PAPER PLAINE

Researchers created a benchmark of over 6,300 scientific figures from top computer-science papers, each scored by experts on five qualities: clarity, layout, how well captions match images, relevance to the text, and whether the figure misleads. They then built an AI system that learned to evaluate figures the same way humans do, achieving 93% consistency with expert judgment—significantly better than existing image-quality tools, which weren't designed for scientific content.

When customers stop buying without announcement, companies use statistical models to guess who's still active. But the paper shows these models are fundamentally unreliable: different reasonable assumptions produce customer counts that vary by a factor of 7.6 on the same dataset, even when predictions about near-term purchases match closely. The problem isn't bad math—it's that the models extrapolate infinitely far into the future, where verification is impossible.

A study of seven major Bitcoin liquidation crashes found no single early-warning signal that works across all events. Price showed telltale signs of instability before five crashes but completely failed to warn of two sudden news-driven collapses; only a compression in trading order-flow emerged as a consistent (though imperfect) precursor across all events. This suggests crypto crashes fall into at least two different types—those that build up gradually and those triggered suddenly by external shocks—each with its own fingerprint.

Current AI-powered robot controllers plan their moves in chunks and can't adjust mid-action, making them clumsy when things change unexpectedly. Researchers created πR², a new approach that lets robots react to what they sense right now while keeping powerful AI models and planning for multiple actions at once. On real robots, it replans about 4 times faster than existing methods and improves success rates by up to 30%.

Companies often can't tell which AI projects will actually pay off—two projects can look equally promising yet deserve opposite decisions. Researchers at real-estate brokerage Compass showed that a simple framework called expected ROI (eROI) breaks this deadlock by asking three separate questions before building anything: How valuable would it be if it worked? How likely is it to work? And what would implementation cost? This sidesteps the catch-22 that you can't estimate ROI without knowing if a project will succeed, yet can't know without building it first.

Researchers created OPERA, a system that combines multiple specialized AI models to analyze medical images from different scanners and hospitals without needing to retrain on new data. By learning how to route each image to the best-suited expert model during a short calibration phase, then adapting slightly at test time, OPERA maintained high accuracy across 9 different medical imaging datasets—including X-rays, CT scans, and MRI images—without the expensive cycle of retraining for each new setting.

Researchers built a single AI model that can take any mix of inputs—images, text, audio, numbers—and produce any type of output, without needing specialized parts for each task. The model performs competitively with systems built specifically for individual jobs, while remaining flexible enough to chain outputs together or verify its own answers by generating data in multiple formats.

Mathematicians proved that when you repeatedly apply a certain algebraic operation to polynomials, their complexity grows no faster than a specific sequence called the N-bonacci numbers. In special cases, polynomials actually reach this maximum growth rate, showing the bound is tight.

A researcher rebuilt a widely-used model that predicts what note comes next in music, translating it from an outdated programming language into modern Python. The new version exposes how the model stores and retrieves memories, making it possible to inspect, visualize, and modify the internal workings that drive its predictions.

Large language model agents that handle mixed sensitive and public data are vulnerable to attacks that trick them into leaking secrets. A new security framework called APPA lets agents temporarily explore untrusted information in a sandboxed branch of their reasoning, then safely bring back only cleaned results to their main context—blocking data theft attempts that were successful 31–50% of the time while recovering much of the work capacity that traditional security methods sacrifice.

Researchers discovered that distributed AI systems combining fast and slow prediction paths are vulnerable to a new type of attack: an attacker can flood the network with dummy requests to delay the slow, accurate path so badly that predictions arrive too late to be used. In a self-driving car test, 4,000 burst requests increased decision latency from 92 milliseconds to 2 seconds, cutting the accuracy benefit of the slower path in half and reducing object detection quality by 7 percentage points on average—with some rare objects like stop signs losing nearly half their detection accuracy.

A new approach to reservoir computing—a fast machine learning technique for forecasting time series—organizes the internal processing around frequencies instead of random connections. By modeling the system as independent oscillators tuned to different frequencies, the method performs as well as or better than traditional random reservoirs and can actually be optimized to improve short-term predictions, something random reservoirs cannot do.