Commonsense Reasoning

Commonsense reasoning is the cognitive ability to apply everyday knowledge about how the world works — understanding physical causality, social norms, typical sequences of events, and implicit assumptions that humans take for granted — to make sense of situations, predict outcomes, and solve problems in ordinary contexts.

What Is Commonsense Knowledge?

• Physical Commonsense: Objects fall down, not up. Water is wet. Fire is hot. Glass breaks when dropped.
• Social Commonsense: People get upset when insulted. You should say "thank you" when someone helps you. Interrupting is rude.
• Temporal Commonsense: You eat breakfast before lunch. Children grow into adults. The past cannot be changed.
• Causal Commonsense: If you don't water plants, they die. Studying improves test scores. Exercise makes you tired.
• Functional Commonsense: Chairs are for sitting. Umbrellas protect from rain. Keys open locks.

Why Commonsense Reasoning Is Hard for AI

• Implicit Knowledge: Commonsense is rarely explicitly stated — "water is wet" doesn't appear in many texts because it's obvious to humans.
• Vast Scope: Commonsense covers an enormous range of everyday knowledge — millions of facts and relationships.
• Context-Dependent: What's "common sense" varies by culture, context, and situation — "it's cold" means different things in Alaska vs. Florida.
• Exceptions: Commonsense rules have exceptions — "birds fly" is generally true, but penguins don't.
• Grounding: Much commonsense knowledge comes from physical interaction with the world — AI systems trained only on text lack this grounding.

Commonsense Reasoning in Language Models

• Modern LLMs have learned substantial commonsense knowledge from their training data — text corpora encode human knowledge and experience.
• Strengths: LLMs can answer many commonsense questions correctly — "Can you fit an elephant in a backpack?" → "No."
• Weaknesses: LLMs still make surprising commonsense errors — especially on questions requiring physical intuition, novel situations, or multi-step commonsense inference.

Commonsense Reasoning Tasks

• Winograd Schema Challenge: "The trophy doesn't fit in the suitcase because it's too big." What is too big? (Requires commonsense about physical size.)
• PIQA (Physical Interaction QA): "How do you cool down hot soup?" → Requires physical commonsense.
• Social IQa: Questions about social situations — "Why did Alex apologize?" → Requires social commonsense.
• CommonsenseQA: Multiple-choice questions requiring commonsense knowledge — "Where would you find a jellyfish?" → Ocean, not desert.

Improving Commonsense Reasoning

• Knowledge Bases: Integrate structured commonsense knowledge — ConceptNet, ATOMIC, etc. — to supplement LLM knowledge.
• Multimodal Learning: Train on images and videos alongside text — grounding language in physical experience.
• Reasoning Chains: Use chain-of-thought prompting to make commonsense inferences explicit — "Why? Because..."
• Few-Shot Examples: Provide examples of commonsense reasoning to guide the model.

Applications

• Dialogue Systems: Understanding user intent and context requires commonsense — "I'm cold" might mean "close the window" or "turn up the heat."
• Story Understanding: Comprehending narratives requires filling in unstated commonsense details — "She opened her umbrella" implies it's raining.
• Question Answering: Many questions require commonsense to answer — "Can fish drown?" → Requires understanding of fish biology.
• Content Moderation: Detecting harmful content requires social commonsense — understanding context, intent, and norms.

Commonsense reasoning is the foundation of human intelligence — it's the vast web of everyday knowledge that lets us navigate the world, and teaching it to AI remains one of the field's grand challenges.