Common-sense reasoning

1 Background and Definition

1.1 Historical roots in philosophy and cognitive science

The concept of common-sense reasoning traces back to ancient philosophy, where Aristotle distinguished between *nous* (intuitive understanding) and *episteme* (scientific knowledge). In the 18th century, the Scottish School of Common Sense (Thomas Reid, Dugald Stewart) argued that certain fundamental beliefs—such as the existence of external objects or the reliability of memory—are innate and necessary for everyday functioning. In the 20th century, cognitive scientists such as Jean Piaget studied how children develop naive theories of physics and psychology, while Marvin Minsky and John McCarthy brought the problem into artificial intelligence, famously noting that machines lack the “obvious” knowledge that humans take for granted.

1.2 Distinction from formal logic and domain-specific reasoning

Common-sense reasoning differs from formal logic in its tolerance for uncertainty, exceptions, and context. Formal logic demands monotonic deduction (conclusions never retracted), whereas common sense relies on defeasible reasoning—e.g., “Birds fly” holds unless the bird is a penguin. It also differs from domain-specific expertise (e.g., medical diagnosis or chess) because common-sense knowledge is broad, shallow, and largely implicit. A person knows that a dropped glass will break without needing a physics degree.

1.3 Role in artificial intelligence

In AI, common-sense reasoning is considered a crucial step toward human-like intelligence. Many early AI systems, such as those for natural language understanding, failed because they could not grasp everyday implications—e.g., that “John put the book on the table” implies the book is now on the table, not under it. Modern AI still struggles with this, making common-sense reasoning a long-standing “hard problem” in the field.

2 Types of Common-Sense Knowledge

2.1 Physical common sense (e.g., gravity, object permanence)

This includes knowledge about how the physical world behaves: objects fall when unsupported, liquids conform to containers, solid objects cannot pass through each other, and small objects can fit inside larger ones. Such knowledge is often learned by infants through interaction and is sometimes called “naïve physics.”

Social common-sense covers norms of human interaction: you should greet acquaintances, thank someone for a favor, avoid interrupting, and reciprocate help. It also includes understanding of intentions, beliefs, and emotions—e.g., if someone looks at a clock and sighs, they might be impatient. This overlaps with “theory of mind.”

2.3 Temporal and causal common sense (e.g., sequences, cause–effect)

People understand that events unfold in time: if you turn a key, a car may start; if you drop an egg, it breaks. Temporal common sense includes ordering (breakfast before lunch), duration (a shower lasts minutes, not seconds), and causality (smoking causes cancer, but correlation is not always causation). Subtypes include:

2.3.1 Naïve physics and naïve psychology

Naïve physics is the intuitive understanding of physical laws (e.g., pushing an object makes it move). Naïve psychology is the intuitive understanding of mental states—beliefs, desires, intentions—that people use to predict behavior. Both are considered core components of common sense.

2.4 Quantitative and qualitative approximations (e.g., “bigger than”, “sooner”)

Instead of precise numbers, common sense uses relative magnitudes: a car is bigger than a cat, a year is longer than a month, a coffee is hot but not as hot as lava. Such qualitative reasoning allows humans to make fast, approximate judgments without measurement.

3 Approaches to Modeling Common-Sense Reasoning

3.1 Symbolic and knowledge-based approaches

These methods attempt to explicitly encode common-sense facts and rules in formal languages, often using logic or frames.

3.1.1 Cyc project and formal ontologies

Started by Douglas Lenat in 1984, Cyc aimed to create a comprehensive knowledge base of everyday common sense, containing millions of assertions and rules. Cyc uses a formal language (CycL) to represent concepts and their relationships, with reasoning engines that apply deduction, induction, and analogy. Despite decades of effort, Cyc remains incomplete and difficult to scale.

3.1.2 Frame representation and scripts

Frames (Marvin Minsky) and scripts (Roger Schank) are structured representations for stereotypical situations. A “restaurant script” includes sequences like entering, ordering, eating, paying. These allow a system to infer missing details—e.g., if someone orders a meal, they intend to eat it. However, scripts are brittle and do not handle novel situations well.

3.2 Statistical and neural approaches

With the rise of machine learning, researchers turned to data-driven methods that learn common-sense patterns from large corpora.

3.2.1 Large language models and in-context learning

Models like GPT, BERT, and T5 have shown surprising ability to answer common-sense questions by learning from vast text. For example, they can infer that “because it was raining” is a plausible reason for someone carrying an umbrella. This “in-context learning” allows them to handle some common-sense reasoning without explicit rules, but they remain unreliable in adversarial cases and can produce absurd answers.

3.2.2 Commonsense knowledge graphs (e.g., ConceptNet, ATOMIC)

Knowledge graphs like ConceptNet (structured as nodes and edges, e.g., “dog – IsA – mammal”) and ATOMIC (focused on everyday inferential knowledge, e.g., “PersonX thanks PersonY – xEffect – PersonY feels appreciated”) provide machine-readable common-sense facts. Neural models can use these graphs to improve reasoning.

