LLMs can’t jump: the abductive gap in AI science

The moment “jump” becomes the bottleneck

Picture the last time you watched a model “solve” something by pattern. A prompt goes in, a plausible answer comes out, and it feels like progress. But science rarely advances by pattern alone. It advances when someone notices a structure no dataset directly hands them—when they leap from messy sensory or experimental reality to a crisp explanatory hypothesis.

A position paper making the rounds in AI-for-science circles argues that modern Large Language Models (LLMs) are excellent at two phases of reasoning—matching patterns (often called induction) and deriving consequences from accepted rules (often called deduction). Where they allegedly fail is the third phase: abduction, the creative move that invents novel explanatory hypotheses. Without that “jump,” the system can become a very good proof machine with a very limited imagination for the premises.

The tricky part is that abduction isn’t “free creativity.” It’s a specific computational role: generating candidates for what the world might be like, especially when evidence is sparse.

Induction and deduction are strong. Abduction is the missing gear.

Let’s define the terms in plain language, because they matter more than the jargon.

• Induction: inferring likely rules from observed examples. In AI, induction often means learning statistical regularities from lots of training data.
• Deduction: using rules to compute what must follow. If you accept premises like “All squares have equal sides” and you know “this figure is a square,” deduction gives “this figure has equal sides.”
• Abduction: inventing the premises. Abduction is the search for an explanation that could make the observations make sense.

LLMs are built to be unusually competent at induction-like behavior (they predict text that continues a pattern) and at deduction-like behavior (they can produce logical steps that follow rules or mimic reasoning). But abduction is not just “generate a new sentence.” It’s “generate a new explanatory model,” the kind of model you could plausibly use to guide further investigation.

That’s why a question-shaped search like “Can an LLM really perform abduction like a scientist?” keeps coming up in communities thinking about artificial scientific invention.

What abduction feels like in real physics

Human scientists do abduction all the time, often before they have enough measurements to “let the data decide.” They build a hypothesis because it explains more than it contradicts and because it fits the world in a coherent way.

Einstein’s path is a canonical example people reach for: the work wasn’t only about fitting known facts, but about proposing a framework that could reshape what counts as the right concepts in the first place. In other words, the “axioms” (the foundational assumptions) weren’t given by a neat supervised dataset.

The position paper’s core claim can be rephrased like this:

• Deduction is a skill LLMs can simulate well.
• Induction is a skill they can approximate through training.
• But abduction requires a mechanism that bridges observations to new premises—and that bridging is exactly where the current systems are argued to be structurally weak.

This is where things get subtle. Many systems can generate “hypotheses.” The real test is whether the generated premises are grounded enough to be systematically testable and reusable—especially when data is scarce.

Why the “premises” step is harder than it sounds

A common misconception is that scientific discovery is “just” a larger version of prediction. If a model can predict the next token, shouldn’t it be able to predict the next theory?

The problem is that premises aren’t passive labels; premises are commitments about structure.

When you abduct, you are choosing:

• what entities exist,
• what constraints they obey,
• and what quantities are meaningful.

Those choices interact with each other. If you pick the wrong primitives, the rest of the reasoning can become consistent while still describing the wrong world.

So the bottleneck isn’t “can an LLM write a new idea?” The bottleneck is “can it generate a globally coherent explanatory framework whose internal dynamics match reality enough to guide further steps?”

That brings us to a phrase that shows up in the position paper: the translation of simulation into formal axioms.

The simulation-to-axiom bottleneck (where LLMs struggle)

Think of simulation as making the world run inside a computer. You pick rules for how objects move and interact, then you observe what happens over time.

Axioms are the formal starting rules of a theory—statements that define how deduction will proceed.

Now imagine the workflow of genuine hypothesis formation:

1. You observe something physical (measurements, trajectories, patterns in experiments).
2. You imagine candidate generative mechanisms (models) that could produce those observations.
3. You notice regularities in the behavior of those candidates.
4. You convert that behavior back into clean statements you can reason with.

