Is your multimodal large language model a good science tutor?

TL;DR

The paper addresses the gap that multimodal LLMs, while proficient at solving scientific problems, are not guaranteed to tutor effectively. It introduces a rubric-based tutoring evaluation with a simulated student to measure tutoring quality across conceptual clarity, scaffolding, suitability for K–12 learners, and encouragement of reasoning. By identifying strong and weak tutors on ScienceQA and constructing a pairwise preference dataset from their outputs, the authors apply preference optimization methods (DPO, ORPO, SimPO) to improve a weak tutor (Qwen2-VL-2B), demonstrating that tutoring quality can be enhanced independently of problem-solving accuracy. The work shows that tutoring-oriented fine-tuning can yield educationally aligned MLLMs, highlighting a path toward models that function as both problem solvers and effective educational assistants in scientific domains.

Abstract

Multimodal large language models (MLLMs) demonstrate impressive performance on scientific reasoning tasks (e.g., ScienceQA). However, most existing benchmarks focus narrowly on the accuracy of the final answer while ignoring other metrics. In particular, when applying MLLMs to educational contexts, the goal is not only correctness but also the ability to teach. In this paper, we propose a framework that evaluates MLLMs as science tutors using a comprehensive educational rubric and a simulated student model that judges the teaching performance of the tutors. Given a list of candidate MLLM science tutors, we use rubric-based student judgments to produce a range of tutor performance scores, identifying both strong and weak tutors. Using the training section of the ScienceQA dataset, we then construct a data set of pairwise comparisons between the outputs of strong and weak tutors. This enables us to apply multiple preference optimization methods to fine-tune an underperforming tutor model (Qwen2-VL-2B) into more effective ones. Our results also show that strong problem-solving skills do not guarantee high-quality tutoring and that performance optimization-guided refinements can yield more educationally aligned tutor models. This approach opens avenues for building MLLMs that serve not only as problem solvers, but as genuinely helpful educational assistants.

Figures (8)

• Figure 1: In our framework, we adopt an MLLM as tutor and another MLLM as student. Given a science question with visual input, the tutor is asked to give tutorial to help the student better understand the question and provide encouragement as well. As long as the student receives the tutorial, a feedback will be given in the form of a review number along with a remark.
• Figure 2: For a science question with visual input, two tutors provide separate tutorials, after which the student selects a preferred one.
• Figure 3: Comparison of tutoring scores and problem solving scores reveals that while a good problem solving ability is necessary for effective tutoring, it is not sufficient.
• Figure 4: Distribution of tutoring scores across baseline models.
• Figure 5: Distribution of tutoring scores for the Qwen2-VL-2B pretrained model and its finetuned variants.