Paper deep dive

EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content

Shuzhen Bi, Mingzi Zhang, Zhuoxuan Li, Xiaolong Wang, keqian Li, Aimin Zhou

Year: 2026Venue: arXiv preprintArea: cs.CYType: PreprintEmbeddings: 72

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%

Last extracted: 4/10/2026, 2:51:57 AM

Summary

EduIllustrate is a new benchmark designed to evaluate Large Language Models (LLMs) on their ability to generate multimodal educational content, specifically interleaved text-diagram explanations for K-12 STEM problems. The benchmark includes 230 problems, a standardized generation protocol using sequential anchoring for visual consistency, and an 8-dimension evaluation rubric. Evaluation of ten LLMs shows Gemini 3.0 Pro Preview as the top performer, while the study validates the effectiveness of sequential anchoring in improving visual consistency and reducing costs compared to parallel generation workflows.

Entities (5)

EduIllustrate · benchmark · 100%Gemini 3.0 Pro Preview · large-language-model · 100%Kimi-K2.5 · large-language-model · 100%Sequential Anchoring · methodology · 100%Manim · software-library · 95%

Relation Signals (3)

EduIllustrate → evaluates → Large Language Models

confidence 100% · a benchmark for evaluating LLMs on interleaved text-diagram explanation generation

Sequential Anchoring → improves → Visual Consistency

confidence 100% · Workflow ablation confirms sequential anchoring improves Visual Consistency by 13%

Gemini 3.0 Pro Preview → performedbeston → EduIllustrate

confidence 100% · Gemini 3.0 Pro Preview leads at 87.8%

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language models are increasingly used as educational assistants, yet evaluation of their educational capabilities remains concentrated on question-answering and tutoring tasks. A critical gap exists for multimedia instructional content generation -- the ability to produce coherent, diagram-rich explanations that combine geometrically accurate visuals with step-by-step reasoning. We present EduIllustrate, a benchmark for evaluating LLMs on interleaved text-diagram explanation generation for K-12 STEM problems. The benchmark comprises 230 problems spanning five subjects and three grade levels, a standardized generation protocol with sequential anchoring to enforce cross-diagram visual consistency, and an 8-dimension evaluation rubric grounded in multimedia learning theory covering both text and visual quality. Evaluation of ten LLMs reveals a wide performance spread: Gemini 3.0 Pro Preview leads at 87.8\%, while Kimi-K2.5 achieves the best cost-efficiency (80.8\% at \\$0.12/problem). Workflow ablation confirms sequential anchoring improves Visual Consistency by 13\% at 94\% lower cost. Human evaluation with 20 expert raters validates LLM-as-judge reliability for objective dimensions ($\rho \geq 0.83$) while revealing limitations on subjective visual assessment.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

71,676 characters extracted from source content.

