While advances in Large Language Models (LLMs) are based on optimizing transformer models using predictive coding, current models overlook crucial elements from neuroscience: actions, compositional structure, and episodic memory. This paper proposes integrating these components into foundation models to achieve safe, interpretable, energy-efficient, and human-like AI. By incorporating actions at multiple scales, a compositional generative architecture, and episodic memory, the authors argue that current LLM deficiencies such as hallucinations, lack of grounding, missing sense of agency, poor interpretability, and energy inefficiency can be addressed.

Sequential structure is fundamental to human cognition and behavior (e.g., language, movement, decision-making) and is also central to AI applications. This paper introduces SymSeqBench, a unified framework comprising SymSeq (for generating and analyzing structured symbolic sequences) and SeqBench (a benchmark suite for rule-based sequence processing tasks). SymSeqBench aims to evaluate sequence learning and processing in a domain-agnostic manner while linking to formal theories of computation. Based on Formal Language Theory (FLT), it offers a practical way for researchers across various domains, including cognitive science, psycholinguistics, neuromorphic computing, and AI, to apply FLT concepts and standardize experiments.

The article highlights four key AI research trends for enterprises to watch in 2026: agentic systems, continual learning, world models, and refinement techniques. These trends signal a shift from AI focused solely on intelligence to practical implementation, emphasizing efficiency, scalability, and the ability to productionize AI applications. The article also touches upon the importance of ethical AI, data reliability, and the evolution of AI infrastructure towards greater efficiency.