Brief

Researchers have introduced Sci-Rho, a new multilingual benchmark designed to test the robustness of visual-language models (VLMs) on STEM problems. This benchmark includes over 4,200 problem templates across five subjects and seven languages, generating more than 42,000 unique instances. Evaluations of 17 state-of-the-art VLMs revealed a significant gap between average and worst-case accuracy, with smaller models showing more performance degradation across languages compared to larger, proprietary models. AI

Oregon State University has developed LaTA, a FERPA-compliant local LLM autograder designed to streamline STEM education. This system allows for secure, zero-cost grading on campus hardware without altering existing LaTeX workflows, significantly reducing grading time. Additionally, intelligent CCTV systems utilizing AI are showing promise in urban safety, with one study indicating a 23% reduction in vehicle speeds in Minneapolis. Meanwhile, the Chinese AI ecosystem is gaining attention, with DeepSeek being praised for its collaborative approach, contrasting with Western models. AI

Researchers have developed MAIC-UI, a system designed to simplify the creation of interactive STEM courseware. This zero-code platform allows educators to generate and rapidly edit educational materials from existing documents like textbooks and PDFs. MAIC-UI utilizes a structured knowledge analysis and a generate-verify-optimize pipeline to ensure pedagogical accuracy and offers sub-10-second editing cycles. A study involving high school students showed that MAIC-UI improved learning outcomes and reduced disparities compared to traditional methods. AI

Researchers have introduced STEM, a new framework designed to enhance knowledge graph-based question answering. This approach tackles challenges related to the structural heterogeneity of knowledge graphs and the lack of global perspective in existing reasoning path retrieval methods. STEM reframes multi-hop reasoning as a schema-guided graph search, improving accuracy and evidence completeness in retrieval tasks. AI

Researchers have developed an interpretable machine learning model to identify instances of mechanistic reasoning within student team conversations. This tool analyzes individual utterances and group contributions to output probabilities of students engaging in such reasoning over time. The model incorporates a specific inductive bias designed to align probabilistic dynamics with domain-specific behavior, which experiments show improves generalization and interpretability. AI

Researchers have developed ArguAgent, a generative AI system designed to improve collaborative learning in STEM classrooms. The system uses AI to group students in real-time based on their argumentation stances and quality, aiming to foster more inclusive and productive discussions. Testing with OpenAI models like GPT-4o-mini and GPT-5.2 demonstrated that prompt engineering significantly improved scoring accuracy, with model upgrades providing a smaller boost. Simulations showed ArguAgent achieved its grouping criteria in 95.4% of cases, a substantial improvement over random assignment. AI