Expand or collapse full text

EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content Shuzhen Bi1,2, Mingzi Zhang3, Zhuoxuan Li3, Xiaolong Wang3, Keqian Li3, Aimin Zhou1,3 1Shanghai Innovation Institute, 2University of Science and Technology of China, 3East China Normal University sa22916003@mail.ustc.edu.cn, 51284102005, lizhuoxuan, wmumu@stu.ecnu.edu.cn, kqli@mail.ecnu.edu.cn, amzhou@cs.ecnu.cn Corresponding author. Abstract Large language models are increasingly used as educational assistants, yet evaluation of their educational capabilities remains concentrated on question-answering and tutoring tasks. A critical gap exists for multimedia instructional content generation—the ability to produce coherent, diagram-rich explanations that combine geometrically accurate visuals with step-by-step reasoning. We present EduIllustrate, a benchmark for evaluating LLMs on interleaved text-diagram explanation generation for K-12 STEM problems. The benchmark comprises 230 problems spanning five subjects and three grade levels, a standardized generation protocol with sequential anchoring to enforce cross-diagram visual consistency, and an 8-dimension evaluation rubric grounded in multimedia learning theory covering both text and visual quality. Evaluation of ten LLMs reveals a wide performance spread: Gemini 3.0 Pro Preview leads at 87.8%, while Kimi-K2.5 achieves the best cost-efficiency (80.8% at $0.12/problem). Workflow ablation confirms sequential anchoring improves Visual Consistency by 13% at 94% lower cost. Human evaluation with 20 expert raters validates LLM-as-judge reliability for objective dimensions (ρ≥0.83ρ≥ 0.83) while revealing limitations on subjective visual assessment. EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content Shuzhen Bi1,2, Mingzi Zhang3, Zhuoxuan Li3, Xiaolong Wang3, Keqian Li3†thanks: Corresponding author., Aimin Zhou1,3 1Shanghai Innovation Institute, 2University of Science and Technology of China, 3East China Normal University sa22916003@mail.ustc.edu.cn, 51284102005, lizhuoxuan, wmumu@stu.ecnu.edu.cn, kqli@mail.ecnu.edu.cn, amzhou@cs.ecnu.cn Figure 1: EduIllustrate generates geometrically accurate visuals for diverse STEM problems. Taking textual K-12 problems from various subjects as input (left), our system produces a progressive sequence of diagrams (right). These diagrams are interleaved with textual reasoning to form multimodal explanations. 1 Introduction Figure 2: Overview of EduIllustrate. Given a K-12 STEM problem, the generation protocol produces a structured explanation document through four stages. The resulting output is evaluated using an 8-dimension rubric validated against human expert raters. With the rapid development of large language models (LLMs), education has emerged as one of the most important and widely adopted application domains. LLMs are already used by millions of learners as on-demand educational assistants (Kamalov et al., 2025), and major providers are actively building tutoring-oriented systems (Wang et al., 2025; Liu et al., 2025). However, prior evaluation of LLMs’ educational capabilities has focused mainly on traditional question-answering and tutoring settings, leaving an important class of real educational tasks underexplored: the generation of multimedia instructional content. Decades of research in Cognitive Load Theory (Sweller, 1988), Dual Coding Theory (Paivio, 1971), and the Multimedia Learning Principle (Mayer, 2002) have established that coordinating visual diagrams with textual reasoning substantially reduces cognitive burden and deepens conceptual understanding. Yet producing such materials at scale remains a formidable challenge—lesson planning and material preparation rank among the most time-consuming aspects of teachers’ professional work (Philipp, 2007; Thompson and Dahlin, 2024), and many educators lack the specialized skills required to create geometrically accurate diagrams (Türközü and Dinçer, 2025; Yao and others, 2026). If LLMs could reliably generate illustrated explanations, the impact on K-12 education would be substantial. But we currently lack the benchmarks to measure whether they can. Existing work falls short in several ways. On the education side, content generation systems bifurcate into text-only approaches (Liu et al., 2025; Wang et al., 2025) and video-based systems targeting university-level theorems (Ku et al., 2025; Chen et al., 2025). On the multimodal generation side, systems such as ANOLE (Chern et al., 2024) and Orthus (Kou et al., 2025) target general-purpose domains and generate photorealistic images, which cannot satisfy the geometric precision required for educational diagrams. Evaluation benchmarks like OpenLEAF (An et al., 2024) and MMIE (Xia et al., 2025) assess interleaved generation but not in educational contexts, while DiagramIR (Kumar et al., 2025) evaluates mathematical diagrams in isolation and EduVisBench (Ji et al., 2025) targets visual reasoning rather than generation quality. No existing benchmark jointly assesses both textual and visual quality of K-12 multimodal educational content. To address this gap, we present EduIllustrate (Figure 2), a benchmark designed to evaluate LLMs on interleaved text-diagram explanation generation for K-12 STEM problems—a setting that more closely reflects real-world educational applications. The benchmark comprises three components: (1) a curated problem set of 230 problems spanning five subjects and three grade levels; (2) a standardized generation protocol with sequential anchoring to ensure cross-diagram visual consistency; and (3) an 8-dimension evaluation rubric grounded in multimedia learning theory that jointly covers textual and visual quality. Our main contributions are threefold: (1) EduIllustrateBench, comprising 230 curated problems across five subjects and three grade levels, with an 8-dimension evaluation rubric addressing the gap in multi-subject, multi-grade multimodal educational content evaluation; (2) a standardized generation protocol with sequential anchoring to ensure cross-diagram visual consistency, serving as both the generation method and an ablation baseline; and (3) a comprehensive empirical study of ten LLMs spanning proprietary and open-weight models, with ablation studies and human evaluation validating LLM-as-judge reliability. 2 Related Work 2.1 LLMs in Education LLM applications in education span intelligent tutoring, automated assessment, and adaptive content generation. Kamalov et al. (2025) reviewed agentic workflows in education, highlighting advancements in automated tutoring while noting constraints in adaptability that multi-agent frameworks address. Wang et al. (2025) introduced GenMentor, an LLM-powered multi-agent framework for goal-oriented learning, demonstrating effective skill gap identification and personalized learning path scheduling. For K-12-specific applications, Liu et al. (2025) proposed COGENT, a curriculum-oriented framework generating grade-appropriate content aligned with curriculum standards. Despite these advances, existing work predominantly focuses on text-only explanations or assessment, leaving multimodal content generation for K-12 STEM underexplored. 2.2 Programmatic Educational Content Generation Recent work has explored generating educational content through executable code. Ku et al. (2025) introduce an agentic approach for generating long-form theorem explanation videos using Manim animations, targeting university-level mathematics and physics. Chen et al. (2025) propose a code-centric agent framework for generating professional educational videos via executable Python code. Both systems target video-based explanations for advanced topics. Our work differs by focusing on static diagram generation for K-12 problem-solving explanations, where textbook-style illustrations enable self-paced learning across five subjects with subject-specific diagram conventions. 2.3 Programmatic Diagram Generation Programmatic diagram generation leverages tools like TikZ, Manim, and Matplotlib. Kumar et al. (2025) introduced DiagramIR, an automatic evaluation pipeline for educational math diagrams using intermediate representations of LaTeX TikZ code, demonstrating higher agreement with human raters than LLM-as-judge baselines. Cui et al. (2025) proposed Draw with Thought, a training-free framework guiding multimodal LLMs to reconstruct scientific diagrams into editable mxGraph XML code through Chain-of-Thought reasoning. 2.4 LLM-as-Judge Evaluation LLMs as evaluators provide practical alternatives to costly human evaluation. Enguehard et al. (2025) revealed in the legal domain that reference-free evaluation protocols correlate better with human expert judgments. Park and Yang (2025) demonstrated AGACCI framework improves accuracy through distributing specialized evaluation roles across collaborative agents. 3 EduIllustrate Benchmark 3.1 Task Formulation Given a K-12 STEM problem (text and an optional diagram), a model must produce an interleaved explanation: a sequence of textual reasoning steps alternating with programmatically rendered diagrams. This task jointly demands (i) mathematically correct and pedagogically coherent text, (i) geometrically accurate diagrams faithful to the problem setup, and (i) visual consistency across multiple diagrams within the same explanation. Failure in any single aspect undermines the explanation’s educational value, making this a challenging multimodal generation task. 3.2 Problem Set We curate 230 problems from K12-Vista (Li et al., 2025), spanning three grade levels (elementary, middle, high school) and five STEM subjects (Table 1, Figure 3). Problems are selected for diagram appropriateness, solution clarity, and topic diversity; full curation details are provided in Appendix A. Grade Math Phys Chem Bio Geo Total Elementary 20 — — — — 20 Middle 30 30 15 15 15 105 High 30 30 15 15 15 105 Total 80 60 30 30 30 230 Table 1: Benchmark distribution across grade levels and subjects. Figure 3: Topic categories by subject. Values in parentheses indicate the number of unique topics mapped in each category. 