LLMs Can't Make the Abductive "Jump"

By the end of 2024, Google DeepMind's AlphaProof solved three of the five International Mathematical Olympiad problems, including the competition's hardest Hubert et al., Nature 2025. By 2025, systems like Aristotle were producing verified solutions to open research questions in mathematics and physics Achim et al., 2025. LLMs are proving theorems. They're finding new chemical catalysts. They're writing papers that get peer-reviewed.

We've been calling this "AI does science" for months. And it is. But it's only half of it.

The Easy Half: Manipulating Symbols You Already Have

Scientific reasoning has three parts. Deduction takes a set of axioms and derives their consequences; each step guaranteed to be true if the premises are. Induction takes a pile of data and finds the patterns that generalize from it. These are the two halves LLMs have already conquered.

Deduction is what AlphaProof does at the IMO. Given the axioms of Euclidean geometry and a problem statement, it derives the proof line by line. Each line is a guaranteed consequence of the ones before it. Induction is what each LLM does every day: it finds patterns in text, compresses observations, and generates the next most statistically likely token. Both are about manipulating symbols you already have.

Here's what most people miss: induction is the easier problem. Finding patterns in data is what LLMs were built for. Deduction is harder because each step must be logically correct; one wrong line and the proof breaks. But the symbols themselves are given to the model. It just has to navigate the space between them.

This is why the current generation of AI systems excels at science: most published scientific work is already in this zone. You have a framework (quantum mechanics, thermodynamics, graph theory), you have data, and you need to derive consequences or find patterns. LLMs are very good at this. They're already doing it at a level that surprises people who remember LLMs as "chatbots."

The Hard Half: Generating Symbols You Don't Have

In 1907, Einstein was writing a review article about relativity when he had what he called his "happiest thought": an observer falling freely from a house would feel no gravity in his immediate surroundings Zahavy, PMLR 306, 2026. He imagined a physicist in a sealed elevator accelerated through deep space — objects released inside would fall with identical acceleration regardless of composition. From this single thought experiment, he derived the Equivalence Principle. From there, in four years, came General Relativity.

Here's the thing about that story that gets lost: Einstein wasn't given the Equivalence Principle as a starting point. It didn't appear in the data. He invented it.

This is abduction. Charles Sanders Peirce called it the third mode of reasoning: given a rule and an unexpected observation, invent the scenario that explains it. Unlike deduction (which guarantees truth) or induction (which finds patterns), abduction is a creative act that invents a cause for a singular phenomenon Peirce, 2026.

Abduction is the hard half. And it's also the half that matters most. Every major scientific revolution; heliocentrism, germ theory, quantum mechanics, General Relativity; began with someone having a thought like Einstein's falling observer: an observation that didn't fit, followed by an instant connection to something nobody had written about before.

Here's my observation: we keep expecting AI to have "eureka moments" because we confuse the easy half with the hard half. We see AlphaProof solving Olympiad problems and think "AI can do science." But those problems already have their symbols. The real work; generating the symbols in the first place; is still outside the model's reach.

Where It's Headed: Interaction Over Observation

LLMs learn by reading the world. Abduction requires interacting with it. Einstein's falling observer came from a head simulation, not a corpus: he imagined cutting an elevator's cable and asked what would happen.

Current video models show intuitive physics as a side effect of statistical correlation Hassabis, 2025. A falling apple appears because falling is the dominant continuation in training data. DeepMind's Genie shifts this: trained unsupervised from internet videos, it learns an action space that lets you intervene Bruce et al., 2024. Replicating Einstein's elevator means being able to cut the cable, not just watch it fall Pearl & Mackenzie, 2018.

An action-controllable world model acts as a synthetic laboratory. An agent manipulates the environment, observes outcomes, and abduces rules from the feedback Zahavy, PMLR 306, 2026. That interactive grounding lets it propose axioms with no symbolic precedent in the training data.

Foundation models often lock onto structurally flawed representations that fit the data but miss the physics Vafa et al., 2025. Without the right inductive bias, the lab produces the wrong physics. The way forward: stop reading about the world. Start interacting with it.

The Last Cognitive Leap

AI has mechanized two of the three cognitive operations that sustain scientific discovery. Induction finds patterns in data. Deduction derives consequences from premises. Abduction; the leap from observation to hypothesis; remains the last barrier.

Einstein discovered General Relativity not by searching over symbols, but by simulating the sensory experience of a falling observer. The formulation of the Equivalence Principle was a self-contained act of physical abduction, where the premises were established solely through internal simulation, independent of immediate external verification Zahavy, PMLR 306, 2026.

The path to automating this leap is emerging: action-controllable world models that provide a "synthetic laboratory" where agents can run counterfactual simulations and test reasoning beyond standard linguistic training data Zahavy, PMLR 306, 2026. These systems are still in their infancy, but they point toward a future where the abductive jump transitions from mystical insight to reproducible algorithmic process.

AI can prove theorems and find patterns. It can't have a eureka moment — not yet, because it can't imagine what it feels like to fall.