Research map / Vision-language and multimodal / Multimodal LLMs
GradingBench: Evaluating End-to-End Compositional Reasoning of MLLMs for Automated Exam Grading
Yuting Wang, Zixian Guo, Weihao You, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide reference answers (answer-based instead of answer-free) (ablation) | Gemini-3.1-Pro · GradingBench L1 | accuracy: 79.60 -> 89.74 | Improved |
| provide reference answers (ablation) | Gemini-3.1-Pro · GradingBench L3 | accuracy: 70.06 -> 81.78 | Improved |
| give gold bounding box coordinates of the target question (L2) versus autonomous full page (L3) (ablation) | Doubao-seed-2.0-Pro · GradingBench answer-free | accuracy: 68.14 -> 46.30 | Hurt |
| decompose L3 grading into 3 rounds (perception, reasoning, grading) (ablation) | Gemini-2.5-Pro · GradingBench L3 answer-free | accuracy: 50.94 -> 58.40 | Improved |
| decompose L3 grading into 3 rounds (ablation) | Qwen3-VL-235B-A22B · GradingBench L3 answer-free | accuracy: 47.42 -> 58.48 | Improved |
| decompose L3 grading into 3 rounds (ablation) | Gemini-3.1-Pro · GradingBench L3 answer-free | accuracy: 70.06 -> 74.38 | Improved |
| decompose L3 grading into 3 rounds (ablation) | Qwen2.5-VL-72B · GradingBench L3 answer-free | accuracy: 23.33 -> 19.44 | Hurt |
| run localization in isolation instead of within the integrated L1 prompt (ablation) | Gemini-2.5-Pro · GradingBench L1 | mAP: 29.34 -> 40.74 | Improved |
| supervised fine-tune Qwen2.5-VL-7B on ~10K grading examples | Qwen2.5-VL-7B · GradingBench L3 answer-free | accuracy: 8.38 -> 52.65 | Improved |
| supervised fine-tune Qwen2.5-VL-7B | Qwen2.5-VL-7B · GradingBench L1 | mAP: 0.93 -> 84.81 | Improved |
| supervised fine-tune Qwen2.5-VL-7B (reasoning accuracy still behind Gemini-3.1-Pro) | Qwen2.5-VL-7B · GradingBench average answer-free | average: 8.20 -> 57.47 vs Gemini 61.27 | Mixed |
Closest papers by meaning
- MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
- MME-SCI: A Comprehensive and Challenging Science Benchmark for Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.87)
- Unveiling the Ignorance of MLLMs: Seeing Clearly, Answering Incorrectly (CVPR 2025 · similarity 0.87)
- BMMR: A Large-Scale Bilingual Multimodal Multi-Discipline Reasoning Dataset (NeurIPS 2025 · similarity 0.86)
- SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.86)
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.