3.3 Generation Protocol To ensure fair comparison, all models generate explanations through a standardized four-stage protocol (Figure 4). The protocol is designed to decouple the generation task into manageable subtasks while enforcing cross-diagram visual consistency. Figure 4: Generation protocol overview. Scene 1 undergoes full sequential processing (outline → implementation planning → code generation); Scenes 2–N skip implementation planning and generate in parallel, each conditioned on Scene 1’s code as a visual anchor. Stage 1: Structured Outline. The model transforms a problem into an XML-based outline with alternating <TEXT_k> (pedagogical explanation) and <SCENE_k> (visual specification) blocks. Scene blocks are generated only when diagrams genuinely aid understanding. This structured format enables modular processing and error recovery. Stage 2: Implementation Planning (Scene 1 only). Planning every scene independently would cause each to invent its own visual conventions, making consistency impossible to guarantee. We therefore restrict planning to Scene 1 only, converting its specification into a detailed implementation plan: intended appearance, spatial constraints, and discipline-specific rendering conventions (e.g., right-angle markers in geometry, vector arrows in physics). The resulting conventions—color scheme, labeling style, line weights—are inherited by all subsequent scenes. Stage 3: Code Generation and Rendering. Scene 1 is rendered first from its plan; its complete Manim code then serves as context for Scenes 2–N, which generate in parallel. This enforces visual consistency without global optimization. Stage 4: Document Assembly. Textual blocks and rendered images are assembled into a Markdown document in alternating order, requiring no LLM involvement. 3.4 Evaluation Framework We propose an 8-dimension rubric grounded in multimedia learning theory (Sweller, 1988; Paivio, 1971; Mayer, 2002). Each dimension is scored on a 0–5 Likert scale and reported as a percentage (0–100%). Text quality (evaluated on extracted <TEXT_k> blocks): Correctness & Completeness—whether the explanation reaches the correct answer through valid reasoning with all intermediate steps; Logical Coherence—whether reasoning flows naturally from premises to conclusions; Pedagogical Effectiveness—whether explanations use grade-appropriate language and effective instructional strategies; Typographic Clarity—proper mathematical notation, consistent formatting, and absence of artifacts. Visual quality (evaluated on rendered diagrams): Diagram–Problem Alignment—whether diagrams faithfully represent the problem’s geometric or physical setup; Element Layout Quality—whether elements avoid overlap, maintain readable spacing, and follow discipline conventions; Visual Consistency—whether multiple diagrams maintain coherent color schemes, labeling, and style; Text–Diagram Coordination—how well textual references integrate with the corresponding diagrams. Automated Evaluation Protocol. We employ Gemini 3.0 Pro Preview (temperature=0) as the judge model, selected for its strong multimodal reasoning and 2M-token context window. Evaluation input varies by dimension type: text-only dimensions receive the problem statement, extracted text, and rubric criteria (plus the gold solution for Correctness & Completeness); multimodal dimensions additionally receive the relevant rendered diagrams alongside surrounding text. For Element Layout Quality, each diagram is scored independently and aggregated via geometric mean ELQ=(∏i=1Nsi)1/NELQ= ( _i=1^Ns_i )^1/N, ensuring a single poor diagram substantially penalizes the overall score. For Visual Consistency, Scene 1 serves as the visual anchor; each subsequent diagram is compared against it, with VC=(∏i=2Ns6,i)1/(N−1)VC= ( _i=2^Ns_6,i )^1/(N-1). Human Evaluation Protocol. To validate LLM-as-judge reliability, we conduct human evaluation on 30 stratified random sampled explanations, scored by 20 expert raters across 7 dimensions (Correctness & Completeness excluded) on a 3-level scale 0, 0.5, 1, producing 4,200 judgments. To enable direct comparison with the automated scores (reported as percentages), LLM scores are normalized to [0, 1] by dividing by 5 before computing Spearman’s ρ. We compute Krippendorff’s α for inter-rater agreement and Spearman’s ρ for human-AI correlation. Full annotation procedures are described in Appendix D. 4 Experimental Results We evaluate ten LLMs on the EduIllustrate benchmark, analyzing performance across dimensions, subjects, and grade levels. We validate our sequential workflow design through ablation studies and assess LLM-as-judge reliability via human evaluation with 20 expert raters. 4.1 Experimental Setup Models Evaluated. We evaluate ten LLMs spanning proprietary and open-weight categories: (1) Gemini 3.0 Pro Preview (Google DeepMind), (2) GPT-5 (OpenAI), (3) Claude Sonnet 4.5 (Anthropic), (4) Kimi-K2.5 (Moonshot AI), (5) Qwen3.5-397B, (6) Qwen3.5-122B, (7) Qwen 3.5-35B (Alibaba Cloud), (8) Mistral-Large-3, (9) Mistral-Small-4, and (10) Ministral-3-14B (Mistral AI). All models use identical prompts with temperature=0.7 on all 230 benchmark problems. 4.2 Main Results Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 87.6% 94.8% 78.4% 98.4% 87.4% 84.6% 89.4% 92.6% 87.8% 97.4% Kimi-K2.5 85.2% 92.2% 73.4% 97.4% 71.4% 68.8% 88.6% 86.2% 80.8% 98.3% Qwen3.5-397B 88.6% 94.4% 70.2% 91.2% 49.2% 61.8% 83.6% 67.8% 72.0% 93.9% Qwen3.5-122B 83.4% 92.0% 71.2% 93.4% 42.8% 58.0% 84.6% 59.4% 68.6% 94.8% Qwen 3.5-35B 83.4% 89.0% 69.2% 87.8% 39.6% 53.6% 88.6% 55.0% 65.8% 83.9% GPT-5 59.0% 66.6% 44.0% 74.2% 44.4% 59.2% 92.0% 63.8% 58.0% 89.6% Claude Sonnet 4.5 55.2% 58.8% 47.6% 77.4% 42.6% 64.0% 84.2% 68.8% 57.8% 96.1% Mistral-Large-3 39.8% 40.0% 32.4% 82.0% 24.6% 47.6% 83.2% 45.0% 43.0% 74.8% Mistral-Small-4 39.8% 38.0% 30.2% 71.6% 23.8% 49.2% 81.6% 39.4% 40.8% 70.4% Ministral-3-14B 37.6% 37.0% 31.0% 77.6% 23.8% 49.6% 92.4% 35.4% 41.0% 17.4% Table 2: Model performance on the full EduIllustrate benchmark (n=230, 0–100% scale). Overall: geometric mean of 8 dimensions. Success Rate: pipeline success rate. Bold indicates best per column. Per-subject and per-grade breakdowns are in Appendix E. Figure 5: Radar charts of all 10 models. (a) By dimension: 8 evaluation dimensions. (b) By subject: overall scores on each of the five subjects. (c) By grade level: overall scores at each grade level. Table 2 presents overall performance across all eight dimensions. The overall score is computed as the geometric mean of all eight dimension scores, ensuring that a single critically low score substantially penalizes the overall result. Gemini 3.0 Pro Preview achieves the highest overall score (87.8%), substantially outperforming all other models. Kimi-K2.5 ranks second (80.8%). The Qwen family forms a middle tier (65.8%–72.0%), while GPT-5 and Claude Sonnet 4.5 score 58.0% and 57.8% respectively. The Mistral family scores lowest (40.8%–43.0%), with Ministral-3-14B severely limited by its 82.6% failure rate. The over 46-point gap between the best (87.8%) and worst (41.0%) models demonstrates that architectural choices significantly impact multimodal educational content quality. All models excel at Logical Coherence and Typographic Clarity, reflecting LLMs’ strength in coherent, well-formatted text generation. Text–Diagram Coordination achieves high scores even for weaker models (Claude: 68.8%, GPT-5: 92.0%), indicating models produce well-integrated textual references to diagrams regardless of diagram quality. Notably, Ministral-3-14B and GPT-5 achieve the two highest Visual Consistency scores (92.4% and 92.0% respectively) despite weak overall performance; models with high failure rates tend to produce fewer scenes per problem, making cross-scene consistency trivially easier to maintain. Critical weaknesses appear in Diagram–Problem Alignment, where scores range from 23.8% to 87.4%. Qwen 3.5-35B’s low score of 39.6% and the Mistral family’s scores near 23.8%–24.6% reflect frequent semantic misalignment. Pedagogical Effectiveness is universally weakest (30.2%–78.4%), indicating all models struggle with grade-appropriate language and scaffolding strategies. Failure rates on the 230-problem benchmark vary substantially: Kimi-K2.5 is most robust (1.7%), followed by Gemini (2.6%), Claude (3.9%), Qwen3.5-122B (5.2%), Qwen3.5-397B (6.1%), GPT-5 (10.4%), Qwen 3.5-35B (16.1%), Mistral-Large-3 (25.2%), Mistral-Small-4 (29.6%), and Ministral-3-14B (82.6%). The Mistral family—particularly Ministral-3-14B—shows substantially higher failure rates, indicating limited robustness to complex Manim code generation. Failed problems are excluded from scoring; reported scores reflect only successfully completed problems. Model Mathe- matics Physics Chem- istry Biology Geog- raphy Subject Avg. Elem. School Middle School High School Gemini 3.0 Pro Preview 89.4% 88.6% 87.9% 85.7% 83.8% 87.1% 86.9% 88.0% 87.8% Kimi-K2.5 83.9% 80.6% 79.8% 78.9% 75.3% 79.7% 80.0% 80.3% 81.4% Qwen3.5-397B 75.0% 72.6% 70.7% 68.3% 67.3% 70.8% 76.3% 72.5% 70.5% Qwen3.5-122B 73.0% 68.9% 66.8% 63.4% 63.0% 67.0% 74.2% 69.2% 67.0% Qwen 3.5-35B 69.2% 64.6% 65.0% 61.5% 61.8% 64.4% 73.4% 65.8% 63.7% GPT-5 60.6% 58.5% 57.3% 55.1% 50.5% 56.4% 68.7% 57.2% 56.4% Claude Sonnet 4.5 62.1% 59.9% 54.6% 53.2% 49.5% 55.9% 64.2% 56.7% 57.5% Mistral-Large-3 46.4% 42.1% 42.6% 37.9% 39.1% 41.6% 47.2% 44.4% 40.7% Mistral-Small-4 45.2% 39.4% 38.5% 37.2% 37.0% 39.5% 42.4% 41.3% 39.9% Ministral-3-14B 44.9% 41.0% 36.2% 29.1% 35.1% 37.3% 35.3% 42.9% 40.5% Table 3: Overall scores (geometric mean, 0–100% scale) by subject and grade level for all 10 models. Subject Average: unweighted mean across five subjects. Bold indicates best per column. Per-subject and per-grade breakdowns with full dimension scores are provided in Appendix E. Table 3 summarizes the overall scores for five representative models across subjects and grade levels. Mathematics consistently achieves the highest scores across models, while geography scores lowest. Elementary school problems consistently yield higher scores than high school problems for most models (e.g., GPT-5: 68.7% vs. 56.4%; Claude: 64.2% vs. 57.5%), suggesting that increasing problem complexity and domain-specific reasoning demands degrade performance. Gemini and Kimi maintain relative consistency across grade levels, while other models show greater degradation on high school problems. 