Large language models are brilliant at following logical rules and predicting patterns, but they're fundamentally missing one critical ingredient for scientific discovery: the ability to generate a genuinely new idea that has no precedent in their training data. That's the core argument of a position paper accepted to the International Conference on Machine Learning (ICML) 2026, titled "Position: LLMs Can't Jump," which has quietly circulated through AI research communities and sparked serious debate about what kinds of thinking machines can and cannot do.

What Does "LLMs Can't Jump" Actually Mean?

The title works on two levels, and both matter. On the surface, it's a metaphor: language models can't make the creative conceptual "leap" that a scientist makes when a genuinely new idea appears. But the paper's author, Tom Zahavy, argues the metaphor points to something literal and physical. LLMs have no body. They've never felt acceleration, never lost their balance, never sat in a bathtub and noticed water rise as they got in. Without sensory experience, the paper contends, they cannot generate the kind of creative hypothesis that requires embodied intuition.

The paper distinguishes between three types of reasoning that scientists use. First, there's induction: statistical pattern matching across enormous datasets. Language models are excellent at this. Second, there's deduction: formal, step-by-step proof from given premises. LLMs have improved dramatically at this over the past two years, especially with reasoning-tuned models that use extended chain-of-thought. But the third type, abduction, is where the paper says models hit a wall. Abduction is the creative act of generating a new explanatory hypothesis or axiom that didn't already exist in some form in the training data.

How Does Einstein's Equivalence Principle Illustrate the Problem?

The paper's central example is Einstein's path to general relativity, specifically the equivalence principle: the insight that acceleration and gravity are locally indistinguishable. Imagine standing in an accelerating elevator with no windows. The sensation feels identical to standing still in a gravitational field. There's no experiment you can run inside the elevator to tell the two apart. This idea did not fall out of existing physics by logical deduction. Newtonian mechanics didn't imply it; nothing in the symbolic apparatus of physics at the time necessitated it.

Einstein himself described arriving at this insight through a specific, embodied thought experience: imagining himself in free fall, imagining the sensation of weightlessness. The paper frames this as abduction in its purest form: a new axiom generated from lived, physical, sensory experience that has no antecedent in any symbolic system. An LLM, the paper argues, can flawlessly deduce every consequence of general relativity once given the field equations. What it cannot do is generate the equivalence principle itself from scratch, because it has never accelerated, never fallen, never felt weight. It has only ever seen the word "gravity" appear next to other words in text.

During the peer-review process, one reviewer pushed the authors for a simpler example than general relativity. The authors added the Archimedes "eureka" story: stepping into a bath, noticing the water level rise in proportion to the volume of his submerged body, and abducing the principle that buoyant force equals the weight of displaced fluid. Like the equivalence principle, buoyancy isn't a deduction from prior axioms. It's a new causal relationship proposed on the spot, from embodied sensory feedback.

What Does the Paper Propose as a Solution?

The paper's proposed fix isn't a bigger language model or a better prompt. It's physically grounded world models. World models are AI systems trained to predict what happens next in the physical world, not just the next word in a sentence. Give one a scene and a hypothetical action, and it shows you what happens next in 3D space, over time, with physics intact.

The distinction matters because world models learn from sensory experience in a way language models cannot. A world model trained on video of robots, vehicles, and physical interactions learns to simulate cause and effect in space and time. It develops an internal sense of how objects behave, how forces work, how bodies interact with environments. This is the kind of grounding the paper argues is necessary for abduction.

How Do World Models Differ From Language Models and Vision-Language-Action Models?

The landscape of AI systems designed to reason about the physical world has expanded rapidly. Large language models are excellent at manipulating text: syntax, semantics, predicting the next word across huge amounts of context. But an LLM can only describe what happens when you drop a glass. It has no internal sense of the glass falling, accelerating, hitting the floor, and shattering.

Vision-language-action models (VLAs) are trained to predict what a robot should do in a given situation, learning from labeled robot demonstrations. But they're narrowly specialized. A world model, by contrast, learns what happens when an action is taken, not just what action to take. This means it can learn from a much wider range of data: egocentric video, third-person footage, even non-robotics internet video. Because a world model is learning how the world evolves rather than mimicking one "correct" action, it can extract more general principles.

What Are the Key Players Building World Models Today?

