PAPER PLAINE

Video generation models trained with Mixture-of-Experts—a technique that lets AI systems use different specialized components for different tasks—usually try to balance the workload evenly across all components. This backfires on video: because videos have repeated patterns and uneven semantic content, forcing uniform expert usage fractures related visual information across disconnected components. A new approach called SplitMoE splits experts into two roles—some specialized for high-level meaning, others for flexible visual details—letting the model naturally group similar content instead of scattering it arbitrarily.

When researchers only have access to coarse data about a network—like trade flows between regions rather than between individual countries—they usually can't reliably figure out what the detailed network actually looks like. This paper shows a new method can reconstruct fine-scale networks from coarse information with high accuracy, even outperforming methods that have access to more detailed data during training.

Three different ways of measuring disorder—Shannon entropy, Fisher information, and quantum kinetic energy—turn out to be the same thing: measures of correlation between variables. When you destroy correlations by treating variables as independent, all three measures decrease in predictable ways. This suggests a fundamental principle: given what we know about a physical system, the most justified description is the one with the least correlation among its parts.

Researchers built a single neural network representation that simultaneously detects when graph neural networks are unreliable, adjusts confidence scores appropriately, and maintains accuracy even when data shifts unexpectedly. In head-to-head tests on 14 benchmark datasets, it outperformed competing methods at measuring prediction confidence without needing extra correction steps, and showed the strongest performance when data characteristics changed.

Researchers tested predictions about how signals flow through the fruit fly brain's 124 million synapses and found that information spreads much more slowly than theory predicted—by roughly tenfold. The culprit: a small cluster of eleven neurons in the fly's nerve cord that act as a nearly sealed-off compartment, trapping signals and slowing the brain's ability to integrate information across regions.

Researchers developed a compression technique that lets ultra-sensitive light detectors (SPADs) send 3D distance measurements using a fraction of the usual memory and bandwidth. The method works by converting individual photon timestamps into compact numerical sketches on the fly, rather than building full histograms — achieving 100–1000× compression while actually improving accuracy. The approach even recovers hidden objects partially blocked by camouflage or semi-transparent materials.

Researchers created a neural network that learns to solve a common family of mathematical problems (elliptic PDEs) on shapes that change size and form. Once trained on a single shape, the same network solves the same equation instantly on new shapes and with different boundary conditions, without retraining—a feat previous AI solvers couldn't achieve.

Markets have been rebuilt five times over—from human traders to computer algorithms to AI agents—each shift changing what traders must reveal and what they can hide. At each step, new rules emerge that try to allocate goods fairly and efficiently, but trade-offs persist: speed, secrecy, and control over order sequencing remain in tension, and no design solves them all.

A new statistical method called GUARD lets hospitals update AI models used for clinical decisions without fully retraining them from scratch, using only small amounts of new labeled patient data. The approach combines information from multiple hospitals and accounts for uncertainty in predictions, keeping models accurate over years even as patient populations and hospital systems shift.

Most AI models freeze after training, but the real world constantly shifts. This paper shows how to let models adapt in real-time to new data by learning a low-dimensional "shortcut" space where Bayesian filtering (a statistical technique for updating beliefs) becomes computationally feasible. On wireless signal detection and image recognition tasks, the method adapts faster and more accurately than existing approaches while staying computationally efficient.

Machine learning systems often fail on unusual cases they've never seen before. This paper introduces a new method called Sinkhorn Boundary Outlier Generation (SBOG) that deliberately creates realistic extreme examples—anomalies in time series data, unusual images, and other edge cases—to stress-test AI systems before deployment. Unlike existing approaches that produce random oddities, SBOG creates outliers that stay semantically grounded while pushing into genuinely risky territory, helping developers spot vulnerabilities their training data couldn't reveal.

Researchers discovered exactly how much information about a pending trade can leak from a private transaction pool before attackers can profitably sandwich it. Even when a pool hides the full transaction size, knowing just the lower bound of that size—and the trade's direction—is enough for attackers to reliably profit, while the upper bound turns out to be irrelevant.