4.3 Workflow Ablation Study EduIllustrate is built upon a modified version of the TheoremExplainAgent codebase (Ku et al., 2025), which originally adopts an all-parallel workflow where all scenes generate implementation plans and code in parallel. We compare our sequential Scene 1 + parallel Scenes 2-N workflow against this all-parallel baseline to quantify the impact of our anchoring strategy. The baseline generates and renders all scenes in parallel without Scene 1 conditioning. Table 4 shows the all-parallel approach incurs 104% higher input token consumption (95.5K vs. 46.8K)—because every scene independently generates a full implementation plan before code generation, whereas our workflow restricts planning to Scene 1 only—92% higher cost ($0.94 vs. $0.49), and 13% worse Visual Consistency (77.8% vs. 89.4%), while achieving lower overall quality (85.8% vs. 87.8%). Manual inspection reveals the all-parallel workflow’s consistency failures stem from independent scene generation preventing style propagation. A side-by-side visual comparison is shown in Figure 6. This validates our sequential anchoring strategy. Figure 6: Visual consistency comparison between our sequential anchoring workflow (top) and the all-parallel baseline (bottom) on the same problem (four-pyramid P-ABCDABCD). In our workflow, Scene 1 and Scene 2 render the identical pyramid structure—Scene 2 simply highlights the PACPAC face on top of the same base diagram—maintaining full geometric and stylistic consistency. In the all-parallel baseline, the pyramid in Scene 1 and Scene 2 is drawn differently, breaking visual coherence and forcing the reader to reconcile two inconsistent representations of the same object. Figure 7: Per-stage cost comparison between our sequential anchoring workflow and the all-parallel baseline (both using Gemini 3.0 Pro Preview). The all-parallel approach incurs substantially higher Implementation Plans cost due to independent per-scene planning. Workflow Tokens (In/Out) Cost ($) Visual Consistency Overall Ours 46.8K / 37.1K 0.49 89.4% 87.8% All-Parallel 95.5K / 71.4K 0.94 77.8% 85.8% Table 4: Workflow comparison using Gemini 3.0 Pro Preview on 50 problems. Our workflow achieves superior visual consistency at lower cost. 4.4 Human-AI Agreement Analysis Table 5 presents inter-rater reliability (Krippendorff’s α) and human-AI agreement (Spearman’s ρ) for 30 explanations evaluated across 7 dimensions by 20 expert raters. Logical Coherence and Diagram–Problem Alignment achieve strong human consensus (α≥0.80α≥ 0.80) and strong human-AI agreement (ρ≥0.80ρ≥ 0.80), validating LLM-as-judge for these dimensions. Logical Coherence’s high reliability reflects that logical gaps are objectively identifiable; Diagram–Problem Alignment’s strong performance is notable given it requires visual reasoning. Typographic Clarity achieves strong-to-moderate agreement (ρ=0.77ρ=0.77, α=0.67α=0.67), as formatting artifacts are largely objective but occasionally ambiguous at borderline cases. Pedagogical Effectiveness and Text–Diagram Coordination show moderate reliability (α≈0.64−0.68α≈ 0.64-0.68, ρ≈0.66−0.70ρ≈ 0.66-0.70), acceptable for comparative studies but requiring human validation for high-stakes decisions. Visual Consistency and Element Layout Quality exhibit weak reliability (α≈0.30α≈ 0.30, ρ≈0.40ρ≈ 0.40). Beyond ill-defined rubric constructs, a key contributing factor is that the LLM judge model is insensitive to low-level visual artifacts—such as overlapping elements, misaligned labels, or rendering glitches—that human raters readily detect when inspecting rendered diagrams. Dim. Spearman Krippendorff Reliability ρ α LC 0.89 0.84 Strong DPA 0.83 0.80 Strong TC 0.77 0.67 Strong-Mod. PE 0.70 0.68 Moderate TDC 0.66 0.64 Moderate VC 0.47 0.30 Weak ELQ 0.39 0.32 Weak Table 5: Human-AI agreement and inter-rater reliability (30 explanations, 20 raters). Correctness & Completeness excluded (objective gold answer verification). 4.5 Cost Analysis Model Avg Input Avg Output Input Output Avg Cost Tokens Tokens ($/M) ($/M) ($) Gemini 3.0 Pro Preview 46,821 37,096 0.86 12.04 0.49 Claude Sonnet 4.5 71,889 23,964 0.65 15.10 0.41 Qwen3.5-122B-A10B 39,487 29,262 0.40 3.20 0.11 qwen3.5-397b 39,937 27,285 0.60 3.60 0.12 Kimi-K2.5 44,537 43,628 0.45 2.22 0.12 GPT-5 34,568 10,178 0.74 10.00 0.13 Mistral-Large-3 44,600 15,081 0.42 1.50 0.04 Qwen 3.5-35B 37,269 24,943 0.25 2.00 0.06 Mistral-Small-4 54,195 17,833 0.13 0.60 0.02 Ministral-3-14B 50,488 18,128 0.17 0.20 0.01 Table 6: Cost analysis per problem (230 problems, average values). Figure 8: Cost vs. quality trade-off across ten models. X-axis: average cost per problem (USD); Y-axis: overall score (0–100%). Kimi-K2.5 offers the best cost-efficiency among high-quality models, achieving 80.8% at $0.12/problem. Figure 9: Per-stage cost breakdown across ten models. Each bar decomposes the average per-problem cost into three pipeline stages: Scene Outline, Implementation Plans, and Scene Rendering. Scene Rendering dominates cost for all models. Table 6 presents cost metrics across all ten models. As shown in Figure 8, Kimi-K2.5 offers the best cost-efficiency among high-quality models at $0.12/problem (80.8% score), representing only 8.0% quality degradation relative to Gemini ($0.49, 87.8% score) at 4.1× lower cost. Among lower-cost options, Ministral-3-14B achieves the lowest cost at $0.01/problem, while Mistral-Large-3 and Mistral-Small-4 offer competitive quality at $0.04 and $0.02 respectively. 5 Conclusion We presented EduIllustrate, a benchmark for evaluating LLMs on K-12 STEM illustrated explanation generation, comprising a 230-problem dataset, a standardized generation protocol with sequential Scene 1 anchoring, and an 8-dimension rubric grounded in multimedia learning theory. Sequential anchoring improves Visual Consistency by 13% over fully parallel generation at 94% lower cost. Human evaluation validates LLM-as-judge reliability on objective dimensions while revealing limitations on subjective visual assessment. Limitations The pipeline currently generates static PNG images via Manim, limiting both output modality and framework generalizability. While the agent framework can be adapted to produce animations by adjusting rendering flags and prompts, our 8-dimension rubric does not cover temporal coherence or animation pacing. The pipeline is also tightly coupled to Manim; extending to target-agnostic intermediate representations (Kumar et al., 2025) would improve flexibility to other frameworks (TikZ, matplotlib, SVG). Weak human-AI agreement on visual dimensions (Element Layout Quality: ρ=0.39ρ=0.39, Visual Consistency: ρ=0.47ρ=0.47) indicates current vision-language models struggle to assess layout quality and stylistic consistency reliably, and the low inter-rater agreement among humans (α≈0.30α≈ 0.30) suggests these rubric constructs require further refinement. More broadly, Pedagogical Effectiveness is the weakest dimension across all models (50.8%–78.4%), as the pipeline delivers static, one-shot explanations with no adaptation to the individual learner. Future work will explore multi-turn Socratic interaction and student profile integration to personalize both explanation strategy and diagram design. Acknowledgments We thank the 20 expert raters (15 STEM graduate students and 5 education doctoral students) for their meticulous human evaluation work. We acknowledge support from [funding agency] under grant [number]. References J. An, Z. Yang, L. Li, J. Wang, K. Lin, Z. Liu, L. Wang, and J. Luo (2024) OpenLEAF: a novel benchmark for open-domain interleaved image-text generation. In Proceedings of the 32nd ACM International Conference on Multimedia, p. . External Links: Document Cited by: §1. Y. Chen, K. Q. Lin, and M. Z. Shou (2025) Code2Video: a code-centric paradigm for educational video generation. arXiv preprint arXiv:2510.01174. Cited by: §1, §2.2. E. Chern, J. Su, Y. Ma, and P. Liu (2024) Anole: an open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135. Cited by: §1. Z. Cui, J. Yuan, H. Wang, Y. Li, C. Du, and Z. Ding (2025) Draw with thought: unleashing multimodal reasoning for scientific diagram generation. