PAPER PLAINE

AI models can identify topological relationships like holes and connectivity when asked directly, but struggle dramatically when they have to plan actions in environments where those relationships matter. Even the best-performing model fell far short of human-level performance, and when models generated their own observations to plan with, they frequently violated the very topological rules they were supposed to be reasoning about.

As generative AI systems like ChatGPT answer customer questions, companies want to know whether being mentioned actually increases sales—but there's no standard way to measure this. Researchers created a new statistical method that combines data on how often a company appears in AI-generated answers with information about whether users actually notice those mentions, then traces the link to real business outcomes. The method works by comparing what would happen under different strategies for getting AI systems to feature a brand.

Researchers created GPU-CFR, a compiler that speeds up counterfactual regret minimization—a mathematical technique for finding optimal strategies in complex games—by 80 times on graphics processors. The key insight: for any fixed game, the entire computation pattern stays the same across iterations, so the compiler can record it once and replay it efficiently rather than re-launching thousands of tiny operations each round.

Researchers proved that one type of quantum building block can exist without another, even though the reverse relationship was known. They constructed an oracle—a hypothetical answering machine—where quantum security states survive intact while classical puzzles vanish entirely, showing these two candidates for quantum cryptography are fundamentally different.

Recent middle-distance world records have benefited from a relatively new starting configuration that places some athletes in front groups on outside lanes, reducing the distance they run on tight curves. Using video analysis and physics modeling, researchers found this advantage ranges from 0.1 to over 1 second depending on the event—enough to be the difference between breaking a world record and running a great but unremarkable race.

Researchers developed a way to customize the step-by-step schedules that guide image-generation AI models like DALL-E, by measuring how well each model predicts at different noise levels. The method produced a 38.6% improvement in image quality for one leading approach on standard benchmarks, and surprisingly, the optimal schedules followed similar patterns across different models and training setups.

A mathematician has solved a long-standing question by proving that some infinitely large networks cannot be split into two groups where no group contains all neighbors of any single point—a property called an unfriendly partition. However, the paper also shows that networks with maximum degree four (where each point connects to at most four others) can always be split this way if they have certain structural features like cycles.

AI data centers are demanding electricity faster than power plants can be built, driving up energy prices with no relief in sight. Using mathematical models that account for real-world uncertainties—from construction delays to the risk of building too much capacity—researchers found that even when investors know demand is soaring, they still hesitate to build enough new power plants, because they fear overbuilding and losing money.

Researchers built a conversational system that lets building managers ask natural questions about energy forecasting models instead of staring at technical dashboards. The system correctly understands 94% of questions asked, compared to 76.8% in previous attempts, and energy experts unanimously preferred talking to it over traditional interfaces.

When hospitals share EEG brain recordings for research, simply removing names isn't enough to protect patients—sophisticated analysis can still re-identify them. This study tested whether adding carefully calibrated random noise to EEG features can hide individual identity while keeping the data useful for AI diagnosis tools, finding that the approach works but requires precise tuning to avoid destroying the signal's clinical value.

Italian researchers mapped the entire network of business-to-business relationships using invoicing data from the country's tax office, revealing that the economy follows a "scale-free" structure where a small number of firms are disproportionately important to the whole system. A handful of suppliers and buyers wield far more influence over production flows than most other companies, and firms have easier access to multiple buyers than to multiple reliable suppliers.

When machine learning models trained on one dataset face new, different data in the real world, they often fail — but predicting exactly how much worse they'll perform has proven theoretically elusive. This paper fixes the broken mathematical definitions used to measure these failures and introduces a new method that actually works across different types of problems, allowing researchers to estimate performance drops before deployment.