Large language models (LLMs) are widely used as automatic judges, with validity typically assessed via alignment with human scores. However, aggregate agreement fails to reveal whether humans and LLMs find the same evaluation cases difficult. In this paper, we study this problem in summarization evaluation from a psychometric perspective. We fit Many-Facet Rasch Models separately to human and LLM ratings to decompose scores into latent summary quality, rater severity, dimension severity, and rating-scale thresholds. Building on this decomposition, we define residual hardness as a model-adjusted measure of judging difficulty and compare whether human and LLM judges share the same hardness structure.

Across 17 open-weight LLM judges on SummEval, we find that moderate alignment in latent summary quality does not imply alignment in residual hardness. Human and LLM judges differ in which summary–dimension units remain difficult, and this mismatch is strongly dimension-dependent. Consistency shows a pronounced LLM-hard shift, whereas coherence shows a human-hard shift. We further show that human-easy but LLM-hard cases are partially predictable from observable source–summary properties. These findings suggest that aggregate human alignment reflects only part of LLM-as-judge reliability, while psychometric residual diagnostics support more informative judge evaluation and more targeted human–LLM collaboration.

Automatic Short Answer Scoring (ASAS) is central to NLP for Education. However, openly available benchmarks remain scarce, and existing datasets largely address how well students answer a question directly rather than how well they master underlying concepts (knowledge elements) such as thermal energy or epistemic activities (skills) such as reasoning or claim.

To address this gap, we introduce Alice, a large-scale, rubric-based German ASAS dataset that is pedagogically aligned and comprises three subtasks: (i) learning performance (Alice-LP), (ii) knowledge elements (Alice-KE), and (iii) skills (Alice-SK).

We further formulate rubric-based ASAS as a rubric-retrieval task and benchmark the dataset with a range of language models, from encoder-only models to lightweight LLMs. We also benchmark the dataset with zero-shot prompting via LLMs and a standard classification baseline. The experiments show that LLMs, in particular, struggle to score knowledge elements and skills in the zero-shot setting. They also indicate that rubric text is often useful, especially for Alice-KE and Alice-SK, while on Alice-LP gains over sample-solution-focused inputs are more modest and vary by model and input format.

Automatic Short Answer Scoring (ASAS) requires models that can score student responses against question-specific criteria while remaining efficient and transferable across rubric sets. We propose RUSPAN, a rubric-conditioned ASAS framework that treats rubric descriptions as semantic label representations. RUSPAN serialises the question context, student answer, and all candidate rubric levels into a single sequence, then scores the levels listwise from the rubric-span and whole-sequence representations produced in a single LM pass. We further introduce RUSPAN-RIM, in which a Rubric-Independent Mask prevents rubric spans from attending to one another, making rubric representations depend only on the answer and question context and preventing overfitting to rubric patterns during training for zero-shot transfer. On six ASAS benchmarks spanning English, German, and Portuguese, RUSPAN improves mono-benchmark scoring over discriminative and generative baselines, while RIM with position reindexing delivers consistent and substantial gains on PT-ASAG, the held-out benchmark with the strongest combined language and rubric-structure shift.

We present the BEA 2026 shared task on rubric-based short answer scoring for German. Rubric-based short answer scoring is a case of automatic short answer scoring (ASAS) that requires models to apply textual scoring rubrics to student answers as a basis for assigning scores. For the shared task, we introduced a novel German-language dataset from multiple STEM domains to provide a comprehensive benchmark for this problem. The dataset was designed to evaluate both performance and generalization (the latter, by distinguishing between seen and unseen questions), as well as coarse- and fine-grained scoring (2-way vs. 3-way). The systems submitted to the shared task cover a wide range of approaches, including fine-tuned large language models, prompt-based methods, human-AI collaboration strategies, or a combination of these. The results show that structured, task-adapted LLM systems achieved the strongest performance across all tracks. The winning system, IWM-DKM, combined LoRA fine-tuning of Qwen models with rubric-aware input structuring, including checklist-style reasoning, rubric reframing as decision trees, background knowledge injection, and ensemble voting. Other systems similarly relied on fine-tuned LLMs, retrieval-augmented prompting, encoder–LLM ensembles, or weighted aggregation strategies. Overall, the shared task results show that rubric-based scoring benefits most from systems that explicitly operationalise rubric semantics, while generalisation to unseen questions remains a central challenge.

