Researchers stretch LeCun's JEPA AI into a universal world model

JEPA-Anything extends LeCun's architecture to seven different fields

The team applied the same core design to seven distinct areas of study. These fields include physics, robotics, weather forecasting, and biology. The researchers tested the model on fluid dynamics, single-cell data, and clinical records. They also examined locomotion planning for simulated robots. The goal was to see if one shared principle could handle all these tasks. Standard models usually need a separate design for each new domain. This approach tries to use a single underlying structure instead.

The team trained the model on different types of data for each field. They did not change the fundamental architecture across any of the seven areas. The only changes involved how they prepared the input data before feeding it in. This allowed the same prediction modules to work on everything from fluid flow to gene activity. The researchers believe this flexibility is key to future AI systems.

The model splits predictions into four orthogonal factors instead of one

JEPA-Anything divides the predicted state into four separate parts. Each part has its own dedicated prediction module within the system. These modules are designed to capture different aspects of the world state. They do not learn the same thing as each other during training. An added constraint forces them to focus on unique features rather than duplicates. The model then reassembles these four partial predictions into a complete picture.

Standard JEPA models funnel everything into a single prediction output. This often causes easy patterns to drown out harder ones inside the system. By splitting the task, the new method avoids this specific weakness. It allows complex interactions to emerge without being lost in noise. The factors remain orthogonal, meaning they are mathematically independent of each other. The system learns which factor handles position and which handles velocity automatically.

Performance gains on physics tests versus pixel dynamics comparisons

The team compared JEPA-Anything against a standard baseline model directly. They used identical data and training conditions for both versions to ensure fairness. Dynamic systems showed the clearest performance gains in these tests. In a simplified Pong environment, prediction error dropped by 35 percent significantly. The model handled unseen interventions with an additional 13 percent drop in error.

JEPA-Anything beat the baseline across ten different test tasks overall. These tasks spanned physics, robotics, and weather forecasting simulations. On the Burgers equation benchmark, error fell by nearly half during evaluation. The advantage held over 50 prediction steps but shrank to about three percent later. The method scored best in water, quartz, acetaminophen, and benzene simulations. It maintained performance even after 100 prediction steps in those cases.

Its biggest gains appear on physical flow equations specifically. Its edge over standard JEPA is barely measurable for pixel dynamics tasks. In single-cell data analysis, the model assigned cell types more reliably. On clinical data, it predicted more than 1,000 possible disease events slightly better. The difference on image tasks remained small compared to other fields. For locomotion planning, JEPA-Anything won in two of three simulated walking environments.

Discovery of a liver cancer treatment candidate in lab and mouse tests

The team's boldest claim comes from their liver cancer research project. They analyzed partial predictions learned from biological data including gene activity. Protein levels and CRISPR screens provided the input for this specific analysis. The top candidate paired IL-18 with blockade of the enzyme CD73. IL-18 is a signaling molecule that activates immune cells in the body. Tumor cells use CD73 to suppress nearby immune responses effectively.

The team tested this combination on liver cancer cells co-cultured with immune cells. They used organoids and tumor tissue from three patients for these tests. In mice, they also ran similar experiments to check safety and efficacy. The combination killed more tumor cells than either IL-18 or CD73 blockade alone. T cells and natural killer cells showed stronger activation with the pair.

The model-proposed combination of IL-18 and CD73 blockade killed the most cancer cells. It activated the most immune cells in organoids and tumor fragments. The study does not establish whether this could become an actual therapy yet. Researchers need to move from simulation to clinical trials before approval. This remains a hypothesis for future drug development efforts only.

Unexpected discovery of Kepler's third law from simulated orbital data

In a second case, the researchers trained the model on simulated orbital data. They did not give the model any physical quantities during this training phase. The learned patterns turned out to be a near-exact match for Kepler's third law. This law states that bodies on larger orbits move much more slowly over time. The law sets orbital frequency at orbit size to the power of minus 1.5 mathematically.

The model landed on an exponent of minus 1.4991 instead of exactly 1.5. This is a very close approximation for such a complex physical system. The team only evaluated one training run during this specific experiment. They picked the single run with the lowest error margin found. No human intervention guided the model toward this specific result directly.

Why it matters for cost, speed, quality, and safety in AI systems

This work matters because it shows a single architecture can handle diverse domains. It reduces the need to build separate models for every new field of study. Engineers might save time by reusing one system instead of many specialized ones. The method filters out irrelevant details while keeping important patterns intact. This could improve quality in areas where standard models struggle with noise. Safety becomes easier if the model explains its partial predictions clearly.

Cost savings come from fewer training runs needed for new applications. Speed increases as data preparation remains consistent across different fields. Quality improves when complex interactions do not get lost in a single prediction stream. Safety gains occur if the model avoids overfitting to specific noise patterns easily. The ability to generalize suggests more robust agents for real-world tasks.

What to do with the publicly available code and models

The team made the code and models publicly available for researchers to use. You can find the repository on GitHub or similar platforms online. Start by reading the paper to understand the four-factor prediction structure. Try training a small version of the model on your own dataset first. Compare its results against standard JEPA implementations directly. Look for orthogonal factors that emerge naturally during your training process.

Check if your data fits the requirement for partial state decomposition well. If you work with robotics, test it on locomotion planning tasks specifically. For biology, examine how it handles gene activity or protein level data. Monitor prediction errors over many steps to see if the advantage holds. Do not assume the model captures real cause-and-effect relationships automatically. Validate its outputs against known ground truth data carefully before deployment.

Source: https://the-decoder.com/researchers-stretch-lecuns-jepa-ai-into-a-universal-world-model-that-works-from-physics-to-biology/

Reflection's Beam beats Chinese models with less compute

Reflection released Beam, an open-weight model that matches GLM 5.2 and Qwen 3.8 on benchmarks while using three to four times less compute.