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

The overall sentiment is mixed and somewhat skeptical of the article's strongest claims. Many commenters agree that present language-only LLMs lack grounding, embodiment, and reliable mechanisms for scientific abduction, but a sizable opposing camp views the paper as speculative, historically shaky, and too narrow because future AI systems can include tools, feedback, and world models. The conversation is mostly constructive rather than hostile, with the strongest agreement around the need for grounding and the strongest disagreement around whether that limitation is fundamental.

In Agreement

• Language is a lossy representation of lived experience, so models trained mostly on text may miss the perceptual and tacit grounding behind human intuition.
• Scientific invention often depends on embodied observation, internal simulation, and experimentation rather than merely recombining written claims.
• LLMs can associate words with sensory concepts such as taste or place without having the direct experience that gives those concepts meaning for humans.
• Current LLMs have not demonstrated landmark autonomous scientific breakthroughs, which some commenters treat as practical evidence for the article's concern.
• Random sampling or higher temperature is not the same as intuition because intuition requires a meaningful reorganization of concepts and a verification loop.

Opposed

• The paper's Einstein example is historically oversimplified because much of the mathematical and physical groundwork already existed before Einstein's synthesis.
• The claim is too absolute because LLMs can be connected to search, tools, robotics, simulators, and laboratory feedback, changing the system being evaluated.
• Breakthroughs often come from recombining existing techniques across domains, and transformer-like systems may be well suited to that kind of broad analogy.
• Strong performance in math, code, and formal reasoning suggests that language models can build useful abstractions even without human-like sensory experience.
• The framing risks moving the goalposts by judging LLMs against rare human genius rather than practical scientific assistance or incremental discovery.