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 5050–5059. Cited by: §2.3. J. Enguehard, M. Van Ermengem, K. Atkinson, S. Cha, A. G. Chowdhury, P. K. Ramaswamy, J. Roghair, H. R. Marlowe, C. S. Negreanu, K. Boxall, et al. (2025) LeMAJ (legal llm-as-a-judge): bridging legal reasoning and llm evaluation. In Proceedings of the Natural Legal Language Processing Workshop 2025, p. 318–337. Cited by: §2.4. H. Ji, S. Qiu, S. Xin, S. Han, Z. Chen, D. Zhang, H. Wang, and H. Yao (2025) From eduvisbench to eduvisagent: a benchmark and multi-agent framework for reasoning-driven pedagogical visualization. arXiv preprint arXiv:2505.16832. Cited by: §1. F. Kamalov, D. S. Calonge, L. Smail, D. Azizov, D. R. Thadani, T. Kwong, and A. Atif (2025) Evolution of ai in education: agentic workflows. arXiv preprint arXiv:2504.20082. Cited by: §1, §2.1. S. Kou, J. Jin, Z. Liu, C. Liu, Y. Ma, J. Jia, Q. Chen, P. Jiang, and Z. Deng (2025) Orthus: autoregressive interleaved image-text generation with modality-specific heads. arXiv preprint arXiv:2412.00127. Cited by: §1. M. Ku, C. H. Chong, J. Leung, K. Shah, A. Yu, and W. Chen (2025) Theoremexplainagent: towards video-based multimodal explanations for llm theorem understanding. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 6663–6684. Cited by: §1, §2.2, §4.3. V. Kumar, S. Mishra, R. Hao, R. Malik, D. Broman, and D. Demszky (2025) DiagramIR: an automatic pipeline for educational math diagram evaluation. arXiv preprint arXiv:2511.08283. Cited by: §1, §2.3, Limitations. C. Li, C. Zhu, T. Zhang, M. Lin, Z. Zhou, and J. Xie (2025) K12vista: exploring the boundaries of mllms in k-12 education. arXiv preprint arXiv:2506.01676. Cited by: Appendix A, §3.2. Z. Liu, S. X. Yin, D. H. Goh, and N. F. Chen (2025) COGENT: a curriculum-oriented framework for generating grade-appropriate educational content. arXiv preprint arXiv:2506.09367. Cited by: §1, §1, §2.1. R. E. Mayer (2002) Multimedia learning. In Psychology of learning and motivation, Vol. 41, p. 85–139. Cited by: §1, §3.4. A. Paivio (1971) Imagery and verbal processes. Holt, Rinehart and Winston, New York. Cited by: §1, §3.4. K. Park and J. Yang (2025) AGACCI: affiliated grading agents for criteria-centric interface in educational coding contexts. arXiv preprint arXiv:2507.05321. Cited by: §2.4. R. A. Philipp (2007) Mathematics teachers’ beliefs and affect. In Second Handbook of Research on Mathematics Teaching and Learning, F. K. Lester (Ed.), p. 257–315. Cited by: §1. J. Sweller (1988) Cognitive load during problem solving: effects on learning. Cognitive Science 12 (2), p. 257–285. External Links: Document Cited by: §1, §3.4. P. N. Thompson and M. Dahlin (2024) Teachers’ time use and well-being: evidence from the american time use survey. Educational Researcher 53 (2), p. 99–110. External Links: Document Cited by: §1. T. Türközü and B. Dinçer (2025) Visual literacy in science education: pre-service teachers’ competencies and challenges in creating instructional materials. Journal of Science Education and Technology. Note: In press Cited by: §1. T. Wang, Y. Zhan, J. Lian, Z. Hu, N. J. Yuan, Q. Zhang, X. Xie, and H. Xiong (2025) Llm-powered multi-agent framework for goal-oriented learning in intelligent tutoring system. In Companion Proceedings of the ACM on Web Conference 2025, p. 510–519. Cited by: §1, §1, §2.1. P. Xia, S. Han, S. Qiu, Y. Zhou, Z. Wang, W. Zheng, Z. Chen, C. Cui, M. Ding, L. Li, L. Wang, and H. Yao (2025) MMIE: massive multimodal interleaved comprehension benchmark for large vision-language models. In International Conference on Learning Representations, Cited by: §1. J. Yao et al. (2026) Instructional material design in k-12 stem: a systematic review of challenges and technological interventions. Computers & Education. Note: In press Cited by: §1. L. Zheng, W. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al. (2023) Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems 36, p. 46595–46623. Cited by: Appendix B. Appendix A Benchmark Construction Details Dataset Curation Process Benchmark construction followed a two-phase procedure. In the first phase, Kimi-K2.5 automatically screened all candidate problems from K12-Vista (Li et al., 2025) for Diagram Appropriateness: for each problem, the model was prompted to judge whether a visual representation would convey spatial, topological, or relational information essential to problem-solving (rather than merely restating the text), and to output a binary decision with a brief justification. In the second phase, two human annotators independently reviewed every problem against all three curation criteria—(1) Diagram-Appropriate Problems, (2) Clear Gold Solutions, and (3) Diverse Topics—and resolved disagreements through discussion, using a dedicated review interface (Figure 10). Problems were retained only when both annotators confirmed all three criteria were satisfied. This two-phase design combines the scalability of LLM screening with the reliability of human expert judgment. Figure 10: Screenshot of the problem review interface used during human annotation. Annotators evaluated each candidate problem against the three curation criteria and recorded their decisions with justifications. Topic Coverage Table 7 lists the complete topic coverage of the 230 benchmark problems, organized by grade level and subject. Level Subject Topics High School Biology (15) Law of independent assortment (×3); PCR amplification; Variants of 9:3:3:1 and 1:1:1:1 ratios (×2); Sex-linked inheritance (×3); Gene locus determination; Chromosomal structural variation; Genetic engineering; Transcription and translation; Ecosystem structure Chemistry (15) Galvanic and electrolytic cell principles (×2); Crystal structure and properties; Organic compound structure and properties (×2); Chemical equilibrium and equivalent equilibrium (×2); Organic molecular formula determination; Structural features of carbon bonding; Isomer enumeration Geography (15) Earth and maps; Latitude/longitude and coordinate grids (×2); Earth’s revolution: noon solar altitude and day-length variation (×3); General characteristics of Earth’s motion (×2); Geographic significance of Earth’s motion; Air pressure belts and wind belts Mathematics (30) Proportional segments and circles (×2); Hyperbola properties; Proposition truth and plane relations (×2); Inscribed polygon properties; Tangent line proof; Conic section common features; Plane-plane perpendicularity; Tangent-chord angle; Inductive reasoning; Parabola properties; Limits of sequences; Pyramid structure (×2); Ellipse properties; Law of sines; Sphere volume and surface area; Area/volume from three-view drawings; Similar triangles (×2); Spatial line-line relations; Arithmetic sequences; Spatial vectors and sphere; (some topics unlabeled) Physics (30) Concurrent force equilibrium and electric field; Work-energy theorem with projectile motion; Work-energy theorem with charged particle in uniform field (×3); Work-energy theorem and mechanical energy conservation (×2); Conservation of momentum (×2); EMF from conductor cutting field lines and Joule’s law (×3); Coulomb’s law and Newton’s second law; Mechanical energy conservation and projectile motion; Ideal gas law and enclosed gas pressure; SHM, vertical projection, and Hooke’s law; Energy conservation, steady current, and force equilibrium; Velocity selector and charged particle in combined fields Middle School Biology (15) DNA replication/transcription/translation calculations; Ecosystem and ecological system; Complete and incomplete metamorphosis; Gene–DNA–chromosome relationships; Gas exchange in lungs; Blood vessels (arteries, veins, capillaries); Classification of algae, mosses, and ferns; Urinary system and urine formation Chemistry (15) Molecules, atoms, ions, elements and their relationships (×2); Molecular properties and acid-base indicators; Four basic reaction types and complex decomposition reactions; Oxygen production and properties; Chemical inference and reaction type identification; Solubility curves; Gas identification and purification; Particle model and conservation of mass (×2); Substance transformation; Substance identification and reaction type determination; Air composition measurement Geography (15) Latitude/longitude and coordinate grids (×3); Earth’s revolution and day-length variation; Earth’s rotation; Day-night alternation and terminator; Seasonal formation; Direction and location using coordinate grids (×2) Mathematics (30) Linear function applications; Triangle angle sum and exterior angles (×2); Triangle circumscribed circle; Quadratic/linear function graphs and coefficients (×2); Triangle congruence (SAS) and Pythagorean theorem; 3D solid nets (×2); Similar triangles (×2); Inequalities and linear/quadratic functions; 30° right triangle, inscribed angles, and diameter; Net folding; Proportional segments, midsegment, and 30° triangle; Rotation and area; Rational number operations; Square properties and coordinates; Similar triangles and equilateral triangle; Golden ratio; Fold transformations; Axial symmetry; Perpendicular bisector properties Physics (30) Force balance and pulley systems; Light refraction ray diagrams; Convex lens imaging rules; Circuit fault diagnosis; Force and motion; Pressure and gravity vs. pressure; Pressure comparison and density; Balanced forces and energy changes; Spring force meter; Friction; Lever equilibrium (×2, including minimum force); Ohm’s law and electromagnet (×2); Ohm’s law applications (×2); Pulley rope tension, work, and power; Ammeter usage; Three circuit states; Dynamic circuit analysis (×3); Magnetic poles and Ampere’s right-hand rule; Archimedes’ principle (×2) Elementary Mathematics (20) Three-view and net diagrams (×2); Observing 3D shapes from different directions; Shape composition; Perimeter of circles and rings; Position and direction (×3); Rotation and coordinate position; Perimeter of composite figures; Clever perimeter calculation; Inverse/direct proportion applications; Composite shape counting; Overlapping problems; Cube/cuboid nets and folding; Area comparison Table 7: Complete topic coverage of the 230 EduIllustrate benchmark problems. Numbers in parentheses after subject names indicate problem count; (×n) after a topic indicates n problems on that topic. Some high school mathematics problems