SciReC: Diagnostic Evaluation of Multimodal Multi-Turn Relational Reasoning with Adaptive Interaction

Relational reasoning requires the process of perceptual understanding, comparing, and integrating the underlying relationships between concepts. This ability encompasses multiple categories, including analogical, structural, and cause-effect, each capturing a distinct aspect of higher-order understanding. To examine the performance of multimodal large language models (MLLM) on these relational inference tasks, we developed SciReC, a model-adaptive multimodal academic dialog benchmark, including visual understanding, exhibiting knowledge, and memory recall questions extracted from open-source textbooks as the relational reasoning process involves multiple representations. Claude 4.6 achieved the best performance on the overall relational score with 73%, followed by GPT 5.4 with 68%. Performance trends indicate that open-source models achieve their lowest scores on spatial relations, while proprietary models struggle more with hierarchical and sequential relations. Across domains, model performance is lowest on Astronomy and highest on Psychology.

Dataset Description

This dataset contains multimodal question-answering examples grounded in textbook figures. Records in the figure-grounded configurations are filtered to include only examples whose referenced image files are present in this release.

• Curated by: Nilay Yilmaz
• Language(s) (NLP): English
• License: cc-by-nc-sa-4.0
• Contact: nyilmaz3@asu.edu

Dataset Sources

• Website: https://scirecc.github.io/SciReC/
• Repository: https://github.com/nlylmz/SciReC
• Paper : Diagnostic Evaluation of Multimodal Multi-Turn Relational Reasoning with Adaptive Interaction

Configurations

• visual: 13791 figure-grounded visual questions with resolved images.
• knowledge: 13501 caption/text-grounded knowledge questions with resolved images.
• relational: 662 cross-figure relational questions.
• memory: 1318 single-figure memory questions from relational examples, without resolved images.
• memory_validation: 1318 validation questions paired with memory questions, sharing the same answer and resolved image.

All configurations expose a single train split.

Loading

from datasets import load_dataset

visual = load_dataset("nlylmz/SciReC", "visual", split="train")
knowledge = load_dataset("nlylmz/SciReC", "knowledge", split="train")
relational = load_dataset("nlylmz/SciReC", "relational", split="train")
memory = load_dataset("nlylmz/SciReC", "memory", split="train")
memory_validation = load_dataset("nlylmz/SciReC", "memory_validation", split="train")

Figure columns contain repository-relative paths to the image files.

Dataset URL: https://huggingface.co/datasets/nlylmz/SciReC

Citation