acceptodds
Under review as a conference paper at ICLR 2027

Where LLM Graders Succeed and Break: Evidence from Two Computer-Science Exams

Abstract

One long-form exam in a large course costs hundreds of grader-hours, and qualified graders are scarce; LLM graders are a tempting alternative. To show its pitfalls we grade a practical Computer Vision exam ( dual-graded students) under configurations spanning closed and open-weights models; the best reaches mean absolute error , below the two human graders achieve against each other. The catch is the prompt: a short ”strict grader” preamble drives of open-weights models out of the graded band (), three stopping grading altogether. The damage traces to the preamble's two credit-withholding sentences, not to tone or model scale; one of them, ”never give partial credit”, alone makes two of three probed models stop grading. The closed flagships of three vendors shift calibration under it but stay in the band. In further configurations on a second, independent Machine Learning exam from another course ( dual-graded students), the preamble worsens ten models, moving three out of the band into collapse and one into refusal, yet improves seven whose neutral prompts over-mark: the vulnerability replicates, but its direction is exam-specific. Light LoRA fine-tuning repairs it: one adapter on the two exams' pooled graded examples brings five small open models to parity or better with a human grader in agreement with the grader pair, and sensitivity to the three harsh personas nearly vanishes ( MAE). We release the anonymised dataset, full ablation grid, and grading, fine-tuning and analysis pipelines.

Primary area:
datasets and benchmarks
Keywords:
LLM-as-a-judge, automated grading, prompt sensitivity, instruction following, persona prompting, LoRa fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rcZnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies LLM grading of long-form programming submissions using two university exams. Across 333 configurations, the authors evaluate model choice, prompt components, grading personas, and decoding settings. Several closed models achieve strong agreement with human grades, while a credit-withholding preamble causes severe under-grading in many open models. Ablations attribute this effect primarily to policy instructions rather than persona adjectives. LoRA fine-tuning on approximately 3,900 graded examples substantially improves five small open models and reduces their aggregate error sensitivity to the tested harsh prompts. Cross-exam experiments also show generally beneficial transfer.

Critical strengths
  • The study provides a strong empirical validation. It uses two real exams, more than 1,600 students, many model families, and extensive prompt ablations, making the observed failure mode convincing and practically relevant.
Critical weaknesses
  • The proposed method seemed to be limited in fine-tuning generalization. The strongest results evaluate new students on exam questions and rubrics seen during training. Thus, the claimed transferable grading ability and robustness on unseen exams remain uncertain.

  • The evaluation measures agreement with two noisy human graders instead of correctness against an independent gold standard. Grader-pair clustering may also make the reported confidence intervals optimistic.

Questions for authors
  • Can authors report persona sensitivity for the cross-exam adapters, i.e., an adapter trained only on one exam and evaluated under neutral and harsh personas on the other?

  • Do the fine-tuned models remain responsive to legitimate, non-conflicting rubric or grading-policy changes? A controlled test would clarify whether the reduced persona sensitivity is desirable robustness or general instruction insensitivity.

  • How do the central parity confidence intervals and verdicts change under a grader-pair/block bootstrap or another analysis accounting for the grading-assignment structure?

  • What institutional review, student consent, data-use authorization, and re-identification audit supported the public release of complete student notebooks and grades?

1 comment

Thank you for the review. New text is blue, reworded text is red.

1. New students, not new exams (W1). You are right. The paper now says the human-level results are for new students of an exam seen in training. For an unseen exam we ran the same test on the adapters trained on the other exam only (new Table 21): they improve grading in 9 of 10 cells, but none is shown to be as good as a human grader.

2. Agreement with noisy graders is not correctness (W2). Agreed. Neither exam has a gold standard. This is now stated in Limitations.

3. Grader-pair bootstrap (Q3). Done (new Table 20). We resample whole grader pairs instead of single students.

  • Fine-tuned on human marks: all 9 "better than a grader" results hold.
  • Fine-tuned with distilled breakdowns: its 2 "better" results drop to "no worse", and 4 cells become inconclusive.
  • Zero-shot closed models: only D01 stays "better"; F02 and D02 drop to "no worse".

So some of our intervals were too optimistic, as you suspected. The text is corrected.

4. Persona sensitivity of the cross-exam adapters (Q1). Done (new Table 23). Each single-exam adapter graded the exam it never saw, under all five personas.

