Common Sense in Large Language Models

This book provides the first comprehensive scientific examination of the structural divide between massive knowledge acquisition and the fluid "common sense" reasoning required for human-level intelligence in machines. Written for researchers, artificial intelligence practitioners, and graduate students, this volume offers a roadmap for building capable intelligence by considering the strengths and limitations of large language models through the important lens of machine common sense-a grand challenge since the founding days of AI in the 1950s. It serves as an essential reference for advanced seminars seeking to bridge the conceptual gaps between fields like cognitive science, the history of computing, and modern-day AI. It also suggests ideas and frameworks for developing the next generation of reasoning-capable machines.

While modern large language models (LLMs) like ChatGPT and Gemini appear deceptively capable in dialogue, they still occasionally fail at the intuitive physical and social tasks that define human common sense. This text synthesizes decades of research-from the symbolic logic of the Cyc project many decades ago to modern neuro-symbolic hybrids-to provide a conceptual grounding for both LLMs' failures and successes in achieving major milestones in machine common sense.

A central focus of the work is the growing crisis in evaluating machine common sense. As LLMs increasingly "solve" benchmarks through surface-level pattern matching and data contamination, traditional performance metrics have become dangerously misleading proxies for intelligence. This text offers a rigorous critique of current testing paradigms, exposing how issues like human annotation noise and statistical artifacts can create a false sense of progress that masks systemic brittleness. It then advocates for a new science of machine common sense that prioritizes rigorous, theory-driven reasoning over raw accuracy scores.