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.