PAPER PLAINE

When Bitcoin and Ethereum futures markets are about to change character—from stable to stressed, or vice versa—the current state of the order book is far more predictive than what's actually being bought and sold. A model based only on order-book snapshots beat interpretable trading models, and adding trade flow helped only modestly for Bitcoin and inconsistently for Ethereum, suggesting market microstructure models should start with liquidity state, not order flow.

A two-agent system won a competition for answering trivia questions with both text and images by learning to judge its own confidence. The key insight: instead of trying to be right about everything, one agent learned when to answer risky questions fast (Tossup rounds) while the other focused on getting the exact answer correct when time wasn't critical (Bonus rounds)—achieving the highest overall score without needing large model ensembles or search tools.

Physicists discovered that certain quantum detectors fail catastrophically when they cross the edge of a black hole, producing mathematical infinities that can't be resolved. The breakdown happens specifically when detectors measure nonlinear properties like momentum or energy density, suggesting the popular "firewall" model of black hole boundaries may be fundamentally incompatible with how these detectors work.

Researchers have discovered how to plant hidden backdoors in neural networks that are mathematically invisible — even when someone inspects the entire trained model with complete access. Someone with this backdoor can instantly break the network's security, while people trying to attack the model without the backdoor would need longer than the age of the universe to do the same thing.

Researchers created SLORR, a lightweight method that nudges neural networks toward simpler, more compressible structures during training—without requiring expensive mathematical operations or architectural changes. When tested on image recognition and large language models, SLORR let researchers compress models by significant amounts while keeping performance intact and adding less than 1% to training time.

Raft is a popular system for coordinating data across multiple computers, but it assumes failures are honest mistakes—it can't handle nodes that deliberately forge elections or tamper with data. Researchers added a trust-and-reputation layer that tracks each node's behavior, penalizes dishonest actions, and blocks untrustworthy nodes from leadership. In testing with 40% malicious nodes, the system kept bad leaders below 5% while losing less than 10% speed.

Mathematicians have proven sharp bounds on how well stochastic methods can approximate solutions to convex optimization problems—the kind used constantly in machine learning and engineering. The work uses a technique called radial dominance to show exactly how the error shrinks as you add more random samples, and proves these rates are the best possible.

Researchers created a new dataset and video generation model that teaches AI systems to solve reasoning problems by generating sequences of video frames rather than text explanations. The model, called Wan-CoF, substantially outperformed existing video generators on four reasoning benchmarks by learning from diverse examples of visual problem-solving and using special tokens to track reasoning across frames.

When governments run anti-poverty programs, the benefits often spread beyond the direct recipients to their neighbors and social connections. Researchers typically guess at how far these spillovers reach, but this paper shows how to let the data reveal the correct distance and shape instead. Applied to two major poverty-reduction programs, the method confirmed some previous estimates but rejected others—and the corrected distances produced substantially different estimates of how well the policies actually worked.

Researchers compared two approaches to finding price manipulation opportunities in financial markets: a traditional method that assumes it knows how prices work, and an artificial intelligence agent that learns patterns from raw data. For moderately volatile markets, the AI agent discovered profitable manipulation strategies using limited training data and actually outperformed the traditional method, even though the traditional method started with correct assumptions about how markets function.

Researchers created MulTTiPop, a collection of 572 pop music segments with matching digital note-by-note transcriptions, to measure how well AI systems can automatically convert recorded music into written musical notation. When tested on the best existing AI models, the results showed significant room for improvement—the top performer only correctly identified 38% of note onsets, the moment each note begins.

Researchers created DrugGen-2, an AI model that designs drug molecules by considering both the disease being treated and the specific protein target involved. When tested on diabetic kidney disease, the model generated drugs with stronger predicted binding to their targets than existing approved treatments like enalapril, while also producing more diverse and chemically valid candidates than earlier AI systems.