have unlabeled topics in the original K12-Vista dataset. Appendix B Judge Model Self-Preference Analysis A potential concern with using Gemini 3.0 Pro Preview as the judge model is self-preference bias: the judge may systematically inflate scores for outputs it generated. To investigate this, we sampled 20 problems and re-evaluated all outputs from all five models plus the all-parallel workflow variant (120 explanations total) using GPT-5 as an alternative judge, then compared the two scoring distributions. Both Judges Exhibit Self-Preference, but Gemini’s Is Smaller. Table 8 reports the mean overall score assigned by each judge. Both judges show self-preference: Gemini-as-judge scores its own outputs 9.2% higher than GPT-5-as-judge does (88.8% vs. 79.6%), while GPT-5-as-judge scores its own outputs 20.6% higher than Gemini-as-judge does (81.8% vs. 61.2%). Crucially, Gemini’s self-preference magnitude (|Δ|=9.2%| |=9.2\%) is less than half of GPT-5’s (|Δ|=20.6%| |=20.6\%). Moreover, GPT-5’s self-preference is severe enough to alter rankings: GPT-5-as-judge ranks itself first, whereas Gemini-as-judge ranks it last. By contrast, Gemini-as-judge and GPT-5-as-judge agree on the top-2 ranking (Gemini >> Kimi), confirming that Gemini’s self-preference does not distort the main comparative conclusions. Kimi-K2.5 serves as a neutral anchor with the smallest cross-judge gap (|Δ|=2.4%| |=2.4\%), further validating the evaluation framework’s consistency on models with no judge affiliation. Dimension-Level Agreement. Table 9 reports Pearson r, Spearman ρ, and MAE between the two judges across all 120 explanations for each dimension. Element Layout Quality (r=0.58r=0.58, ρ=0.56ρ=0.56) and Text–Diagram Coordination (r=0.56r=0.56, ρ=0.58ρ=0.58) show the highest agreement, reflecting relatively objective visual criteria. Diagram–Problem Alignment has the largest MAE (1.15), indicating the two judges disagree most on diagram–problem semantic matching. Typographic Clarity shows near-zero correlation (r=0.10r=0.10, ρ=0.12ρ=0.12, both non-significant), suggesting this dimension’s assessment is highly judge-dependent. The overall score achieves moderate agreement (r=0.50r=0.50, ρ=0.43ρ=0.43, MAE=0.61=0.61), indicating directionally consistent but numerically divergent scoring. Implications. Self-preference bias is present in both judge models, consistent with prior findings on LLM-as-judge self-preference (Zheng et al., 2023). We select Gemini over GPT-5 as the primary judge because (1) its self-preference is substantially smaller, (2) it does not alter the model ranking, and (3) the top-2 ranking is robust across both judges. Nevertheless, the moderate cross-judge agreement on overall scores (Spearman ρ=0.43ρ=0.43) underscores that absolute score values should be interpreted with caution, and we recommend future work report multi-judge robustness checks. Model (Generator) Gemini GPT-5 Δ Judge Judge Gemini 3.0 Pro Preview 88.8% 79.6% −-9.2% Kimi-K2.5 81.2% 78.8% −-2.4% Qwen 3.5-35B 67.8% 74.6% +7.0% Claude Sonnet 4.5 63.4% 72.8% +9.4% GPT-5 61.2% 81.8% +20.6% Table 8: Mean overall scores (0–100%) assigned by Gemini 3.0 Pro Preview and GPT-5 as judge models on the same 20 problems (120 explanations). Δ = GPT-5 judge −- Gemini judge. GPT-5-as-judge exhibits strong self-preference (Δ=+20.6% =+20.6\%). Dimension Pearson Spearman MAE r ρ ELQ 0.58 0.56 0.62 TDC 0.56 0.58 0.74 VC 0.48 0.46 0.45 LC 0.45 0.38 0.74 DPA 0.40 0.42 1.15 C&C 0.39 0.39 0.87 PE 0.29 0.27 0.90 TC 0.10† 0.12† 0.77 Overall 0.50 0.43 0.61 Table 9: Dimension-level agreement between Gemini and GPT-5 judges (120 explanations). †Non-significant (p>0.05p>0.05). Dimensions sorted by Pearson r. Appendix C A Gallery of Generated Explanations We present representative high-quality and low-quality explanations generated by EduIllustrate to illustrate the range of outputs across different subjects and failure modes. Each entry shows the problem statement, key solution steps, and the corresponding rendered diagrams. High-Quality Explanations Middle School Physics — Circuit Analysis. This example demonstrates strong Visual Consistency and Text–Diagram Coordination. All three scenes use identical circuit drawing conventions, and each scene incrementally modifies the previous one—adding or removing a single wire—so that the student can track exactly what changed. The progressive structure mirrors effective pedagogical scaffolding. Problem: The circuit is shown. Which analysis is incorrect? (A) After S is closed, the circuit short-circuits. (B) After S is closed, L1L_1 and L2L_2 are in parallel and both light up. (C) Removing wire b makes L1L_1 and L2L_2 series-connected. (D) Moving wire M from B to A makes L1L_1 and L2L_2 parallel. Approach: Trace current paths from the positive terminal to determine circuit topology under each scenario, then evaluate each option. Scene 1: Original circuit diagram with current path analysis. Wire b short-circuits the bulbs, creating a zero-resistance path. Step 2: Removing wire b forces current through L2→L1L_2→ L_1 sequentially—series connection confirmed. Scene 2: Modified circuit with wire b removed. Single current path through both bulbs in series. Step 3: Moving wire M from B to A creates two independent paths. Both bulbs connect between the same high/low potential nodes—parallel connection. Scene 3: Circuit with wire M moved to A. Both bulbs span identical potential difference—parallel. Answer: B (bulbs do not light up; short circuit prevents current through loads). Middle School Biology — DNA Replication. This example showcases consistent color coding (blue for 15N strands, orange for 14N strands) maintained across all three scenes. Each scene builds directly on the previous one, progressing from molecular structure to base composition to a replication tree. The color scheme serves as a visual thread that ties the explanation together—a hallmark of high Diagram–Problem Alignment. Problem: A 15N-labeled eukaryotic gene has Cytosine (C) accounting for 30% of all bases. Which statement is correct? (A) Helicase acts on sites ① and ②. (B) In one strand, (C+G)/(A+T) = 3:2. (C) After T→AT→ A mutation at site ①, after n replications the mutant gene is 1/4. (D) After 3 replications in 14N medium, DNA with 14N is 3/4. Approach: Identify bond types at sites ① and ②; apply Chargaff’s rules; apply semi-conservative replication. Scene 1: DNA double helix with sites ① (hydrogen bonds between bases) and ② (phosphodiester bonds in backbone) clearly labeled. Step 2: Chargaff’s rules: C=G=30%, so A=T=20%. Within one strand, C+G=60%, A+T=40%, giving ratio 3:2—Statement B is correct. Scene 2: Base composition diagram showing C/G/A/T proportions derived from Chargaff’s rules. Step 3: Semi-conservative replication after 3 rounds yields 23=82^3=8 DNA molecules; all 8 contain 14N (since new strands use 14N); 2 of 8 also retain 15N. So 14N proportion = 8/8 = 100%, not 3/4—Statement D is wrong. Scene 3: Semi-conservative replication tree after 3 rounds, showing isotope distribution across 8 daughter molecules. Answer: B. High School Math — Geometric Series & Inscribed Circles. The three scenes progressively construct nested hexagons and circles, each building on the previous geometric structure. Annotations remain consistent (same label positions, line styles), and the final scene visualizes the infinite nesting—an abstract concept made concrete through visual accumulation. This illustrates how sequential anchoring preserves spatial coherence across increasingly complex diagrams. Problem: A regular hexagon is inscribed in a circle of radius 1 m. Its inscribed circle is drawn, inside which another regular hexagon is inscribed, and so on infinitely. Find the sum S of the areas of all circles. Approach: Find the ratio between successive circle radii using the apothem of a regular hexagon, identify the geometric series common ratio, then apply the infinite sum formula. Scene 1: First circle (R1=1R_1=1) with inscribed hexagon and its apothem shown. The apothem equals 32R1 32R_1, giving R2=32R_2= 32. Step 2: Area ratio =(R2R1)2=34= ( R_2R_1 )^2= 34. The areas form a geometric series with first term π and common ratio 34 34. Scene 2: First two circles and hexagons overlaid, illustrating the scaling relationship. Step 3: S=π1−34=4πS= π1- 34=4π m2. Scene 3: Infinite nesting visualized with converging circles and hexagons. Answer: S=4πS=4π m2. High School Math — Solid Geometry & Perpendicular Planes. This example requires multi-step spatial reasoning. The three scenes incrementally reveal the proof: Scene 1 establishes the rhombus base, Scene 2 proves BD⟂BD plane PACPAC, and Scene 3 derives the perpendicularity condition. Each diagram preserves the same 3D viewpoint and labeling, allowing the reader to follow the logical chain without re-orienting spatially—strong evidence of both Visual Consistency and Logical Coherence. Problem: In quadrangular pyramid P-ABCDABCD, PA⟂PA base ABCDABCD, all base edges are equal (rhombus). M is a moving point on PCPC. State a condition on M such that plane MBD⟂MBD plane PCDPCD. Approach: Use the plane-perpendicularity theorem: find a line in plane MBDMBD perpendicular to plane PCDPCD. Show BD⟂PCBD PC using the rhombus diagonal and PA⟂PA base properties, then determine where MBDMBD must intersect PCDPCD. Scene 1: Pyramid with rhombus base ABCDABCD, PAPA perpendicular to base, diagonal BDBD highlighted. Step 2: Since ABCDABCD is a rhombus, AC⟂BDAC BD. Since PA⟂PA base, PA⟂BDPA BD. Thus BD⟂BD plane PACPAC, hence BD⟂PCBD PC. Scene 2: Plane PACPAC highlighted, showing BD⟂BD plane PACPAC and therefore BD⟂PCBD PC. Step 3: BD⟂PCBD PC and BD⊂BD⊂ plane MBDMBD; if M is the foot of the altitude from B to PCPC (i.e., MB⟂PCMB PC), then BD⟂BD plane