  • No adapter collapses. A harsh persona moves MAE by -1.45 to +1.18; the untuned models move by up to +19.75.
  • They are not flat either. 20 of 30 harsh cells stay within 0.5 MAE, but Gemma-4 (CV-trained, strict, ML exam) gets worse by 1.18.
  • The lenient persona raises MAE by up to 3.18.

So the stability we reported holds for adapters trained on the exam being graded, and only partly for an unseen exam. Section 8.4 now says this.

5. Do fine-tuned models still follow legitimate policy changes (Q2)? No. We ran the controlled test you suggested (new Table 24). One sentence was added to the neutral prompt: deduct 2 marks per question as a late penalty, or add 2 marks as an approved adjustment.

  • Fine-tuned (pooled) models apply -22% to 16% of the requested change (median 2%).
  • Untuned models are erratic: from -125% to 144%.

So your second reading is the right one: fine-tuning makes the grader ignore grading instructions in the prompt, legitimate ones included. The abstract, Section 8.4, the Discussion and the Conclusion now say this, and we recommend applying policy changes outside the model.

Both experiments use the paper's own adapters. As a check we also retrained all adapters with the same recipe and repeated both experiments; the retrained adapters match the originals to within 0.14 MAE on average (0.63 at most) and give the same picture.

6. Ethics (Q4). The statement now says the data are pseudonymised, not anonymous, and names the remaining risk. Grading was part of normal assessment and the release has the instructors' permission. Students agreed to this use of their work when they signed up for the course, and the release was approved by the head of research ethics of the department where the courses took place. All students were undergraduates; none was a minor.

Re-identification audit: the release went through several rounds of manual checking, and we also ran an automated scan of every notebook. Neither found any student identifier (names, emails, ID numbers, account names, personal file paths). The Ethics Statement now says this.

Reviewer_EBtwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the accuracy and prompt sensitivity of LLM graders on two university computer-science exams with 570 and 1,038 dual-graded students. Across 171 CV-exam and 162 ML-exam configurations, it compares closed and open models, prompt components, personas, and decoding choices. A strict preamble containing two credit-withholding instructions severely degrades many open models; controlled ablations attribute most of the effect to those instructions rather than the persona adjective. The second exam reproduces the vulnerability but reverses its direction for several over-marking models. Finally, LoRA training on approximately 3,900 pooled examples substantially improves five small open models and reduces the increase in aggregate MAE under the tested harsh personas. The submission releases the exam data, grading outputs, and analysis and training pipelines.

Critical strengths
  1. The empirical design makes the failure mode substantially more informative than a generic demonstration of prompt sensitivity. Matched neutral/strict comparisons span 17 open models from six families, and the adjective controls and S1/S2 factorial probes isolate the consequential change in grading policy (Sections 5.1–5.2; Tables 3–5). The second exam is particularly valuable: seven models improve under the same preamble, showing why an apparently helpful persona can be accidental bias correction rather than transferable robustness. This is useful evidence for evaluating LLM judges beyond a single prompt or benchmark.

  2. The dual-human labels and released artifacts make this a useful empirical resource, and the repair is practically credible. The pooled marks recipe uses human labels without a closed-model teacher, improves all five models on both held-out exams, and is accompanied by cross-exam transfer and matched base-model comparisons (Section 8; Tables 1, 18–19). I independently reproduced the two human-disagreement values, D01's CV MAE, and the pooled marks MAEs and paired-comparison intervals from the supplied tables. These checks support the numerical evidence behind the main findings.

Critical weaknesses
  1. The reported robustness measure does not establish the degree of grading stability suggested by "immunises" and "sensitivity ... nearly vanishes." Tables 16 and 18 measure changes in aggregate MAE against human labels, which can be small even when different students' grades move considerably. The paper should distinguish stable average accuracy from stable individual decisions and report paired score changes and their tails. The distinction is also important because the strict preamble changes grading policy rather than merely paraphrasing it.

  2. The statistical rule for human "parity" is too strong for the test performed. Section 8.2 and Tables 11/19 label a confidence interval containing zero as parity, then use this to support the claim that every pooled model is no worse than a human grader. A nonsignificant difference is not evidence of equivalence or noninferiority. Several breakdown-recipe intervals still permit a positive disagreement gap. The authors appropriately acknowledge that agreement with human labels is not grading correctness and improve on the original averaged-target comparison with Eq. (1). However, a justified noninferiority margin or narrower "no significant difference detected" wording is still needed. This is fixable and does not undermine the large base-to-tuned improvements or the significantly negative marks-recipe comparisons.