World models have moved quickly from being a research concept to an active product in the last two years. Several major labs are setting the pace, each with a different angle on the technology:

• Google DeepMind: Focused on real-time interactive environments, building the Genie series, including Genie 3, which generates playable, persistent world environments in real time.
• World Labs: Founded by Fei-Fei Li, building models like Marble that generate consistent, navigable 3D environments with emphasis on spatial intelligence and 3D generation.
• NVIDIA: Developing the Cosmos platform, open-weight world foundation models aimed at giving robotics and autonomous vehicle teams a simulation base to build on.

Each of these organizations takes a different bet on where the value sits: DeepMind is optimizing for interactivity, World Labs for spatial consistency, and NVIDIA for open infrastructure that teams can build on.

How Do World Models Actually Work Under the Hood?

Strip away the branding, and every world model solves three core problems: seeing, remembering and predicting, and acting. The model needs a compressed, usable representation of what's in front of it, pulled from camera frames, LiDAR, or other sensors. This is the raw input layer, and its quality sets a ceiling on everything downstream.

From that perception, the model builds an internal representation of the scene that persists and updates over time, then uses it to forecast what happens next. This is what makes it a model of the world, rather than a single-frame classifier. Predictions aren't generic. They're tied to a specific hypothetical action: if the gripper closes now, if the vehicle steers left, if the leg plants here. That conditioning is what turns a passive prediction engine into a planning tool.

Processing full-resolution video frame by frame is too slow for a robot that needs to react in real time. This is why architectures like Joint Embedding Predictive Architecture (JEPA) map raw visual input into compact, simplified representations instead. This approach ignores irrelevant noise like a flickering screen or a moving shadow, focuses on object affordances like where an object can be gripped, and runs the math faster because predicting a shift in a compact space takes a fraction of the compute that pixel-perfect video rendering would need.

Where Is the Real Bottleneck in World Model Development?

The biggest constraint on world model quality isn't the model architecture anymore. It's the quality and diversity of the data behind it. Before a world model can learn anything, its training data has to be filtered, annotated, deduplicated, and organized at scale, then converted into the compact tokens it actually trains on.

This step isn't glamorous, but it's the single biggest lever on model quality. A world model trained on poorly curated data will confidently predict outcomes that don't make physical sense. The data curation and tokenization stage sets the ceiling on model quality. Poor data means confidently wrong predictions. Pretraining then builds the model's baseline understanding of physics and space using massive, general data. Post-training is where real competitive advantage gets built, through fine-tuning the generalist model on proprietary, domain-specific data.

How Should Researchers and Engineers Approach World Model Development?

For teams looking to build or deploy world models, the practical implications flow directly from understanding what these systems can and cannot do. Here are the key considerations:

• Prioritize Data Quality Over Architecture: Invest heavily in data curation, annotation, and deduplication before scaling model size. A smaller model trained on clean, diverse data will outperform a larger model trained on messy data.
• Combine World Models With Other AI Systems: World models increasingly act as infrastructure around vision-language-action models, generating training data, running evaluations, and simulating outcomes before a policy ever touches a real robot.
• Choose the Right Prediction Architecture for Your Use Case: Autoregressive transformers generate the future frame by frame, strongest for sequential decision-making and long-horizon planning. Diffusion transformers start from noise and denoise the entire scene at once, strongest for visual fidelity and generating rich, realistic synthetic environments.

Does This Mean AI Can't Do Science?

The ICML paper is careful to clarify what it is and isn't claiming. It's a position paper, not an empirical proof. Position papers are judged on whether they make a clear, well-supported, falsifiable argument, not on whether they settle the question. The paper argues that one specific creative step, the abduction of genuinely new axioms, is out of reach for language models alone. It does not argue that AI has no role in scientific discovery.

During peer review, all three ICML reviewers rated the paper as Accept or Borderline Accept, and all raised their scores after the authors' rebuttal. One reviewer pushed back on whether better prompting, context engineering, or scaffolding might unlock the same kind of leap. The authors' response drew a sharp line: prompting and context engineering improve reasoning within an already-existing symbolic space. A better prompt can help a model find a proof it already had the pieces for, structure a deduction more reliably, or retrieve a relevant analogy from training data. What no amount of prompting supplies is sensory grounding, the actual physical experience that generated Einstein's or Archimedes' hypothesis in the first place. You can't prompt your way into having a body.

The paper's position is that abduction of genuinely new axioms is bottlenecked on grounding, not on elicitation technique, and those are different problems with different fixes. For researchers and engineers building AI systems, this distinction matters. It suggests that the path to AI systems capable of genuine scientific discovery may run through embodied, physically grounded models rather than through larger language models alone.