The Abduction Gap: Why LLMs Cannot Leap to New Scientific Axioms

This paper argues that LLMs are incapable of the 'abductive jump' required for fundamental scientific discovery because they lack sensory grounding. While current AI excels at pattern recognition and logical proofs, it cannot generate new physical axioms from intuition as humans like Einstein did. The author proposes that developing physically consistent world models is the key to unlocking true artificial scientific invention.

Key Points

• LLMs excel at induction and deduction but lack the mechanism for abduction, the generation of novel explanatory hypotheses.
• Scientific discovery requires an intuitive 'jump' from sensory experience to axioms, a process LLMs cannot replicate due to a lack of physical grounding.
• The case of Einstein's General Relativity shows that major breakthroughs can occur without abundant observational data, relying instead on physical intuition.
• LLMs are limited by their existence within a purely symbolic space, preventing them from originating principles derived from raw physical experience.
• Physically consistent, multimodal world models are proposed as the necessary substrate to enable 'manipulative abduction' in AI.

Sentiment

Skeptical and analytical, with a focus on historical context and the 'moving goalposts' of AI capability definitions.

In Agreement

• LLMs are currently like 'brains in a vat' and require an external environment to sample from and act upon to make real forward progress.
• Intuition in LLMs is effectively 'locked' after training, meaning they can only generate 'new' ideas through generalization of training data or chain-of-thought reasoning.
• The lack of native adjacency resolution and physical grounding makes certain types of problem-solving significantly harder and more sample-inefficient for LLMs.
• There is a potential hierarchy where sensory experience and language development are necessary precursors to abstract abductive reasoning.

Opposed

• The paper's historical premise is flawed; Einstein's work was a physical interpretation of existing math (Lorentz transformations) rather than a jump from nothing.
• Claims that 'LLMs can't' do something have a poor track record, as these limitations are often overcome by scaling or architectural improvements.
• The concept of 'abduction' or 'jumping' is not a well-defined computational class, making the paper's argument more rhetorical than empirical.
• Humans themselves may be quite poor at abstract creative leaps, and we only perceive them as special because we lack better benchmarks.
• Multimodal grounding has not yet proven to be a 'silver bullet' for general reasoning, as evidenced by models like Gemini lagging behind text-heavy models in logic.