Questions for authors
  1. Can you report paired absolute changes in individual scores, including a high quantile, for the pooled models across personas? In particular, how do you interpret the 30B ML/exacting example above, and would you restrict the conclusion to robustness of aggregate error rather than near-invariance of grading decisions? This can be computed from the existing outputs.

  2. What practically meaningful noninferiority margin would you use for the third-grader comparison, and which cells satisfy it? Alternatively, would you revise the parity language to distinguish demonstrated lower disagreement from an inconclusive difference?

3 comments

Thank you for the careful review and for re-deriving our numbers. Both of your points were right and both changed the paper. New text is blue, reworded text is red.

1. Stable MAE is not the same as stable grades (W1, Q1). We did the analysis you asked for: for each held-out student, the score under a persona minus the score under the neutral prompt, same model (new Table 22).

CV exam (/35)ML exam (/65)
Fine-tuned (pooled), harsh personas: mean score change0.2 to 1.3 marks0.5 to 2.3 marks
Untuned base, harsh personas: mean score changeup to 21.5 marksup to 40.5 marks

Your 30B / ML / exacting example is correct. Scores drop by 2.24 marks on average, the bias goes from +0.46 to -1.78, and MAE only moves from 3.45 to 3.72.

So we now claim less: fine-tuning removes the collapse and keeps MAE stable, but a persona still nudges individual grades. The words "immunises" and "nearly vanishes" are gone.

2. "Parity" (W2, Q2). We removed the word. A confidence interval that contains zero is now called "no clear difference".

We also added a margin: a model is "no worse than a grader" only if the upper end of its interval is below 10% of the human disagreement (0.28 marks on the CV exam, 0.52 on the ML exam). In words: the model disagrees with a grader at most a tenth more than a second grader does. Results (new Table 20):

Result
Fine-tuned on human marks (10 cells)better than a grader in 9, no worse in 1
Fine-tuned with distilled breakdowns (10 cells)no worse in 6, inconclusive in 4
Zero-shot D01better than a grader
Zero-shot F02, D02no worse
Zero-shot B03, gpt-5.5 (both runs)inconclusive

The 10% is a convention, so the appendix repeats the verdicts at 5% and 20%. No "better" result depends on the margin. The distilled-breakdown adapter is no worse in 4, 6 and 10 of its ten cells at 5%, 10% and 20%. B03 becomes no worse at 20%; the gpt-5.5 runs stay inconclusive.

3. Ethics. The statement no longer says students cannot be identified. It says the data are pseudonymised and names the remaining risk (someone who knows the cohort might recognise a coding style). The notebooks sent to the external APIs were code cells only, with students already replaced by numbers.

The release also went through several rounds of manual checking and an automated scan of every notebook. Neither found any student identifier (names, emails, ID numbers, account names, personal file paths).

On consent and approval: students agreed to this use of their work when they signed up for the course, and the release was approved by the head of research ethics of the department where the courses took place. All students were undergraduates; none was a minor. The Ethics Statement now says this.

Thank you for the detailed response. The individual score-change analysis and revised robustness claims address my first concern. Replacing “parity” with margin-based comparisons and adding the grader-pair bootstrap also substantially address my statistical concern. I'm raising my rating to weak accept.

Thank you for reading the response and for raising your rating.

Reviewer_umJfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper evaluates LLMs for grading computer science exams. This is important but also well-studied. The paper argues that such real world assessments are rare, which is debatable as there are a good number of works on autograding using LLMs and even early LLMs such as Bert that date back to 2022 and before. Also the writing style is poor. The abstract provides lots of details without clearly saying what the finding is, sth like "the catch is the prompt" and "a short stricter grader" might be a good finding for a teacher's conference but the value for a conference like ICLR is not evident as presented by the author. Prompt engineering is known to have an impact on many tasks. A strong point is certainly that they evaluated many models, but using 2 graders and 2 exams is a weakness (which they acknowledge, but that does not make it go away). So in short, the contribution and other shortcomings make it not yet ready to publish at this venue. But the topic certainly has potential.

