arXiv AI

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

The study evaluates large language model (LLM) graders on two computer‑science exams, testing 171 configurations of closed‑ and open‑weights models. While the best LLM configuration achieved a mean absolute error of 1.64/35—better than the 2.61/35 error between two human graders—its performance was highly sensitive to the prompt. A short "strict grader" preamble caused most open‑weight models to exceed acceptable error thresholds or stop grading entirely, whereas fine‑tuning with a single LoRA adapter restored parity with human graders and reduced sensitivity to harsh prompts.

arXiv AI
Aug 19

Grading Needs a Rubric, Not Intelligence

Small language models can grade open‑ended exam answers as reliably as much larger models when they use an explicit rubric. In experiments with six cost‑efficient model configurations, the rubric decouples grading from judge intelligence, with answer identity explaining 95.6% of score variance and judge identity only 0.2%. Removing rubric criteria or the official answer collapses reliability and inflates scores, showing the rubric’s essential role.

By Jhen-Ke Lin
arXiv AI
Aug 20

Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme

The paper introduces THPT‑Ladder, a benchmark based on Vietnam’s 2025 National High School Graduation Examination’s convex grading scheme, which rewards partial credit non‑additively. It shows that standard accuracy metrics inflate model scores because they treat partial knowledge proportionally, whereas the official rubric penalizes incomplete correct sets. Using the benchmark, the authors demonstrate that this discrepancy can shift a model’s percentile ranking by up to 13 points among over 480,000 candidates.

By Nguyen Quoc Hung, Nguyen Dang Minh, Le Nhu Quynh, Tran Khanh Linh, Nguyen Kieu Linh
arXiv Machine Learning
Sep 18

How Far Can Sub-3B Open Language Models Go in Zero-Shot Essay Scoring on an 8 GB Consumer GPU?

The study evaluates zero‑shot essay scoring using sub‑3B open language models that run locally on a single 8 GB consumer GPU. Four instruction‑tuned models (Qwen2.5‑0.5B, 1.5B, 3B and SmolLM2‑1.7B) were tested on all eight ASAP‑AES prompts, comparing rubric‑decomposed versus holistic prompting, different aggregation methods, and trait‑mapping strategies. Results show rubric‑decomposed prompting consistently outperforms holistic prompting, trait‑mapping is sensitive to calibration, and longer essays reduce error, yet the best local configuration (macro QWK 0.388) still falls short of human agreement and a length‑only baseline, suggesting these models are best suited for formative, human‑supervised feedback.

By Nguyen Dung Son, Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, Nguyen Thai Anh
arXiv Computation and Language
Nov 4

Same Outputs, Different Scores: Auditing the Unchecked Graders of Multimodal LLM Benchmarks

The paper investigates how the scoring rules used in multimodal large‑language‑model (LLM) benchmarks can dramatically alter reported accuracy. By reproducing official scorers for eight benchmarks and comparing them to a blinded human‑plus‑LLM adjudication, the authors show that the same model outputs can yield accuracy differences of up to 22 points depending on the rule. They argue that no single scoring rule is reliable and recommend that benchmarks report non‑compliance rates and use at least two distinct scoring rules.

By Ahmad Jarjis Hasan, Jonathan Lwowski
arXiv AI
Jul 17

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

arXiv:2607. 14552v1 Announce Type: cross Abstract: A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors.

By Jungseob Lee, Seungyoon Lee, Suhyune Son, Dongyub Jude Lee, Sungbin Han, Sugyeong Eo, Heuiseok Lim
arXiv Machine Learning
1d ago

How Much Is Left Matters: False Positives and False Negatives in Partial-Prompt Contamination Tests

The paper evaluates partial‑prompt completion tests—where a model is given the first part of a benchmark problem and checked for reproducing the rest—on seven open‑weight models using MATH‑500 and recent AIME 2025/2026 problems as true negatives. It finds that memorization flags are highly sensitive to the amount of text left: problems with fewer than eight remaining words are flagged far more often than those with 25 or more, and most ROUGE‑based flags arise from short leftovers. On true negatives, even a high ROUGE‑L threshold flags up to 6.7% of problems at a 60% prefix, and exact‑match flags 8.6% at an 80% prefix, while models distilled from Qwen2.5‑Math show almost no recall of contamination. The authors recommend length‑stratified reporting, limiting prefixes to 60%, using style‑matched post‑cutoff controls, and cross‑family calibration to reduce false positives.

By Abhishek Anand
arXiv AI
Oct 12

A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning

The paper examines on‑policy self‑distillation, where a teacher scores a student’s rollout to provide dense token‑level credit. Experiments show that the raw scoring signals neither localize errors nor effectively penalize incorrect tokens, because locating errors is a search problem that the teacher is not explicitly asked to perform. Introducing a critique‑conditioned approach—first generating a critique of the rollout and then conditioning the teacher on it—produces signals that accurately localize errors, correct erroneous spans, and guide the policy toward backtracking tokens, improving performance on math problems and in cold‑start GRPO settings.

By Mayug Maniparambil, Ayush Rai, Ekansh Verma, Noel E. O'Connor, Fergal Reid