PCDPCD, giving plane MBD⟂MBD plane PCDPCD. Scene 3: Plane MBDMBD and plane PCDPCD shown with their intersection line, confirming perpendicularity condition. Answer: MB⟂PCMB PC (i.e., M is the foot of the perpendicular from B to PCPC). Middle School Math — Quadratic Function Coefficients. The two scenes effectively split the reasoning into visual subproblems: Scene 1 reads the signs of a and b from the parabola’s shape, and Scene 2 plots the resulting linear function. Consistent axis styles and color coding across scenes reinforce the algebraic-to-graphical connection. The missing quadrant is clearly shaded, making the answer visually self-evident. Problem: The graph of y=ax2+bx+1y=ax^2+bx+1 is shown. Which quadrant does y=ax+by=ax+b not pass through? A. First B. Second C. Third D. Fourth Approach: Read signs of a and b from the parabola (opening direction, axis of symmetry position), then determine the linear function’s slope and intercept to identify the missing quadrant. Scene 1: Parabola opens downward (a<0a<0), axis of symmetry at x<0x<0, so −b/(2a)<0⇒b<0-b/(2a)<0 b<0. Step 2: For y=ax+by=ax+b: slope a<0a<0 (decreasing), y-intercept b<0b<0. A decreasing line with negative intercept passes through Quadrants I, I, IV—it does not pass through Quadrant I. Scene 2: Linear function y=ax+by=ax+b plotted with a<0a<0, b<0b<0, passing through Q2, Q3, Q4 only. Answer: A (First quadrant). Middle School Math — Square & Rhombus Shaded Area. A consistent color scheme (blue square, orange rhombus, green shaded region) is maintained across all three scenes, each progressively zooming into the region of interest. Scene 2 decomposes the unshaded area into computable triangles, and Scene 3 highlights the final answer. This example demonstrates effective use of color as a pedagogical anchor for the subtraction method. Problem: Square ABCDABCD has area 25; rhombus PQCBPQCB has area 20. Find the area of the shaded region. A. 11 B. 6.5 C. 7 D. 7.5 Approach: Determine side lengths from given areas; use the subtraction method to express the shaded area as the square minus an unshaded triangular region formed by the overlap of the square and rhombus. Scene 1: Square ABCDABCD (side=5) and rhombus PQCBPQCB overlaid; the shaded region and the unshaded overlap region clearly distinguished by color. Step 2: Side of square = 5. Height of rhombus = area/base = 20/5 = 4. The unshaded region inside the square is a right triangle with legs 5 and 5−4=15-4=1; its area = 12×5×1=2.5 12× 5× 1=2.5. Shaded area =25−2×2.5×…=25-2× 2.5×… (subtraction applied per configuration). Scene 2: Decomposition of the unshaded region into computable triangles. Scene 3: Final shaded region highlighted with computed area. Answer: D (7.5). Low-Quality Explanations Elementary Math — Cube Net Folding (Correctness Failure). The model applies the opposite-face pairing rule correctly in principle but misidentifies the symbol placement during the mapping step, arriving at answer C instead of the gold answer A. This is a pure Correctness & Completeness failure: the textual reasoning framework is sound, but a single perceptual error in reading the net topology propagates to an incorrect conclusion. Such failures highlight the gap between procedural competence and visual perception accuracy. Problem: Xiaoxiao made a cubic gift box where opposite faces share the same pattern (⋆ , ♡ , ⊙ ). Which net (A/B/C/D) folds into this cube? Approach: Apply opposite-face pairing rule for cross-type nets: positions (1,3), (2,4), and (top, bottom) are opposite pairs. Scene 1: Cross-type net with positions labeled and opposite-pair rule illustrated. Step 2: Tests each option by mapping symbols to opposite-pair positions. Concludes that option C satisfies the constraint. Scene 2: Net options annotated with opposite-face verification. Final Answer: C × Gold answer: A Failure mode (Correctness & Completeness): The reasoning process applies the opposite-face pairing rule correctly in principle, but misidentifies the symbol placement in option A vs. C during the mapping step, leading to an incorrect final answer. Middle School Math — Isosceles Triangle (Diagram–Problem Alignment Failure). Although the textual reasoning is correct (cos⁡36∘=5+14 36 = 5+14), the rendered diagrams draw an acute scalene triangle with visibly unequal sides, directly violating the given AB=ACAB=AC constraint. The 72∘72 base angles are not rendered at vertices B and C. A student relying on the diagrams would form incorrect spatial intuitions about the triangle’s shape, undermining the pedagogical value despite textual correctness. Problem: In △ABC ABC, AB=AC=4AB=AC=4, ∠C=72∘ C=72 . Point D is the midpoint of ABAB; point E lies on ACAC; DE⟂ABDE AB. Find cos⁡A A. Options: A. 5−12 5-12 B. 5−14 5-14 C. 5+14 5+14 D. 5+12 5+12 Approach: Use isosceles triangle properties (AB=AC⇒∠B=∠C=72∘AB=AC B= C=72 ), compute ∠A=180∘−144∘=36∘ A=180 -144 =36 , then evaluate cos⁡36∘ 36 . Scene 1: The diagram draws an acute scalene triangle with visibly unequal sides, violating the given AB=ACAB=AC constraint. The 72∘72 angles are not rendered at vertices B and C. Step 2: ∠A+72∘+72∘=180∘⇒∠A=36∘ A+72 +72 =180 A=36 . Scene 2: The highlighted triangle still shows inconsistent proportions. The isosceles constraint (AB=ACAB=AC) is visually absent, making the diagram misleading. Step 3: cos⁡36∘=5+14 36 = 5+14 Final Answer: C ✓ (textually correct) Failure mode (Diagram–Problem Alignment): Although the textual reasoning is correct, the rendered diagrams do not reflect the isosceles structure of the triangle—a fundamental geometric constraint of the problem. A student relying on the diagrams would form incorrect spatial intuitions. High School Physics — Gas Laws (Element Layout & Alignment Failure). A critical geometric parameter (3H/43H/4, the nitrogen column height in cylinder A) is mislabeled as H in Scene 1, and this error propagates into Scene 2. The textual solution computes correct volumes using 3H/43H/4, but a student using the diagram to set up equations would obtain wrong initial volumes. This exemplifies how Element Layout Quality failures can actively mislead learners even when the text is correct. Problem: Two connected vertical cylinders A (closed top, diameter 2dB2d_B, cross-section 4S4S) and B (open top, diameter dBd_B, cross-section S), both height H. Piston a in A is H/4H/4 from the top; piston b in B is at mid-height. N2N_2 fills below both pistons; O2O_2 fills above piston a. (1) Heat N2N_2 until piston b reaches the top: find N2N_2 temperature. (2) Continue heating until piston a rises H/16H/16: find O2O_2 pressure. Approach: Stage I: N2N_2 expands at constant pressure P0P_0 (piston b free); O2O_2 isothermal (constant volume). Stage I: N2N_2 volume in B fixed; O2O_2 compressed isothermally by Boyle’s law. Scene 1: The height label below piston a in cylinder A is marked H instead of the correct 3H/43H/4, contradicting the given initial geometry and making the diagram inconsistent with the problem. Step 2: Stage I—N2N_2 volume changes from VN0=4S⋅3H4+S⋅H2=7SH2V_N0=4S· 3H4+S· H2= 7SH2 to VN1=3SH4⋅4S+SH=4SHV_N1= 3SH4· 4S+SH=4SH. By pV∝TpV T: T1=T0⋅4SH7SH/2=87T0T_1=T_0· 4SH7SH/2= 87T_0. Scene 2: After piston b reaches the top, the diagram still carries the incorrect label from Scene 1, propagating the error into Stage I. Scene 3: Stage I diagram showing piston a risen by H/16H/16; layout partially corrected but inconsistent with Scene 1. Failure mode (Element Layout Quality / Diagram–Problem Alignment): The mislabeled height H (should be 3H/43H/4) in Scene 1 violates the problem’s geometric constraints. A student using the diagram to set up equations would obtain wrong initial volumes and incorrect answers. Middle School Geography — Latitude/Longitude Reading (Logical Coherence Failure). The model arrives at the correct answer (B) through systematically flawed reasoning: it misidentifies point A’s latitude zone and falsely claims the diagram lacks longitude reference markers. This “right answer, wrong method” failure is pedagogically dangerous—if adopted by students, it instills incorrect map-reading habits. The case reveals that Correctness & Completeness alone is insufficient to assess educational value; Logical Coherence and Diagram–Problem Alignment must be evaluated jointly. Problem: From a polar-view diagram of Earth with latitude circles and longitude markings, determine which statement is correct: (A) Point A receives direct sunlight only once a year. (B) Point B may receive direct sunlight twice a year. (C) Point B is in the Southern Hemisphere. (D) A is in the Eastern Hemisphere, B is in the Western Hemisphere. Key facts: Direct sunlight occurs only between 23.5∘23.5 S and 23.5∘23.5 N. Points on the tropics receive it once; points inside receive it twice. Hemisphere boundaries: equator (N/S), 20∘W–160∘E (E/W). Scene 1: Polar-view diagram. The model incorrectly identifies point A as lying outside the tropics (mid-latitude), concluding A receives no direct sunlight—in fact, A is on the equator and receives direct sunlight twice a year. Flawed reasoning: The model misreads A as a mid-latitude point, eliminating option A with wrong logic. It then claims the diagram lacks explicit 0∘0 /180∘180 longitude markings to determine the E/W hemisphere—but the diagram does contain computable longitude references, making option D verifiable. The model eliminates options A, C, D through incorrect intermediate steps and selects B by elimination. Scene 2: Hemisphere analysis diagram. The longitude reference lines drawn are inconsistent with the original diagram’s markings, reflecting the model’s failure to correctly read the given coordinate information. Final Answer: B ✓ (coincidentally correct) Failure mode (Logical Coherence / Diagram–Problem