, reference source, error cause, and knowledge-point mapping, so the first wrong step can be found.

How the product is organized

I closed the demo into four sequential processing stages. Each stage produces a savable intermediate state, so problems are not only found in the final report.

1. Automatic splitting is not one model call. It is routing with a quality bar

The main path first cleans noise blocks through layout analysis, restores reading order, and builds candidates from large-item numbers; sub-item numbers are not mistakenly split into new large items. The system checks result quality by candidate count, coverage area, and block height: if the main path is unreliable it falls back to multimodal whole-page splitting; only if that also fails does even splitting keep the flow from stopping.

For multi-page papers, the system sends the bottom block of the previous page and the top block of the next page into a cross-page judgment; they merge only when position, number continuity, and semantic evidence all hold, and the merge reason and confidence are saved.

2. Figures cannot be associated only by “who is nearest”

When a stem says “as shown” or “as in the table,” but the current item has no matching visual block, the system recalls candidate figures from the same page and adjacent pages. A rule layer first compares page, number hints, figure type, context keywords, and spatial position, then a vision model judges among a limited set of candidates; after a high-confidence hit, the item crop is concatenated with the candidate figure and re-recognized.

This design is not “is there an image.” It is “which question does this image belong to.” If evidence is thin, the main flow keeps a missing-figure state and continues, rather than forcing a candidate that looks reasonable.

3. Per-item feedback and the whole-paper report finish separately

After splitting, each item independently goes through stem and handwriting recognition, model pre-solving, formal marking, error-cause or completion advice, and knowledge-point mapping. Items are processed concurrently; whichever finishes first returns first. Students do not wait for the whole-paper summary before seeing the first item.

The front end organizes each item as three tabs: answer status, deep explanation, and knowledge-point mapping. Compound items keep a judgment for every sub-question; when no answer is detected, the student side hides the reference answer and item analysis, so a blank answer does not become an answer-getting entry.

4. Knowledge-point mapping must be constrained by a knowledge graph

If the model freely generates knowledge-point names, the same concept can appear in several spellings and cannot be aggregated stably. I used “semantic extract → rule recall of candidates → model chooses under constraint” to map each item to a primary knowledge point and test point in the matching subject graph.

Mapping results also save confidence, match reason, and evidence spans. When evidence is thin, the system can return abstain; only high-confidence mappings enter weak-point stats. The summary layer then connects lost-score knowledge points with prerequisites and test points, giving study advice a structured basis.

Which complex paths the demo actually ran

The local task directory holds 30 completed-task snapshots, covering 364 structured questions. These records include repeated debugging and replay, so they cannot be treated as scale or accuracy metrics. They can prove the main product chain is not stuck in a PRD and a static prototype.

Of those, 292 items produced a high-confidence knowledge-point mapping. That number still only describes mapping state on the current local sample. It cannot be derived as mapping accuracy; formal evaluation still needs a human-labeled question–knowledge ground truth.

How failure returns to a checkable state

• 01
  Layout service unavailable

  Fall back to multimodal whole-page splitting; if that still fails, use even splitting as a last resort and keep the source for later investigation.

• 02
  No standard answer

  Let the model solve independently first as a reference, then compare with the student answer; the reference source is saved with the result.

• 03
  Thin knowledge-point evidence

  Allow abstain, so low-confidence results do not enter weak-point stats.

• 04
  Student did not answer

  Keep an unanswered judgment, but hide the reference answer and deep analysis on the student side, so feedback does not change answering behavior.

Results and limits

This demo has validated that multi-image upload, paper structuring, cross-page merge, missing-figure matching, per-item concurrent marking, progressive SSE return, knowledge-point mapping, and a summary report can form an end-to-end chain. It also made clearer that a formal product cannot replace process quality with “the final answer looks correct.”

The current version is still single-tenant product validation on local JSON storage: no account and permission system, no class management or teacher-review workbench, and no long-term knowledge profile across assignments. Marking, splitting, and knowledge-point mapping still need separate human-labeled sets to evaluate error boundaries, high-confidence false calls, and fallback paths — not only overall average accuracy.

What I took from the project

This experiment made me recast AI marking from a “large-model capability question” into an observable product chain: once input structure is wrong, every later model can keep giving a reasonable explanation on the wrong object. What product design actually has to do is make question boundaries, figure–text relations, reference sources, mapping confidence, and failure fallbacks all visible, so a problem can be traced to the first place it went wrong.

Scope note: This is product validation and a local demo, not described as a production system. Task snapshots include debugging and reruns. They prove functional paths were actually run, not marking accuracy or scaled use.