Language models excel at statistical pattern recognition and formal derivation, but they fall short in generating novel scientific theories, according to a position paper by Google DeepMind's Tom Zahavy. Zahavy argues that these models lack the cognitive mechanism for 'manipulative abduction,' the creative leap required to formulate new axioms. This gap, he suggests, is what prevents language models from sparking scientific revolutions.
Zahavy bases his argument on a framework proposed by Albert Einstein, who described discovery as a cycle involving sensory experience, intuitive leaps, and logical deduction. He contrasts deduction and induction—reasoning methods language models can already handle—with abduction, which requires inventing explanations for unexpected phenomena. While ordinary abduction is within the models' capabilities, the more complex form of abduction, which involves creating entirely new concepts, remains beyond their reach. Zahavy highlights that systems like AlphaProof, Gemini, and GPT-5 have achieved gold-level scores on mathematical problems, but they cannot formulate the foundational assumptions needed for groundbreaking theories.
Zahavy illustrates the challenge by referencing Einstein's development of general relativity. He argues that AI models, which rely on error signals to adjust predictions, would have struggled to overturn Newtonian physics without observable anomalies. Einstein's 'happiest thought,' where he imagined a freely falling observer, came from embodied simulation rather than pure calculation. Zahavy compares language models to John Searle's 'Chinese Room' thought experiment, where symbols are manipulated without understanding their meaning. He suggests that world models, which allow agents to intervene in simulations and run counterfactual experiments, could provide the necessary feedback loop for scientific innovation.
Source: thedecoder
