PAPER PLAINE

Quantum computers can break certain types of encryption far faster than classical computers when given hints about the answer—but nobody had a reliable way to measure exactly how much faster. This paper provides the first practical method to calculate the quantum speedup precisely, showing it can reach nearly 4 times faster in realistic cryptographic scenarios, compared to the theoretical maximum of 2 times previously thought possible.

Researchers showed how to create quantum computers that naturally remember recent information without external help—a critical feature for processing time-dependent data. By carefully controlling how quantum states decay in arrays of neutral atoms, they achieved computational memory with just 5 atoms and demonstrated the system can solve complex real-world prediction problems like forecasting chaotic time series.

Researchers discovered a way to hijack AI agents that rely on the Model Context Protocol by making fake tools look attractive and trustworthy. On standard benchmarks, the attack method achieved a 93.6% success rate at tricking agents into using attacker-controlled tools, and could inflate computational costs by over 30 times while stealing data or corrupting the agent's reasoning.

A new method lets groups of decision-makers work together effectively even when each person has incomplete information, nobody knows how the system works ahead of time, and updates reach the team with delays. The approach doesn't require a central coordinator—each member independently learns the situation and computes their own strategy, yet still converges on decisions nearly as good as if they had perfect information and central planning.

Mathematicians have improved a classic image-denoising technique by using a logarithmic correction that activates only where image gradients are soft and diffuse. The new method eliminates the blocky "stair-step" artifacts that plague standard denoising while preserving sharp edges and keeping computational costs reasonable.

When researchers test whether AI models can correctly call tools (like calculators or databases), the results depend as much on the serving software as on the model itself. Different serving platforms reject or format the same request differently, and measurement errors—like miscounting failures as model mistakes—can artificially inflate or deflate reported success rates by up to 55 percentage points.

Researchers trained artificial intelligence systems to automatically identify and outline five key structures in high-speed videos of the larynx—the vocal folds, arytenoid cartilages, epiglottis, aryepiglottic folds, and the opening between the vocal folds—achieving over 95% accuracy. The breakthrough was applying this method to natural connected speech rather than just sustained vowels, where tissue movement and video quality make the task far harder.

Researchers built SAiFE-gym, a simulation tool that lets computer scientists train artificial intelligence agents to make trading decisions in concentrated liquidity markets—a newer type of crypto exchange where liquidity providers can flexibly adjust where they deploy their capital to maximize fees. The tool decomposes these complex markets into manageable parts and runs fast enough to train reinforcement learning models, and tests show AI agents can learn effective strategies even when market conditions are unpredictable.

When scientists need to compare experimental groups, the way they define what counts as the same—a culture, a treatment, a reference point—changes everything. Researchers tested an AI component called Jev that makes these semantic choices before calculations happen, comparing it against five other approaches across ten real scientific scenarios. Jev matched the best alternatives at picking the right definitions, but the team found that wrong choices in one area could silently change the numbers downstream without breaking the final answer—a hidden risk.

Researchers combined two AI foundation models to identify eight different hyperkinetic movement disorders from video alone, without requiring special markers or sensors. Trained on just 25 people, the system transferred directly to new pediatric and adult datasets with zero false positives for dystonia and near-perfect detection of tremor, outperforming standard keypoint-tracking methods on high-confidence diagnoses.

Researchers developed a way to automatically tune optical sensors without needing to deliberately vary parameters or know how the instrument responds. By watching how natural fluctuations affect measurement accuracy, the system can learn optimal configurations on its own—reaching the theoretical limits of what quantum physics allows for tasks like measuring star positions and separations.

Researchers built a system that combines touch and vision to help robots perform delicate manipulation tasks like in-hand object rotation and manipulation. The system outperformed existing approaches on all six tested tasks, nearly doubling the success rate (70.6 versus 38.0), because it actually modeled how contact forces change over time rather than just sensing them passively.