AlignMerge Advances LLM Safety While MOBIMEM Enables Agent Self-Evolution

Researchers are developing advanced frameworks to enhance AI capabilities and safety across various domains. For LLM merging, AlignMerge ensures alignment preservation by treating merging as a geometry-constrained operation, outperforming existing methods like Fisher soups and SafeMerge in maintaining safety and performance across multiple model families. In agentic AI, MOBIMEM enables self-evolution without retraining through specialized memory primitives and OS-inspired services, achieving significant improvements in profile alignment and task success rates. For AI safety, Prefix Probing offers a lightweight method for harmful content detection with near first-token latency, while QuadSentinel provides a machine-checkable control framework for multi-agent systems using sequents to enforce safety policies.

The integration of AI into scientific research and education is expanding. The TIB AIssistant platform supports researchers across the entire life cycle, from ideation to writing, by providing modular AI assistants and access to scholarly services, aiming for transparency and reproducibility. For education, a system combining topic modeling and LLMs automatically discovers and categorizes AI policies in course syllabi, promoting responsible GenAI use. Furthermore, the concept of "Cyber Humanism in Education" advocates for centering human agency, positioning educators and learners as algorithmic citizens who critically shape AI-enabled learning environments.

New benchmarks and methodologies are emerging to evaluate and improve AI performance. KalshiBench assesses LLM epistemic calibration using prediction market questions, revealing systematic overconfidence across frontier models. PediatricAnxietyBench evaluates LLM safety under parental anxiety and pressure in pediatric consultations, highlighting vulnerabilities to realistic adversarial pressures. For multimodal understanding, AMUSE and RAFT address agentic reasoning in multi-speaker scenarios, improving performance in tasks like dialogue summarization and speaker grounding. In scientific computing, PDE-Agent automates PDE solving using LLM-driven agents and toolchains, while Anubuddhi designs and simulates quantum optics experiments from natural language prompts.

Efficiency and specialized applications are key areas of focus. Small Language Models (SLMs) are shown to outperform LLMs in agentic tool calling through targeted fine-tuning, drastically reducing infrastructure overhead. For generative art, ParamExplorer is an interactive framework that aids in exploring complex parameter spaces using human-in-the-loop feedback. In wireless communications, weighted K-harmonic means clustering (WKHM) offers a stable and interpretable method for fractional user association. For LLM compression, TOGGLE uses temporal logic to formally specify and enforce linguistic properties, enabling efficient deployment on edge devices with significant reductions in computational costs and model size.