WORLD MODELS, NOT LANGUAGE MODELS, MAY DRIVE SCIENTIFIC BREAKTHROUGHS

In a position paper titled "LLMs can't jump," Zahavy outlines fundamental limitations in how language models operate. These systems excel at pattern recognition and text generation but cannot produce genuinely novel insights required for scientific advancement. The core issue: language models work within existing knowledge frameworks, recombining and refining established concepts. True scientific revolutions demand the ability to conceptualize entirely new frameworks and mechanisms. World models represent a different approach. Rather than processing language alone, they build internal representations of how systems behave and interact—essentially learning the underlying rules governing physical or logical systems. This capability could enable AI to propose genuinely novel hypotheses and identify unexpected patterns humans might miss. The distinction matters for AI development priorities. While language models continue advancing incrementally, investing in world model research may be essential for AI to contribute meaningfully to scientific discovery.