PAPER PLAINE

When engineers test the ground to predict earthquake damage, they often deploy multiple measurement arrays at different sizes to probe deeper. This paper solves a longstanding problem: how to properly measure the uncertainty in those velocity measurements when data comes from mismatched array sizes. The fix works well for both simple and complex soil structures, producing reliable uncertainty estimates even when the raw velocity values themselves are uncertain.

Researchers automated the process of shrinking and optimizing the artificial intelligence that controls nano-drones, cutting memory use in half and speeding up the drone's decision-making by 1.6 times. The optimized system let a Crazyflie nano-drone fly twice as fast as before, avoid obstacles more sharply, and navigate turns—all while using less than 2% of the drone's power budget.

AI agents waste enormous amounts of computing power on simple tasks by re-reading files and dependencies they've already seen, treating a one-line code edit like a full codebase audit. Researchers developed E3, a method that makes agents estimate task difficulty first, then expand their search only if something goes wrong—cutting costs by 85% and file inspections by 92% while maintaining 100% success rates on code-editing tasks.

When charged particles emit photons, they don't have a single well-defined energy boundary—instead they fade out gradually at the edge, following a mathematical power law. Rey shows this fuzzy edge comes directly from soft photons we can't measure individually, and that the exact shape of this edge encodes information about how those unmeasured photons behave.

Scientists have developed a new method called the Ensemble Controlled-flow Filter that updates forecasts from dynamical systems when observations are complex, indirect, or only accessible through simulation. Unlike traditional filtering methods that assume observations are clean and straightforward, this approach works with messy real-world measurement mechanisms—including those that produce multiple possible outcomes or require running expensive computer simulations to interpret.

A validated system for automatically sorting teacher feedback comments by topic and sentiment, built in 2019, still works well in 2026 even with newer AI models. Surprisingly, the newest frontier models don't improve sentiment classification over much cheaper alternatives, and the system transfers successfully to English feedback, suggesting the original protocol is robust across time and languages.

When an IoT device starts behaving suspiciously, security systems can now detect it and boot it from the network in about a third of a second—without needing specialized equipment. The system uses standard protocols already installed in most networks, combines three different detection methods into a single model, and can permanently revoke a device's access while also disconnecting its active session.

Researchers built an AI system to schedule work in factories that build house modules, where long waits for concrete to cure and paint to dry create bottlenecks that traditional scheduling ignores. The AI reaches within 4% of the best possible schedule and outperforms both standard scheduling rules and genetic algorithms, even as factories get more congested.

When countries sever trade relationships, the pain is rarely equal. Using data from the entire global supply network in 2022, researchers measured how much economic damage each country would suffer from breaking ties with major partners—and found stark imbalances: the United States has leverage over all its trading partners, China over all but one, and countries on the economic periphery like Belarus face devastating losses from severing ties with neighbors while barely denting those neighbors' economies. The asymmetry stems not from trade imbalances but from a country's position in the global network: central hubs can easily find alternatives, while dependent nations cannot.

Drug-discovery AI models perform much worse than expected when tested on chemically unusual molecules, even when using supposedly rigorous evaluation methods. Researchers found that standard testing approaches hide these "structural frontier" failures — inflating measured accuracy by 87% to 130% — and that adding penalty-based safeguards to training doesn't fix the underlying problem.

Small vision-language models can match larger ones on multilingual image questions through test-time scaling, but not in the way researchers expected. The real bottleneck isn't choosing between different reasoning paths—it's simply giving the model enough tokens to finish writing its answer. Fixing how the model formats its response and doubling the thinking budget from 1,000 to 2,000 tokens recovered most gains, while elaborate verification methods added almost nothing.

Researchers tested 13 different methods for measuring how mirror-symmetric an image is, comparing traditional computer vision techniques against modern deep learning approaches. Deep learning won on harder tasks, but a classical method called HOG came surprisingly close while running 300 times faster on standard computers—suggesting that for practical symmetry measurement, the speed advantage of classical methods may outweigh deep learning's modest performance gains.