Research map / Multi-agent systems and games / LLM agents
From EduVisBench to EduVisAgent: A Benchmark and Multi-Agent Framework for Reasoning-Driven Pedagogical Visualization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EduVisBench, a 1,154-question STEM benchmark scoring pedagogical visualizations across five rubric dimensions, and EduVisAgent, a multi-agent framework that generates interactive learning webpages, reporting 81.6 average vs 58.2 for the best baseline (v0).
Problem
Foundation models such as diffusion models and LVLMs produce pedagogically weak visual explanations for STEM problems, and no benchmark systematically measures visual pedagogical reasoning.
Approach
A benchmark with a 0-5 rubric over five pedagogical dimensions (context visualization, diagram design, text-graphic integration, thought guidance, interactivity), plus a multi-agent pipeline (task planning, conceptual mapping, reasoning decomposition, metacognitive review, visualization) that builds interactive webpages via the v0 system.
Key results
- EduVisAgent achieves an average score of 81.6 on EduVisBench, versus 58.2 for the best baseline v0 and 54.6 for Claude 3.7 Sonnet (Webpage).
- Diffusion models score lowest, from 13.8 (Flux.1-dev) to 21.8 (SDXL) average.
- GPT-4o judgments show cosine similarity 0.9655 and MSE 0.5702 against human evaluators on 50 samples per subject.
Surprising findings
- Claude 3.7 Sonnet scores 61.2 on Maths Easy with SVG output but only 26.7 on Maths Medium, showing large variance across difficulty.
Limitations
- Evaluation relies on GPT-4o as judge, with agreement to human raters measured on only 50 samples per subject.
- Ablation of individual agents is not reported.
Experimental setup
Datasets: C-MHChem-Benchmark high-school-physics MATH-500 Illustrative Mathematics
Benchmarks: EduVisBench
Metrics: five-dimension rubric score (0-5 each, normalized to 0-100)
Baselines: Flux.1-dev SD3.5 SDXL Claude 3.7 Sonnet GPT-4o Gemini 2.0 Flash v0 Qwen2.5-VL-72B
Builds on: Concrete-Representational-Abstract model FOPS strategy metacognitive theory
Closest papers by meaning
- SketchMind: A Multi-Agent Cognitive Framework for Assessing Student-Drawn Scientific Sketches (NeurIPS 2025 · similarity 0.87)
- Explain with Visual Keypoints Like a Real Mentor! A Benchmark for Multimodal Solution Explanation (AAAI 2026 · similarity 0.86)
- When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.86)
- EduDiag: A Benchmark for Educational Diagnostic Reasoning with Error Tracing and Correction on Large Multimodal Models (CVPR 2026 · similarity 0.86)
- Seeing as Experts Do: A Knowledge-Augmented Agent for Open-Set Fine-Grained Visual Understanding (CVPR 2026 · similarity 0.85)
- ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.