Rubrics are the primary reference for manual scoring of constructed responses, and there is growing interest in their use in automated scoring methodologies. In this work, we propose Aspect-Grounded Rubric–Answer Alignment (AGRAA), a rubric-based end-to-end scoring framework that models rubric descriptors as latent aspect spaces. Concretely, rubric descriptors are represented as low-dimensional subspaces derived from contextualised transformer embeddings, and student responses are scored according to how strongly their representations align with these rubric-induced spaces relative to the residual space outside them. This formulation provides a geometrically grounded interpretation of rubric-based scoring while enabling end-to-end training with standard transformer encoders. We introduce three distinct architectural variants and evaluate them on multiple short-answer and essay scoring datasets. Across these tasks, AGRAA achieves predictive performance highly competitive with strong neural and feature-based baselines. In addition, the framework yields interpretable intermediate representations that expose which rubric-defined aspects contribute to scoring decisions, enabling decision-aligned explanations grounded in rubric descriptors.

Automated short answer grading (ASAG) with large language models (LLMs) is commonly evaluated with aggregate metrics such as macro-F1 and Cohen’s kappa. However, these metrics provide limited insight into how grading performance varies across student responses of differing grading difficulty. We introduce an evaluation framework for LLM-based ASAG based on item response theory (IRT), which models grading correctness as a function of latent grader ability and response grading difficulty. This formulation enables response-level analysis of where LLM graders succeed or fail and reveals robustness differences that are not visible from aggregate scores alone. We apply the framework to 17 open-weight LLMs on the SciEntsBank and Beetle benchmarks. The results show that even models with similar overall performance differ substantially in how sharply their grading accuracy declines as response difficulty increases. In addition, confusion patterns show that errors on difficult responses concentrate disproportionately on the partially_correct_incomplete label, indicating a tendency toward intermediate-label collapse under ambiguity. To characterize difficult responses, we further analyze semantic and linguistic correlates of estimated difficulty. Across both datasets, higher difficulty is associated with weaker semantic alignment to the reference answer, stronger contradiction signals, and greater semantic isolation in embedding space. Overall, these results show that item response theory offers a useful framework for evaluating LLM-based ASAG beyond aggregate performance measures.

How students use generative AI is a configuration of distinct behaviors, not one skill, yet scoring typically collapses it onto a single proficiency axis, conflating process with product. We develop a multidimensional Bayesian IRT model with rater effects and post-hoc Varimax-permutation identification, surfacing three substantive dimensions: prompting effort, content delegation, AI-delivered citations. Domain familiarity shifts students toward more active engagement and away from delegation.

This paper presents our contribution to the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-Powered Tutors. The objective of this shared task was to assess the quality of conversational feedback provided by LLM-based math tutors to students regarding four facets: whether the tutors 1) identified mistakes, 2) identified the mistake’s location, 3) provided guidance, and whether they 4) provided actionable feedback. To leverage information across all four labels, we approached the problem with FLAN-T5 models, which we fit for this task using a multi-step pipeline involving regular fine-tuning as well as model merging using the DARE-TIES algorithm. We can demonstrate that our pipeline is beneficial to overall model performance compared to regular fine-tuning. With results on the test set ranging from 52.1 to 68.6 in F1 scores and 62.2% to 87.4% in accuracy, our best models placed 11th of 44 teams in Track 1, 8th of 31 teams in Track 2, 11th of 35 teams in Track 3, and 9th of 30 teams in Track 4. Notably, the classifiers’ recall was relatively poor for underrepresented classes, indicating even greater potential for the employed methodology.

This paper describes a contribution to the BEA 2024 Shared Task on Automated Prediction of Item Difficulty and Response Time. The participants in this shared task are to develop models for predicting the difficulty and response time of multiple-choice items in the medical field. These items were taken from the United States Medical Licensing Examination® (USMLE®), a high-stakes medical exam. For this purpose, we evaluated multiple BERT-like pre-trained transformer encoder models, which we combined with Scalar Mixing and two custom 2-layer classification heads using learnable Rational Activations as an activation function, each for predicting one of the two variables of interest in a multi-task setup. Our best models placed first out of 43 for predicting item difficulty and fifth out of 34 for predicting Item Response Time.