AI TIPS 2.0 Advances Governance While SCOPE Enhances Planning

Recent advancements in AI are tackling complex challenges across various domains, from operationalizing AI governance with the AI TIPS 2.0 framework to enhancing planning capabilities. AI TIPS 2.0 addresses critical governance gaps in risk assessment, actionable controls, and scalable operationalization, offering a comprehensive approach to trustworthy AI. For planning, an end-to-end framework integrates agentic LLMs with PDDL, enabling natural language specifications to be translated into executable plans, demonstrated on benchmarks like Google NaturalPlan and PlanBench. Architectures for agentic AI emphasize reliability through principled componentization and disciplined interfaces, proposing a taxonomy of agents and design guidance for safety and governance.

In the realm of AI reliability and assessment, RIFT (Reinforcement Learning-guided Intelligent Fault Targeting) offers a scalable methodology for LLM accelerator fault assessment, achieving significant speedups and reduced test vector volume with superior fault coverage compared to traditional methods. ARTEMIS, a new agent scaffold, demonstrates strong performance in penetration testing, outperforming most human participants on a large university network while offering cost advantages, though AI agents still struggle with GUI-based tasks and exhibit higher false-positive rates. For continuous action spaces in planning, a Gaussian Process aggregation method improves Monte Carlo Tree Search performance over existing strategies. Furthermore, SCOPE provides an efficient hierarchical planning approach for text environments by using LLM-generated subgoals for pretraining a lightweight model, drastically reducing inference time compared to LLM-based agents.

Research also delves into the fundamental nature of AI and human cognition. A categorical framework analyzes Large Language Models (LLMs), arguing they circumvent rather than solve the symbol grounding problem. Visual categorization studies compare human and AI performance on degraded stimuli, highlighting parallels and divergences in perception and reasoning, and motivating neuro-symbolic architectures. A cognitive-geometric model frames belief and meaning as linear transformations within personalized value spaces, offering insights into communication, miscommunication, and leadership. In materials science, the MatSci-YAMZ platform integrates AI with human-in-the-loop crowdsourcing to accelerate metadata vocabulary development, aligning with FAIR and open-science principles. Finally, a qualitative study on LLM hallucinations reveals context-sensitive trust calibration, influenced by factors like expectancy, prior experience, user expertise, perceived risk, and intuition, leading to recommendations for responsible LLM use.