That last step is the hard one. Converting “the simulated behavior looks like X” into “therefore the axioms should be Y” is not mere text rewriting. It’s an alignment problem between:

• a messy, numeric, dynamical world inside the simulator, and
• a symbolic, compositional structure outside it.

An LLM is powerful at symbolic surface forms, but the argument is that it doesn’t have the deep mechanism to construct the right premises from sparse evidence in a way that is physically consistent and systematically improvable.

In less poetic terms: LLMs are good at producing candidate deductions, but the jump to the premises that make those deductions about the world is the missing muscle.

Why “data compression creativity” can miss the point

Another idea people often borrow when thinking about creativity is data compression: the notion that the “best” explanation is the one that compresses the observed data efficiently.

Compression can be a useful lens, but it has a failure mode. If the data are limited, “compressibility” may not uniquely point to the correct underlying mechanism. Two different theories can both explain the same small set of observations, while predicting different futures.

Abduction is forced to care about counterfactual structure: what would happen next if the hypothesis were true. Compression alone doesn’t guarantee that.

So the position paper’s complaint is not that inductive pattern learning is useless. It’s that induction-as-compression can’t replace the abductive responsibility of proposing premises that remain coherent under new conditions.

From LLM text to world models: adding the missing sensory grounding

The proposed bridge is physically consistent, multimodal world models.

Let’s unpack those words.

• A world model is a system that learns to represent how the environment behaves—how things change over time.
• Multimodal means using more than one kind of signal: for example, visual data (images), temporal data (video or sequences), and potentially structured signals like motion or coordinates.
• Physically consistent means the model respects constraints from physics rather than hallucinating motion that looks plausible in a purely statistical sense.

The narrative hope is straightforward: if the system can simulate grounded dynamics, it gets a “sensory” substrate analogous to the way humans reason from experiments. Then abduction becomes a matter of extracting stable, general premises from what the model consistently reproduces.

This doesn’t magically make abduction solved, but it changes what the model has access to. Instead of only operating on language regularities, it can search over mechanisms that produce realistic dynamics, and then translate the stable behavior into formal structure.

That translation step is still nontrivial, but at least the candidate hypotheses are being filtered through a world-shaped bottleneck.

A practical mental model for building toward abduction-like behavior

To make this concrete, imagine an “abduction pipeline” that doesn’t pretend an LLM is doing everything.

• Phase A: candidate mechanisms from grounded simulation. A world model generates and tests dynamical hypotheses against observations or constraints.
• Phase B: invariant discovery. The system looks for stable patterns across simulated scenarios—things that remain true when you vary irrelevant details.
• Phase C: formalization into premises. Those invariants are expressed as axioms, constraints, or governing equations that can support deduction.

In that pipeline, the LLM can play a role, especially in Phase C, where symbolic representation matters. But the abductive leap is no longer purely “text from text.” It becomes “world-consistent mechanism to formal premise.”

That distinction is the whole story. Without grounding, an LLM can produce theories that are linguistically elegant yet dynamically arbitrary. With grounding, the search space changes, and abduction becomes less like freestyle writing and more like structured hypothesis testing.

Closing: what “LLMs can’t jump” should teach us

The claim that LLMs “can’t jump” is easy to over-literalize. No one is arguing that these models are incapable of generating new text or novel-sounding ideas.

The sharper point is that scientific invention has a specific computational demand: selecting premises that explain sparse observations and predict new outcomes. That demand is abductive. And abduction, as the argument goes, requires a bridge between simulation-shaped experience and axiom-shaped formal theory.

When that bridge is missing, you get a system that can reason outward from assumptions but struggles to invent the right assumptions in the first place. The path forward, then, isn’t to demand that language models become miracle discoverers. It’s to give them a world to reason within—then let formalization and deduction follow from dynamics the system can actually model.