Alignment): Correct final answer reached through systematically flawed reasoning—wrong latitude identification for A, and incorrect claim about missing longitude markers. This is a “right answer, wrong method” failure that would mislead students. Appendix D Human Annotation Process Human evaluation was conducted through a dedicated annotation website (Figure 11). Raters were presented with a K-12 STEM problem, its gold-standard solution, and the generated illustrated explanation (including rendered diagrams). For each explanation, raters scored 7 dimensions (Logical Coherence through Typographic Clarity) on a coarse 3-level ordinal scale 0, 0.5, 1, where 0 indicates poor quality, 0.5 indicates acceptable quality, and 1 indicates high quality. Correctness & Completeness was excluded from human evaluation because solution correctness can be determined objectively by comparing against the gold-standard answer, making subjective human judgment unnecessary and potentially inconsistent. Each page displayed one explanation at a time, with the rubric description and score criteria shown alongside. Raters could zoom into individual diagrams before scoring visual dimensions. Figure 11: Screenshot of the human annotation website used for expert evaluation. Raters scored 7 dimensions on a 3-level scale after reviewing the problem, gold solution, and generated explanation with diagrams. Rater Training. Before formal scoring, all 20 raters underwent a calibration session. For each of the 7 dimensions, we presented two anchor examples—one high-quality explanation demonstrating exemplary visual and textual clarity, and one low-quality explanation exhibiting common failure modes (e.g., overlapping diagram elements, misaligned labels, or pedagogically ineffective step sequencing). Particular emphasis was placed on the aesthetic dimensions (Element Layout Quality and Visual Consistency), as these proved most subjective. Raters discussed borderline cases together to align their interpretation of the 3-level scale. Pilot Scoring. Following training, raters completed a pilot round on 5 held-out explanations not included in the final evaluation set. Inter-rater agreement was computed via Krippendorff’s α after the pilot. Raters whose individual agreement with the group fell below an acceptable threshold were given additional feedback before proceeding. The pilot Krippendorff’s α across all dimensions exceeded 0.60, confirming sufficient calibration to proceed with formal scoring. Formal Scoring. Each of the 30 explanations in the final evaluation set was independently scored by all 20 raters, producing 30 × 7 × 20 = 4,200 human judgments. Figure 11 shows a screenshot of the annotation interface. Appendix E Full Per-Subset Benchmark Results Tables 10–17 report complete 8-dimension scores for each subject and grade-level subset of the EduIllustrate benchmark. Column headers and scoring follow Table 2. Bold indicates best per column. Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 90.0% 93.4% 79.0% 100.0% 90.8% 86.6% 88.4% 94.8% 89.4% 100% Kimi-K2.5 91.0% 95.0% 74.8% 99.0% 77.2% 71.8% 88.2% 88.2% 83.8% 100% Qwen3.5-397B 93.6% 95.0% 71.8% 93.6% 54.8% 66.6% 83.8% 68.0% 75.0% 95.0% Qwen3.5-122B 90.2% 95.2% 74.4% 96.6% 48.4% 62.8% 87.2% 60.6% 73.0% 95.0% Qwen 3.5-35B 89.0% 91.2% 70.4% 91.2% 45.8% 57.2% 90.0% 55.6% 69.2% 88.8% GPT-5 66.2% 72.6% 45.2% 74.6% 48.2% 62.2% 91.2% 62.4% 60.6% 97.5% Claude Sonnet 4.5 64.4% 63.0% 52.6% 75.8% 50.2% 67.0% 85.2% 68.6% 62.0% 97.5% Mistral-Large-3 48.8% 45.4% 38.4% 79.6% 28.2% 50.4% 81.8% 45.0% 46.4% 80.0% Mistral-Small-4 51.0% 45.6% 37.0% 69.6% 27.0% 53.4% 82.2% 41.2% 45.2% 75.0% Ministral-3-14B 41.2% 43.8% 35.0% 80.0% 27.6% 54.8% 94.6% 39.6% 45.0% 20.0% Table 10: Results on the Mathematics subset (n=80). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 90.2% 97.0% 80.0% 98.6% 85.4% 84.4% 92.0% 91.0% 88.6% 98.3% Kimi-K2.5 88.6% 94.4% 76.0% 98.0% 67.6% 63.6% 92.0% 82.6% 80.6% 100% Qwen3.5-397B 93.8% 95.6% 76.0% 93.0% 45.2% 56.6% 85.4% 69.0% 72.6% 91.7% Qwen3.5-122B 86.4% 92.6% 73.8% 94.2% 40.8% 55.6% 86.6% 58.8% 68.8% 98.3% Qwen 3.5-35B 86.2% 90.2% 73.0% 88.6% 34.0% 48.6% 87.0% 54.0% 64.6% 85.0% GPT-5 61.0% 69.2% 47.4% 72.2% 43.8% 52.4% 91.0% 67.4% 58.4% 98.3% Claude Sonnet 4.5 60.4% 64.2% 54.8% 72.8% 41.8% 60.4% 88.2% 70.8% 59.8% 96.7% Mistral-Large-3 36.2% 37.6% 32.0% 87.6% 24.6% 44.6% 82.4% 44.2% 42.2% 80.0% Mistral-Small-4 37.8% 36.8% 27.8% 69.0% 22.8% 44.8% 84.2% 38.6% 39.4% 63.3% Ministral-3-14B 40.0% 38.6% 34.2% 78.6% 22.4% 46.4% 88.8% 32.2% 41.0% 23.3% Table 11: Results on the Physics subset (n=60). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 91.8% 98.6% 84.4% 95.6% 80.6% 84.0% 90.4% 89.6% 87.8% 90.0% Kimi-K2.5 86.2% 93.2% 75.8% 93.8% 65.6% 65.8% 90.0% 86.0% 79.8% 96.7% Qwen3.5-397B 86.8% 94.4% 72.4% 88.2% 48.8% 55.8% 85.0% 64.8% 70.8% 96.7% Qwen3.5-122B 80.0% 89.6% 69.6% 89.6% 39.4% 56.2% 85.6% 61.0% 66.8% 96.7% Qwen 3.5-35B 82.2% 94.2% 72.8% 87.8% 34.2% 49.2% 89.6% 55.4% 65.0% 93.3% GPT-5 56.2% 63.0% 47.0% 73.0% 46.8% 55.4% 91.8% 61.4% 57.4% 86.7% Claude Sonnet 4.5 50.4% 57.0% 46.0% 80.0% 36.4% 62.0% 81.2% 65.0% 54.6% 90.0% Mistral-Large-3 41.0% 40.0% 29.0% 78.2% 22.2% 47.8% 89.6% 49.6% 42.6% 73.3% Mistral-Small-4 31.8% 37.6% 29.4% 70.6% 21.2% 46.0% 77.6% 40.2% 38.4% 56.7% Ministral-3-14B 32.0% 24.0% 20.0% 68.0% 20.0% 49.8% 93.6% 36.0% 36.2% 16.7% Table 12: Results on the Chemistry subset (n=30). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 81.4% 93.6% 76.4% 94.2% 91.2% 83.6% 87.2% 91.2% 85.8% 93.3% Kimi-K2.5 75.8% 86.4% 68.6% 94.2% 75.4% 72.8% 84.0% 91.6% 78.8% 93.3% Qwen3.5-397B 80.0% 92.6% 63.4% 85.4% 48.2% 61.4% 74.8% 70.0% 68.2% 100% Qwen3.5-122B 79.4% 90.0% 67.4% 90.0% 40.0% 51.2% 74.0% 53.0% 63.4% 100% Qwen 3.5-35B 78.0% 81.0% 61.0% 79.0% 40.6% 52.4% 85.2% 53.4% 61.6% 70.0% GPT-5 50.0% 57.0% 38.0% 78.0% 43.2% 62.2% 93.2% 66.0% 55.0% 66.7% Claude Sonnet 4.5 39.2% 51.4% 35.0% 83.6% 41.0% 65.8% 80.6% 72.4% 53.2% 93.3% Mistral-Large-3 28.4% 32.6% 25.2% 79.0% 20.2% 41.6% 82.6% 43.6% 37.8% 63.3% Mistral-Small-4 32.8% 28.2% 24.6% 76.4% 21.2% 43.8% 82.2% 37.8% 37.2% 73.3% Ministral-3-14B 20.0% 20.0% 20.0% 80.0% 20.0% 20.0% 100.0% 20.0% 29.0% 3.3% Table 13: Results on the Biology subset (n=30). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 78.6% 91.4% 69.4% 99.4% 84.6% 80.8% 88.4% 94.4% 83.8% 100% Kimi-K2.5 69.6% 84.8% 66.8% 98.0% 64.6% 70.4% 85.2% 83.4% 75.4% 96.7% Qwen3.5-397B 74.6% 92.4% 59.2% 89.2% 43.2% 66.2% 87.6% 65.6% 67.2% 86.7% Qwen3.5-122B 62.4% 85.8% 60.8% 90.8% 37.0% 59.4% 83.6% 64.6% 63.0% 80.0% Qwen 3.5-35B 65.4% 80.0% 60.0% 83.6% 38.4% 60.0% 90.0% 56.0% 61.8% 73.3% GPT-5 40.0% 53.0% 33.0% 76.6% 31.4% 68.2% 97.6% 60.2% 50.4% 76.7% Claude Sonnet 4.5 40.0% 45.4% 34.6% 82.6% 32.2% 63.8% 80.4% 65.0% 49.6% 100% Mistral-Large-3 28.4% 35.8% 24.2% 83.2% 20.0% 52.0% 83.8% 42.8% 39.0% 63.3% Mistral-Small-4 27.2% 29.6% 23.2% 76.8% 21.2% 53.4% 78.6% 37.2% 37.0% 83.3% Ministral-3-14B 25.0% 25.0% 20.0% 75.0% 20.0% 46.4% 93.4% 33.0% 35.2% 13.3% Table 14: Results on the Geography subset (n=30). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 84.0% 91.0% 73.0% 100.0% 86.8% 85.8% 90.4% 96.4% 86.8% 100% Kimi-K2.5 81.0% 89.0% 68.0% 99.0% 69.6% 75.2% 86.6% 93.6% 80.0% 100% Qwen3.5-397B 91.0% 97.0% 68.0% 88.0% 59.0% 76.6% 86.4% 70.4% 76.2% 100% Qwen3.5-122B 88.0% 97.0% 71.0% 96.0% 51.2% 65.8% 87.4% 65.4% 74.2% 100% Qwen 3.5-35B 92.0% 97.0% 69.0% 90.0% 56.4% 60.8% 90.6% 61.6% 73.4% 100% GPT-5 70.6% 80.0% 49.4% 83.2% 59.0% 73.4% 91.6% 74.2% 68.6% 95.0% Claude Sonnet 4.5 65.0% 69.0% 55.0% 85.0% 45.6% 72.6% 85.8% 69.6% 64.2% 100% Mistral-Large-3 47.2% 40.0% 41.4% 84.2% 25.4% 60.8% 89.0% 45.6% 47.2% 70.0% Mistral-Small-4 40.0% 40.0% 30.6% 74.2% 24.8% 61.0% 83.8% 37.8% 42.4% 85.0% Ministral-3-14B 20.0% 46.6% 20.0% 86.6% 20.0% 33.4% 92.4% 33.4% 35.2% 15.0% Table 15: Results on the Elementary School subset (n=20). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 87.2% 94.8% 79.6% 98.8% 88.2% 84.8% 88.2% 93.2% 88.0% 96.2% Kimi-K2.5 84.8% 93.0% 73.2% 97.6% 69.4% 69.6% 88.6% 83.6% 80.2% 99.0% Qwen3.5-397B 91.0% 94.8% 71.8% 92.2% 50.0% 60.8% 82.0% 69.8% 72.6% 95.2% Qwen3.5-122B 85.2% 94.8% 73.8% 94.0% 43.0% 56.6% 83.4% 60.0% 69.2% 94.3% Qwen 3.5-35B 86.6% 90.6% 72.8% 88.8% 37.6% 53.2% 88.4% 53.2% 65.8% 86.7% GPT-5 56.0% 64.6% 43.0% 75.6% 43.4% 59.2% 93.4% 65.4% 57.2% 91.4% Claude Sonnet 4.5 53.6% 58.0% 45.2% 79.8% 40.2% 64.2% 83.6% 69.4% 56.8% 94.3% Mistral-Large-3 42.0% 45.2% 33.8% 82.0% 25.2% 47.4% 83.6% 45.8% 44.4% 76.2% Mistral-Small-4 38.6% 38.4% 31.2% 73.4% 23.6% 50.0% 83.2% 42.0% 41.2% 68.6% Ministral-3-14B 35.0% 33.8% 30.0% 77.6% 27.6% 56.6% 95.6% 42.6% 42.8% 15.2% Table 16: Results on the Middle School subset (n=105). Text Quality Visual Quality Model Correctness & Completeness Logical Coherence Pedagogical Effectiveness Typographic Clarity Diagram–Problem Alignment Element Layout Quality Visual Consistency Text–Diagram Coordination Overall Success Rate Gemini 3.0 Pro Preview 89.0% 95.4% 78.0% 97.4% 86.8% 84.2% 90.4% 91.4% 87.8% 98.1% Kimi-K2.5 86.2% 92.0% 74.8% 96.8% 73.6% 66.8% 89.0% 87.4% 81.4% 97.1% Qwen3.5-397B 85.6% 93.6% 69.2% 90.6% 46.4% 59.8% 84.6% 65.2% 70.4% 91.4% Qwen3.5-122B 80.4% 88.2% 68.6% 92.6% 40.8% 57.8% 85.2% 57.8% 67.0% 94.3% Qwen 3.5-35B 77.8% 85.4% 65.4% 86.4% 37.6% 52.4% 88.4% 55.4% 63.6% 78.1% GPT-5 59.6% 66.2% 44.0% 71.0% 42.4% 56.2% 90.8% 59.8% 56.4% 86.7% Claude Sonnet 4.5 54.8% 57.4% 48.6% 73.8% 44.4% 62.2% 84.6% 68.0% 57.6% 97.1% Mistral-Large-3 36.2% 34.6% 29.4% 81.6% 24.0% 45.6% 81.8% 44.0% 40.8% 74.3% Mistral-Small-4 40.8% 37.0% 29.4% 69.4% 23.6% 46.0% 79.8% 37.4% 39.8% 69.5% Ministral-3-14B 42.0% 38.0% 33.4% 76.2% 21.6% 46.4% 90.0% 30.2% 40.6% 20.0% Table 17: Results on the High School subset (n=105).

Instant research discovery