---
title: 'TutorBench: LLM Tutoring Skills Benchmark'
url: https://www.emergentmind.com/topics/tutorbench
type: topic
---

# TutorBench: LLM Tutoring Skills Benchmark

Searching arXiv for TutorBench and closely related benchmarking papers.
TutorBench is a dataset and evaluation benchmark designed to rigorously evaluate the core tutoring skills of large language models (LLMs), rather than treating tutoring as ordinary answer generation. It comprises 1,490 samples curated by human experts, focuses on high-school and AP-level curricula, and organizes evaluation around three common tutoring tasks: adaptive explanation generation, assessment and feedback, and active learning support. Its central methodological choice is to score model responses against sample-specific rubrics using an LLM-judge, thereby targeting tutoring behaviors such as adaptivity, actionable feedback, and hinting without answer leakage [2510.02663].

## 1. Scope, curriculum, and benchmark rationale

TutorBench was introduced from the premise that tutoring requires a different skill set from ordinary reasoning benchmarks. In the benchmark’s framing, a good AI tutor must identify the core needs of students, be adaptive, provide personalized guidance, and be accurate. Existing evaluations were described as too narrow, often focusing on one subject, text-only interactions, or final-answer correctness rather than the human-centered behaviors that make tutoring effective. TutorBench was therefore constructed as a broader, more realistic, and more fine-grained evaluation of tutoring capability [2510.02663].

The benchmark targets high-school and AP-level STEM curricula across six subjects: Biology, Physics, Chemistry, Statistics, Calculus, and Computer Science. It is multimodal: 828 samples include images of handwritten work, typed work, diagrams, or screenshots. This multimodal design matters because several tutoring scenarios require interpretation of student artifacts rather than only textual prompts, and the benchmark is intended to evaluate how models handle both text and visual input [2510.02663].

## 2. Task structure and sample composition

TutorBench organizes its samples around three tutoring tasks, each corresponding to a distinct instructional function.

| Task | Core requirement | Educational role |
|---|---|---|
| Adaptive explanation generation | Tailor an explanation to a student’s follow-up confusion | Diagnose and explain |
| Assessment and feedback | Evaluate work, identify mistakes, classify error type, explain improvement | Critique and correct |
| Active learning support | Give hints or scaffolding without directly revealing the answer | Preserve student agency |

Adaptive explanation generation requires the model to respond to a student’s follow-up confusion by tailoring the explanation to that student’s knowledge gap. Assessment and feedback requires the model to evaluate a student’s solution, identify mistakes, classify the type of error, and explain how to improve. Active learning support requires the model to provide hints or scaffolding that help the student take the next step without directly revealing the answer. This task design makes the benchmark explicitly pedagogical: it does not only ask whether a model knows the material, but whether it can support learning in ways appropriate to the student’s situation [2510.02663].

A further design choice is that each sample is accompanied by sample-specific rubrics written by human subject experts. These rubrics are intended to be self-contained, mutually exclusive, and verifiable, and each example contains between 3 and 39 criteria. Across the full benchmark there are 15,220 rubric criteria. Annotators had at least a bachelor’s degree and tutoring or professional experience in the subject they were writing for, and they also wrote the prompt items and a “golden tutoring response” that informed the rubric criteria [2510.02663].

## 3. Rubric design, weighting, and automatic judging

TutorBench uses an automatic LLM-judge rather than requiring human grading for every model release. The judge is Claude Sonnet 4, selected because it aligned best with human ratings. It evaluates a model response criterion-by-criterion using the sample-specific rubric and assigns each criterion a binary pass/fail rating [2510.02663].

Rubric aggregation uses a weighted scheme. Critical rubrics receive weight $+5$, ordinary rubrics receive weight $1$, and some critical rubrics can receive $-5$ to penalize undesirable behavior, such as directly revealing the final answer in an active-learning setting. The per-example score is the weighted average rubric rating $ARR^j$:
$$
ARR^j = \frac{\sum_{i=1}^{N_j} w_i^j \cdot \mathbbm{1}_{r_i^j}}{\sum_{i=1}^{N_j} w_i^j \cdot \mathbbm{1}_{w_i^j > 0}}
$$
where $N_j$ is the number of rubric criteria for example $j$, $w_i^j \in \{-5,1,5\}$ is the rubric weight, and $r_i^j \in \{0,1\}$ is the fail/pass label for criterion $i$. The overall model score is the average $ARR$ across all examples. In effect, the benchmark asks how many required tutoring behaviors were satisfied, and how important those behaviors were, rather than only whether a response was “correct” in a conventional QA sense [2510.02663].

The judge was also validated against expert ratings. On a 250-sample subset, three human expert ratings per rubric criterion were collected, totaling 2,475 rubric criteria. The judge achieved mean agreement with human ratings of 0.78, mean inter-human agreement of 0.75, and $F1 = 0.82$ against majority vote after filtering out critical criteria. These results were presented as evidence that the automatic evaluation is reasonably faithful to expert judgment [2510.02663].