3.2.2.1 Graph neural networks for reasoning

Graph neural networks (GNNs) can process knowledge graphs by propagating information along edges, enabling reasoning over paths—e.g., from “John is thirsty” through “thirsty implies desire for drink” to “John wants water.” GNN-based approaches have been applied to benchmark tasks but still struggle with complex multi-step reasoning.

3.3 Hybrid and neuro-symbolic methods

Recognizing the strengths and weaknesses of both paradigms, hybrid systems combine symbolic knowledge bases with neural networks. For example, a neural model might generate candidate inferences that are then verified by a symbolic reasoner. Neuro-symbolic approaches aim for better explainability and reliability while retaining the flexibility of learned representations.

4 Benchmarks and Evaluation

4.1 Winograd Schema Challenge

The Winograd Schema Challenge features sentences like “The trophy would not fit in the brown suitcase because it was too big.” Resolving the pronoun “it” requires common-sense knowledge (the trophy, not the suitcase, was too big). The challenge is designed to avoid statistical cues, making it a test of genuine reasoning. Significant progress has been made, notably by large language models, though some argue they still rely on superficial correlations.

4.2 Physical Reasoning Tasks (e.g., PIQA, PhysicalQA)

PIQA (Physical Interaction Question Answering) presents multiple-choice questions about everyday physical interactions—e.g., “Which tool makes it easier to peel a potato?” PhysicalQA contains similar questions about object properties and forces. These tasks test physical common sense, and performance is often compared to human baselines.

Social-IQ is a dataset of social situations requiring inference about emotions, intentions, and norms. For example, “Alice saw Bob drop his books. What should she do?” Answering correctly requires understanding of prosocial behavior. Such benchmarks reveal gaps in models’ grasp of social common sense.

5 Applications in AI Systems

5.1 Natural language understanding and dialogue

In chatbots, virtual assistants, and question-answering systems, common-sense reasoning improves coherence. For instance, if a user says “I’m hungry,” a system with common sense can infer the user wants food suggestions, not a biology lesson. It also helps resolve ambiguity, such as understanding that “Can you pass the salt?” is a request, not a question about ability.

5.2 Robotics and autonomous agents

Robots operating in unstructured environments need common-sense physics to manipulate objects, avoid obstacles, and anticipate outcomes—e.g., knowing that a cup on the edge of a table might fall if bumped. Social common sense is also essential for human-robot interaction, such as respecting personal space.

5.3 Common-sense in humor and internet culture (e.g., meme interpretation)

Understanding humor often relies on violating common-sense expectations or puns (e.g., “Why did the scarecrow win an award? Because he was outstanding in his field”). Memes require knowledge of social norms, references, and typical scenarios—e.g., a “distracted boyfriend” meme presupposes knowledge of romantic relationships and attention. AI systems that interpret such content must leverage broad common-sense understanding to generate appropriate responses.

6 Challenges and Limitations

6.1 Breadth and open-endedness of everyday knowledge

Common-sense knowledge is vast and unbounded. Any attempt to manually encode it will inevitably miss countless facts and exceptions. The open-ended nature means that models must handle an infinite variety of situations with finite resources.

6.2 Context dependence and ambiguity

The same action can have different common-sense implications depending on context. “John kicked the ball” is normal on a playground but strange during a funeral. Words and phrases are ambiguous: “bat” could be an animal or sports equipment. Disambiguating context requires sophisticated models that often fail at the edge cases.

6.3 Representing stereotypes without reinforcing biases

Common sense often includes stereotypes—e.g., “nurses are usually women.” Encoding such knowledge can perpetuate harmful biases. Researchers must carefully design systems that can use statistical regularities when appropriate but avoid overgeneralization and discrimination. This remains an active area of ethical concern.

7 Future Directions

7.1 Lifelong learning and dynamic knowledge updates

Current common-sense models are static; they do not adapt to new information or changing social norms. Future systems may incorporate lifelong learning, updating their common-sense knowledge from experience or curated updates, similar to how humans learn from new situations.

7.2 Integration with embodied experience

Humans develop common sense partly through physical interaction with the world. Robotics researchers argue that embodied agents—those with sensors and actuators—can ground abstract concepts in sensorimotor experience, potentially leading to more robust common-sense reasoning. For example, a robot that knocks over a cup learns the consequences firsthand.

7.3 Explainable common-sense reasoning

As AI systems are deployed in sensitive applications, the ability to explain inferences becomes crucial. Future work may focus on producing reasoning chains (e.g., “Because the cup is on the edge, touching it could cause it to fall”) that humans can verify. This will build trust and allow debugging of faulty common-sense rules.