Critical strengths

see above

Critical weaknesses

see above

Questions for authors

None

3 comments

Thank you for the review. In the revised PDF, new text is blue and reworded text is red.

1. "LLM grading is well studied." You are right, and our wording suggested otherwise. We added a related-work paragraph on automated grading, from BERT-era short-answer grading to GPT-4, and now say plainly that LLM grading is not new. What is new here is narrower: long programming exams, every student graded by two humans, many prompts and models, and a second exam.

2. "The abstract does not say what the finding is." Rewritten, its first sentence is now the finding: an LLM grader that looks accurate can be broken by one small prompt edit. It then says that the same edit can help on another exam, and what fine-tuning fixes and what it costs.

3. "Prompt sensitivity is known; why ICLR?" Our point is not that prompts matter. It is three things:

  • The failure comes from one instruction (withhold partial credit), not from tone or model size.
  • The same instruction improves 7 of 17 models on the second exam, because it cancels their over-marking. So a prompt that "works" on one benchmark can be luck. We think this matters for anyone who uses an LLM as a judge.
  • Fine-tuning on about 3,900 human-graded examples removes the failure, but at a price. In a new experiment (Table 24) we gave the fine-tuned graders a legitimate instruction: deduct 2 marks as a late penalty, or add 2 as an approved adjustment. They applied 2% of the requested change (median). So the fine-tuned grader is robust because it largely ignores grading instructions in the prompt, good or bad. We think this trade-off matters for anyone who fine-tunes an LLM judge.

4. "Two exams, two graders." This is a real limitation and the paper says so. For context: the two exams cover 1,608 students and 69 graders, and the second exam is a different course with its own rubric and graders.

This is an improvement. Thanks. Also the number of students is clearly a strength of the paper compared to many other papers. But key concerns remain and simply cannot be easily eliminated. All insights are of some value, but tone, model size and framing of prompts are at this point all very well studied. Also the impact of fine-tuning (including through soft-prompts) is also well-known. That is not to say that there should not be more studies on this, but they really need to do more than early studies that could more or less say "Look if you add or remove this to the prompt something strange happens ...". The papers findings would need much more evidence from more domains and even within the education domain, relying on 2 exams is not very convincing.

Thank you for the follow-up and for acknowledging the improvement.

We agree that tone, model size and fine-tuning are well studied. Our claims: the failure comes from one grading-policy instruction rather than tone or size, the same instruction helps on the second exam by cancelling over-marking, and fine-tuning removes the failure only by largely ignoring grading instructions, including legitimate ones (Table 24).

We agree that two exams is a limitation, and the paper states it. However, to our knowledge this is the largest public dataset of full programming exams in which every student is graded by two humans: 1,608 students and 69 graders. The closest public graded dataset we know of, Menagerie, has fewer than 300 graded submissions from a single assignment. We see this as a substantial step towards closing the gap, and we agree that it can be extended further to more exams and courses.

Comments on the paper

1 comment

We thank the three reviewers. In the revised PDF, new content is blue and reworded text is red. This note lists what changed.

Claims we now make more carefully

  • "Parity" is gone. An interval that contains zero is called "no clear difference", and a model is "no worse than a grader" only inside a stated margin
  • "Immunises" and "nearly vanishes" are gone. Fine-tuning removes the collapse and keeps MAE stable, but a persona still moves individual scores by 0.2 to 2.3 marks on average
  • The human-level results are for new students of an exam seen in training, not for a new exam

New analyses, all asked for by the reviewers

  • Table 20: the third-grader test with a margin and with a grader-pair bootstrap.
  • Table 21: the same test on an exam never seen in training
  • Table 22: per-student score changes under each persona
  • Table 23: personas on an exam never seen in training
  • Table 24: a controlled test of legitimate policy changes

One new finding Fine-tuned graders largely ignore grading instructions in the prompt, legitimate ones included. Asked to deduct 2 marks as a late penalty, or to add 2 as an approved adjustment, they apply 2% of the requested change (median). So fine-tuning buys robustness by making the grader insensitive to the prompt. The paper states this as a limitation and recommends applying policy changes outside the model.

Other changes

  • Abstract rewritten and cut by a third; it now opens with the finding.
  • Related work on automated grading added.
  • Ethics Statement rewritten: the data are pseudonymised, not anonymous, and the statement gives the consent basis, the approval and the re-identification check.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.