Grading is where TA hours disappear. TAs describe hundreds of submissions, half a minute or more per question, and makeup quizzes that send you back to grade again. Speed and quality are not opposites here: most of what makes grading faster also makes it more consistent, because you make each decision once and apply it the same way to everyone. This page collects those habits, the Gradescope features that support them, and what research says about how consistent graders really are. Your instructor's grading rules come first; anything that changes the grading scheme is their call.
Defaults that work. If you only take four things from this page:
- Grade one question for everyone before you start the next.
- Put your feedback in the rubric items, and write it once.
- Calibrate before you start and spot-check after, with the other graders.
- Fix the rubric, not one student's score, when you find a mistake, so everyone gets the fix.
How consistent are graders, really?
Less than we think. In a 2025 study, 28 experienced graders scored the same 40 intro Java programs in groups of four. Their agreement was very low on every skill they graded: correctness, elegance, readability and documentation. When one program came back two weeks later, almost none of the graders gave it the same grades again, although most believed they were consistent. A study of CS1 exams found that instructors disagreed on points "even for very simple programs." Studies of essay scoring also find that the paper you just graded pulls your score on the next one toward it.
None of this means grading is hopeless. It means consistency comes from the process: a clear rubric, calibration, grading by question and regular checks, not from trying harder.
Before you start
- Look at a sample first. Read 10 or so submissions before you finalize the rubric. You will see the common mistakes and can write a rubric item for each.
- Calibrate with the other graders. Everyone grades the same few submissions on their own, then compares and agrees on the differences. Fifteen minutes here saves hours of regrade requests.
- Agree on who may edit the rubric. In Gradescope the instructor can set Rubric Permissions to all staff, instructors only, or no one. Agree on it before grading starts, so one grader's edit does not quietly change everyone else's grades.
- Set a pace. Decide a target rate before you start, for example a number of problems per hour, and put hard cases aside to grade at the end.
While you grade
- Grade by question, not by student. You keep one rubric in your head, you get fast at it, and you do not see names. In Gradescope: Grade Submissions, choose the question, and use Next Ungraded (the Z key), which skips submissions another grader has open.
- Learn the keyboard shortcuts. Number keys apply rubric items, Z goes to the next ungraded submission, and the arrow keys move between submissions. The Gradescope guide lists the rest.
- Use answer groups on fixed-template exams and quizzes. Gradescope can group identical or similar answers (multiple choice, fill-in-the-blank, math fill-in) so you grade each group once. Review the suggested groups before grading them, and grade the ungrouped answers one by one at the end. This needs the institutional license and works on PDF and bubble-sheet assignments, not online assignments.
- Write feedback into rubric items. A rubric item like "−2: off-by-one in the loop bound; trace it with n = 1" gives every student the same clear comment with one keystroke. Save personal comments for work that will be revised or that misses a key goal.
- Comment on what matters most. One or two points tied to the learning goals help a student more than a dozen small marks, and take you less time.
- Re-anchor. After a break, look over a few submissions you already graded before you continue. When you finish, regrade the first few and see whether you still agree with yourself.
- Take breaks. Long, unbroken grading sessions make drift worse.
When you find a mistake in the rubric
In Gradescope, editing a rubric item changes the score of every submission that already has it. That is what you want for a real mistake, because everyone gets the fix at once. For a change that should apply to one student only, use a submission-specific adjustment instead, and leave a comment saying why. The magnifying glass next to a rubric item shows every submission it was applied to, which helps you check a change before you make it. Tell the other graders and the instructor when the rubric changes.
Code plus a report
- Let the autograder check correctness, and spend your time on what it cannot judge: design, readability, the report. Research on autograding tools finds that most check correctness through unit tests, and very few judge quality.
- In Gradescope, programming assignments can be autograded, hand-graded with a rubric and inline comments, or both (turn on Enable Manual Grading). Rerunning the autograder keeps the manual grading.
- Don't grade the report alone and assume the code matches. A quick look at the test results for each submission keeps the two honest.
Regrade requests
- Restate the grading policy when grades come out. It heads off requests that the rubric already answers.
- Answer through the rubric: change a rubric item if it was wrong for everyone, or add an adjustment if it was wrong for this student, then close and respond.
- Ask your instructor to set start and end dates for regrade requests, so they do not stay open all term.
- See Handle It or Hand It to the Instructor? for which regrades are yours to decide.
Makeups and multiple versions
- Import the rubric from the main version for any question that a makeup shares, so it is graded with the same items.
- Grade makeups with the same rubric, ideally by the same grader who graded that question on the main exam.
Bring this to your instructor
If grading still runs over your hours, the fix is a design decision, so raise it with your instructor rather than cutting corners on your own. Designing a rubric for grading time, on the rubric page, goes through it step by step.
- A coarser scale. Graders disagree most on fine point scales. A few clear levels per criterion (for example full, partial, none) are faster and more consistent.
- Specifications grading: pass or not yet against clear specifications, often with a chance to revise.
- Grade a sample of practice problems in detail and the rest for completion.
- More autograding for the parts that can be checked mechanically.
Keep a record of your grading time. "Grading Homework 4 took 9 hours for 120 submissions" is the kind of number that gets a workload changed.
Further reading
Gradescope guides
- Grading submissions with rubrics: grading by question, rubric items, adjustments and comments.
- Submission navigation and keyboard shortcuts.
- AI-assisted grading and answer groups.
- How can I edit the rubric for this assignment?: rubric permissions.
- Managing Regrade Requests.
- Grading a Programming Assignment: autograding and manual grading together.
Teaching centers
- UC Berkeley GSI Teaching & Resource Center, Grading and Time Management: pacing, sampling, re-anchoring and comment habits.
- UC Berkeley GSI Teaching & Resource Center, Grading Student Work.
- CMU Eberly Center, Creating and Using Rubrics.
Research
- Messer, Brown, Kölling and Shi, "How Consistent Are Humans When Grading Programming Assignments?": the 28-grader study.
- Albluwi, "A Closer Look at the Differences between Graders in Introductory Computer Science Exams" (IEEE Transactions on Education, 2018).
- Zhao et al., "Sequential Effects in Essay Ratings" (Frontiers in Psychology, 2017): how the previous paper pulls the next score.
- Messer et al., "Automated Grading and Feedback Tools for Programming Education: A Systematic Review".
- Nilson and Packowski, Specifications Grading 2.0 (Routledge, 2025).
Related pages: Course Tools for TAs, Coordinating Graders and UCAs, Giving Feedback Students Use, Creating Rubrics and Grading to Them, Working With Your Instructor and Handle It or Hand It to the Instructor?.