## 4. Empirical performance and reported model behavior

TutorBench is reported as difficult and unsaturated. No frontier model scores above 56% overall. On the reported leaderboard, Gemini 2.5 Pro is best overall at 55.65%, followed by GPT-5 at 55.33%; other reported scores include o3 Pro at 54.62%, o3 Medium Effort at 52.76%, Claude Opus 4.1 (Thinking) at 50.78%, and GPT-4o at 36.12%. For text-only-only models reported separately, gpt-oss-120b reaches 56.01% on text-only, gpt-oss-20b reaches 49.01%, and DeepSeek-R1 reaches 48.38% [2510.02663].

Performance also varies by tutoring task. Average scores across the three use cases are 47.16% for adaptive explanation generation, 51.56% for assessment and feedback, and 54.07% for active learning support. Adaptive explanation is therefore the hardest of the three benchmarked skills. The paper also reports a family-level pattern: Claude models do relatively better on active learning support than other model families, but lag in overall performance and in the other two use cases. The authors treat this as a substantive behavioral difference rather than a leaderboard artifact [2510.02663].

The benchmark’s interpretation is that frontier LLMs still fall short in exhibiting the full range of tutoring skills needed to guide, diagnose, and support students effectively, with all the frontier models achieving less than a 60% pass rate on rubric criteria related to these skills. A related study, using public TutorBench artifacts as part of a broader measurement argument, likewise concluded that educational-impact evaluation should not treat task success as a sufficient proxy for learning support [2606.16206].

## 5. Evaluation dimensions, tutoring skills, and learner agency

TutorBench measures tutoring quality not only through task labels but also through finer-grained rubric tags. Rubrics are tagged with evaluation dimensions including instruction following, style and tone, truthfulness, visual reasoning, visual perception, conciseness and relevance, student-level calibration, and emotional component. Within these dimensions, Gemini 2.5 Pro is reported as especially strong at recognizing student emotions and adjusting tone/style, while GPT-5 and o3 Pro are strongest on truthfulness, student-level calibration, and instruction following [2510.02663].

The benchmark also aggregates rubrics into eight high-level tutoring skills: identifying core difficulty, identifying students’ correct steps, identifying students’ incorrect steps, recalling and stating knowledge, providing alternative solutions, including examples, asking questions to guide students, and providing step-by-step help. Models do best on identifying correct steps at 53.7% and identifying incorrect steps at 53.0%, but they are weaker on alternative solutions at 41.9% and examples/analogies at 32.8%. This reported pattern indicates stronger performance on diagnosis than on richer pedagogical scaffolding [2510.02663].

A diagnostic analysis of the public TutorBench sample sharpened this interpretation by examining whether the rubrics themselves encode agency-preserving tutoring behavior. That sample contains 30 scenarios and 324 rubric items across adaptive explanation, assessment/feedback, and active learning. The analysis maps rubric metadata deterministically into agency-dimension items and agency-skill items, and reports the following rubric densities: adaptive explanation has 108 rubrics with agency-dimension density 0.769 and agency-skill density 0.139; assessment/feedback has 109 rubrics with densities 0.954 and 0.294; active learning has 107 rubrics with densities 1.000 and 0.411. Active learning therefore contains the strongest concentration of agency-preserving criteria. One representative criterion requires the tutor to “provide hint[s] ... by asking at least 2 guiding questions without completely showing next steps,” making explicit that the benchmark rewards guiding questions, calibrated hints, and non-disclosive scaffolding rather than mere answer delivery [2606.16206].

## 6. Limitations, related benchmarks, and bibliographic clarification

TutorBench is broad compared with prior work, but its scope remains bounded. The paper explicitly notes that it covers only three tutoring scenarios, evaluates final responses rather than fully dynamic multi-turn tutoring adaptation, does not test image generation or broader UI/UX interaction, focuses only on STEM, and does not include some useful tutor tasks such as generating practice problems or introducing new concepts. These limitations mean that TutorBench is a benchmark for core tutoring skills under a specific response-based evaluation setting, not a complete measure of all tutoring activity [2510.02663].

Within the broader literature, TutorBench sits alongside a growing family of pedagogy-aware tutoring benchmarks. MathTutorBench is an open-source benchmark for holistic tutoring model evaluation in mathematics; it contains seven tasks spanning expertise, student understanding, and pedagogical ability, and reports that subject expertise, indicated by solving ability, does not immediately translate to good teaching [2502.18940]. SafeTutors extends the evaluation problem to joint pedagogy and safety across mathematics, physics, and chemistry, and argues that tutoring harm is often a quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding [2603.17373]. These adjacent benchmarks help situate TutorBench as part of a broader shift away from final-answer evaluation and toward measurement of adaptive, pedagogically grounded instructional behavior.

A separate bibliographic clarification is necessary because one cited arXiv record can be misread as a TutorBench source. The provided document for arXiv:2302.12623 is not a TutorBench paper or a tutoring-systems paper; it is the AAAI 2023 LaTeX author-kit/instructions template. It contains no tutoring task, no tutor-student dialogue model, no instruction grounding method, no multi-task learning objective, no dataset, no evaluation, and no benchmark result to extract. Accordingly, it has no direct bearing on TutorBench beyond illustrating the need for careful source verification in benchmark surveys [2302.12623].

Source: https://www.emergentmind.com/topics/tutorbench