Step-level partial credit: why STEM grading needs more than a score
A final score tells students almost nothing about where their reasoning broke down. Here is how step-level grading changes what you can teach from an exam, and why it matters most in multi-step STEM problems.
Most grading tools treat a single exam question as a single number. A student either gets the point or they do not. In a humanities essay that compression is defensible — the work is argued, not derived. In a multi-step thermodynamics or circuits problem, it throws away the most valuable information on the page.
What a score hides
Consider a student who sets up an energy balance correctly, applies the right assumption, then makes a sign error in the final substitution. Their answer is wrong. But their understanding is mostly there.
A score of 0/10 and a score of 0/10 for a student who never wrote a valid governing equation look identical in the gradebook. They are not the same student, and they do not need the same feedback or the same remediation.
Step-level partial credit recovers that signal. Instead of one mark, each problem becomes a short sequence of scored checkpoints:
- choosing the correct governing principle
- setting up the relationship correctly
- carrying the algebra through
- arriving at a physically sensible result
The total still rolls up into a grade. But the breakdown is what lets you see whether a whole cohort is struggling with this specific step.
Why it matters more in STEM than elsewhere
In writing-heavy disciplines the artifact already encodes the reasoning. A reader can follow the argument. In quantitative work, the final answer is a lossy compression of a long chain of decisions — and the chain is the point. The whole reason we teach derivations is that the reasoning is transferable when the numbers change.
Partial credit is not a softer grade. It is a higher-resolution grade.
That resolution is what makes a question like “did students actually understand the boundary conditions, or did they get lucky?” answerable at all.
The cost of doing it manually
The honest reason most courses do not grade at step level is time. Reading a handwritten derivation, locating each checkpoint, and scoring it consistently across 120 students and three TAs is slow, and consistency drifts over an evening of grading.
That is the specific problem GradeFlow is built for: it reads handwritten equations, derivations, graphs, and diagrams, and scores them step by step — so the partial-credit detail that used to be too expensive to capture becomes the default output.
What to do with the data
Once each problem is scored step by step, a few things become practical:
- Per-concept analytics. Roll checkpoints up by concept and see which one the cohort missed, not just which question.
- Consistent feedback. Every student on the same step gets the same note, regardless of which TA opened their booklet.
- Targeted review. Spend the next class on the step that failed, instead of re-teaching the whole topic.
None of this requires changing your exam. It requires grading the exam you already write at the resolution your students actually reason at.