cs.CYApr 6, 2026
SaveEduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content
Organizations: Shanghai Innovation Institute · University of Science and Technology of China · East China Normal University
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 \ρ\geq 0.83$) while revealing limitations on subjective visual assessment.
Figures & tables
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.
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.
| 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.
Figure 4: Generation protocol overview. Scene 1 undergoes full sequential processing (outline implementation planning code generation); Scenes 2– skip implementation planning and generate in parallel, each conditioned on Scene 1’s code as a visual anchor.
| 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% |
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.
| 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% |
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.
Figure 6: Visual consistency comparison between our sequential anchoring workflow (top) and the all-parallel baseline (bottom) on the same problem (four-pyramid - ). In our workflow, Scene 1 and Scene 2 render the identical pyramid structure—Scene 2 simply highlights the 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.
| Dim. | Spearman | Krippendorff | Reliability |
| 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 |
Table 5: Human-AI agreement and inter-rater reliability (30 explanations, 20 raters). Correctness & Completeness excluded (objective gold answer verification).
| 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 |
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.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
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.
| 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 |
Table 7: Complete topic coverage of the 230 EduIllustrate benchmark problems. Numbers in parentheses after subject names indicate problem count; (× ) after a topic indicates problems on that topic. Some high school mathematics problems have unlabeled topics in the original K12-Vista dataset.
| Model (Generator) | Gemini | GPT-5 | |
| 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). = GPT-5 judge Gemini judge. GPT-5-as-judge exhibits strong self-preference ( ).
| Dimension | Pearson | Spearman | MAE |
| 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 |
Table 9: Dimension-level agreement between Gemini and GPT-5 judges (120 explanations). † Non-significant ( ). Dimensions sorted by Pearson .
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.
| 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% |
Table 10: Results on the Mathematics subset (n=80). Mathematics yields the highest overall scores across all models. Qwen3.5-397B achieves the best C&C (93.6%), slightly surpassing Gemini. All top-tier models exceed 90% on LC, reflecting that mathematical reasoning is well-structured and easier for LLMs to present coherently.
| 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% |
Table 11: Results on the Physics subset (n=60). Gemini achieves near-perfect LC (97.0%) and the highest PE (80.0%) among all subsets, suggesting physics problems elicit more pedagogically structured explanations. DPA scores drop sharply for the Qwen family (34.0%–45.2%), indicating difficulty in rendering force diagrams and circuit layouts accurately.
| 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% |
Table 12: Results on the Chemistry subset (n=30). Chemistry is among the most visually demanding subjects, requiring structural formulas and reaction diagrams. Gemini dominates all dimensions, achieving the highest PE (84.4%) across all subject subsets. Visual quality drops steeply for smaller models: Ministral-3-14B collapses to a 3.3% success rate, and even among successful runs its DPA is only 20.0%, reflecting near-total failure to render chemical structures accurately.
| 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% |
Table 13: Results on the Biology subset (n=30). Biology problems involving anatomical diagrams and biological cycles pose unique spatial rendering challenges. Gemini leads comprehensively with an 85.8% overall score and the highest DPA (91.2%) among all subject subsets. Kimi-K2.5 is a close second (78.8%), while the Qwen family shows a pronounced DPA gap ( 48.2%), suggesting difficulty mapping biological structures to code. Ministral-3-14B achieves only a 3.3% success rate, the lowest across all subsets.
| 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% |
Table 14: Results on the Geography subset (n=30). Geography problems require spatial reasoning over maps and topographic features. Gemini leads with 83.8% overall and the top TDC (94.4%), while Kimi-K2.5 achieves the highest VC (85.2%) among non-Gemini models. The Qwen family maintains comparatively strong LC ( 80.0%) but falls short on DPA ( 43.2%), indicating that spatial layout of geographic elements is a persistent weakness. GPT-5 shows the highest VC (97.6%) among all models on this subset despite a relatively low overall score (50.4%).
| 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% |
Table 15: Results on the Elementary School subset (n=20). Elementary problems cover simple arithmetic and basic geometry and are the easiest grade level for all models. The Qwen family achieves the highest LC scores among non-Gemini models ( 97.0%), and all five top models reach 100% success rate. Despite the simpler content, DPA remains a distinguishing dimension: Gemini leads at 86.8%, while smaller models cluster below 60%, indicating that even simple visual tasks expose rendering capability gaps.
| 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% |
Table 16: Results on the Middle School subset (n=105). Middle School is the largest grade-level subset and spans the widest subject variety. Gemini achieves the highest overall score (88.0%) with strong leads in PE (79.6%) and DPA (88.2%). Kimi-K2.5 is the only other model exceeding 80% overall (80.2%) and reaches the best success rate (99.0%). The Qwen models exhibit a characteristic pattern: strong text quality (LC 90.6%) paired with weak visual alignment (DPA 50.0%), a gap that widens as problem complexity increases from elementary to high school.
| 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% |
Table 17: Results on the High School subset (n=105). High School problems are the most complex grade level, covering advanced physics, chemistry, and mathematics. Overall scores are lower than middle school across the board, confirming that harder content degrades visual generation quality. Gemini and Kimi-K2.5 remain the top two models (87.8% and 81.4% respectively). Claude Sonnet 4.5 achieves the highest success rate (97.1%) on this subset despite a relatively lower overall score (57.6%), suggesting it defaults to simpler but syntactically valid diagrams. The DPA gap between Gemini (86.8%) and the next-best model (Kimi-K2.5, 73.6%) is the largest of any grade level, highlighting a growing visual rendering challenge at higher difficulty.
More in this topic
Generative AI in Education
Generative AI in education uses language and multimodal models to support teaching, learning, feedback, assessment, and educational design.
See all 287 papers