Research map / Graph learning / Graph self-supervised learning
GRAB: A Challenging GRaph Analysis Benchmark for Large Multimodal Models
Jonathan Roberts, Kai Han, Samuel Albanie
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 |
|---|---|---|---|
| use multiple-choice format instead of single answer (ablation) | Claude 3.5 Sonnet · GRAB Properties | accuracy (%): 41.8 -> 31.8 | Hurt |
| use multiple-choice format instead of single answer (ablation) | Claude 3 Haiku · GRAB Properties | accuracy (%): 14.2 -> 23.6 | Improved |
| use multiple-choice format instead of single answer (ablation) | LLaVA-1.5 13b · GRAB Properties | accuracy (%): 5.0 -> 22.4 | Improved |
| use multiple-choice format instead of single answer (ablation) | Gemini 1.0 Pro Vision · GRAB Properties | accuracy (%): 20.2 -> 12.7 | Hurt |
| use Seaborn instead of Matplotlib plots (ablation) | Claude 3.5 Sonnet · GRAB 100-question subset | accuracy (%): 30.0 -> 26.0 | No clear effect |
| replace exact matching with LLM answer extraction (ablation) | LMM evaluation · GRAB | accuracy (%) | No clear effect |
| increase task complexity from 0 to 3 functions/series/transforms (ablation) | Claude 3.5 Sonnet and Gemini 1.5 Pro · GRAB Functions/Series/Transforms | accuracy (%) | Hurt |
| require decimal precision instead of integer precision (ablation) | LMMs · GRAB | accuracy (%) | Hurt |
| newer model release (Gemini 1.5 Flash to Gemini 2.0 Flash) | Gemini Flash · GRAB Real task | accuracy (%): 17.1 -> 19.6 | Improved |
| newer model release (Gemini 1.5 Flash to Gemini 2.5 Flash) | Gemini Flash · GRAB Real task | accuracy (%): 17.1 -> 29.4 | Improved |
| newer frontier reasoning models vs older baselines | GPT-5 vs GPT-4o · GRAB-Lite | accuracy (%): 12.6 -> 46.2 | Improved |
| newer models on Transforms task | GPT-5 · GRAB-Lite Transforms | accuracy (%): 6.0 -> 63.0 | Improved |
Closest papers by meaning
- Mosaic of Modalities: A Comprehensive Benchmark for Multimodal Graph Learning (CVPR 2025 · similarity 0.90)
- MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks (ICLR 2025 · similarity 0.87)
- ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models (ICML 2026 · similarity 0.87)
- SO-Bench: A Structural Output Evaluation of Multimodal LLM (CVPR 2026 · similarity 0.87)
- SciFIBench: Benchmarking Large Multimodal Models for Scientific Figure Interpretation (NeurIPS 2024 · similarity 0.87)
- GLBench: A Comprehensive Benchmark for Graph with Large Language Models (NeurIPS 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.
