Monday, October 19
1,575 new papers appeared. We read every one. These 20 are worth your time.
Today's edition
Picked for broad interest, a real step forward and honest claims, with no field allowed to take over.
Bad Diagnosis: Analyzing Consent and Data Collection Practices of Mobile Medical Apps
We present the first comprehensive analysis of consent and data collection practices across 10,000+ mHealth apps on the Google Play Store, combining static analysis and privacy policy examination.
AI summary
Researchers analyzed over 10,000 health apps from the Google Play Store to study how they handle user consent and collect personal data. They created a strict definition to tell true health apps apart from general wellness programs. Nearly 93% of apps fail to give users a consent form, breaking rules like GDPR. The authors say stronger enforcement of privacy rules is urgently needed for health apps.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useEasy to readGenerative AI Produces and Transmits Beauty Bias
Generative AI can reproduce racial hierarchies and gendered patterns of beauty and alter the context in which real people are judged.
AI summary
The researchers tested 27 text-to-image generators by creating 51,300 faces using simple prompts that never mentioned beauty. Human raters judged outputs from all 23 tested generators as significantly more attractive than real photographs. Groups with darker skin tones tended to receive lower attractiveness scores than groups with lighter skin tones. In a hiring simulation, seeing highly attractive AI faces reduced trust and interview chances for real candidates shown with photographs.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintBroad interestReal-world useEasy to readAtmospheric Chemistry-Climate Modeling with an AI Foundation Model
Here, we present the first foundation model to capture interactions between emissions, atmospheric composition, and radiation.
AI summary
The researchers built the first AI foundation model that links emissions, air composition, and radiation together. They adapted an existing AI model called Aurora using data from a detailed chemistry-climate model named UKCA. The AI model accurately reproduced daily weather, air composition, and radiation, with 23 of 27 predicted fields scoring above 0.85. It runs about 100 times faster than UKCA and stays stable over 50 years of simulated time. The authors note this does not replace process-based models, but shows AI can learn from observations at low computational cost.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useDNA-Encoded Library-Guided Discovery of a First-in-Class Molecular Glue of the NOS1AP-SCRIB Complex in Alzheimer's Disease Patient-Derived Neurons
With an in vitro ADME profile consistent with CNS exposure, NOSTRA is the first molecular glue for a NOS1AP complex, establishing interaction stabilization as a viable NOS1AP pharmacology.
AI summary
Researchers screened a DNA-encoded library of 1.4 billion molecules to find compounds that bind NOS1AP, a protein linked to Alzheimer's disease. They developed NOSTRA, which stabilizes the interaction between NOS1AP and SCRIB, making them stick together 30 times more strongly. In neurons from Alzheimer's patients, NOSTRA reduced toxic amyloid and tau markers and improved how the cells communicate. NOSTRA is the first molecular glue for a NOS1AP complex and can enter the central nervous system.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useMultiscale Quantum-Classical Simulation of Protein Autoproteolysis in a Million-Atom Membrane Environment
These results demonstrate how quantum computers can serve as accelerators within scalable scientific workflows, extending correlated electronic-structure methods to biologically relevant systems.
AI summary
The authors developed a hybrid method combining molecular dynamics, density functional theory, and correlated wavefunction methods. They used CPUs, GPUs, and quantum processing units together to run simulations at multiple scales. They studied ADGRG1 autoproteolysis, a maturation reaction in a receptor, using a system of over one million atoms. The simulation included the full receptor in a membrane and used a 54-qubit quantum processor and up to 1,200 GPUs. The results show how quantum computers can accelerate scientific workflows for biologically relevant systems.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepIonic switching between supermagnetic states
Without resynthesis or continued stimulus, we show, for the first time, a reversible magneto-ionic switching between superferromagnetic (SFM) and superspin-glass (SSG) states using room-temperature Li-ion transport.
AI summary
Supermagnetic states are usually fixed when a material is made. The authors used lithium ion movement at room temperature to switch between two supermagnetic states without remaking the material. Adding lithium creates tiny nickel networks that produce a superferromagnetic state with stronger memory. Removing lithium partially restores nickel oxide and creates a superspin-glass state with sharper memory at specific temperatures. This switching worked through all tested cycles of adding and removing lithium.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health
We developed ECG-SDI, a single-lead electrocardiography (ECG) framework for continuous sleep-depth profiling and digital biomarker extraction.
AI summary
The authors developed ECG-SDI, a system that uses a single-lead ECG to track sleep depth continuously. The system captured sleep depth changes that standard sleep staging misses.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useTrojan Horse: Automated Repository Poisoning against Coding Agents
We present TrojanForge, the first automated framework for repository poisoning.
AI summary
Coding agents can now work on their own, which makes them vulnerable to attacks that poison entire code repositories. The authors created TrojanForge, the first automated system for poisoning repositories. TrojanForge analyzes code structure and execution traces to find the best hiding spots for malicious code. It learns better disguise strategies by testing against target agents and copying successful tactics to new repositories. Tests on 384 poisoned repositories show over 60% poisoning success and over 90% backdoor injection, and even commercial agents fail to defend.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useWater-Energy-Data nexus reveals sustainability gaps and priorities for AI infrastructure
The evidence map identifies a mismatch between the interconnected activities supporting AI infrastructure and boundaries used to assess them.
AI summary
AI growth relies on water, energy, and data systems that can worsen resource insecurity and social inequities worldwide. The authors created a water-energy-data framework and used AI to analyze 12,825 studies across four AI infrastructure activities. They found researchers study operational energy more than water use and overlook connections between locations and life-cycle stages. These gaps hide how resource demands shift between regions and how data benefits compare to their costs.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useMicro-Localization: High-Precision Mobile Robot Localization Using Microscopic Ground Texture
We demonstrate high-precision global localization using natural ground texture viewed through microscope optics, achieving localization in the tens-of-micrometers regime.
AI summary
They tested it on various surfaces with 1,007 queries each, achieving position errors of 28 to 73 micrometers. A second measurement method agreed with their results within about 9 micrometers. The system also tracked tiny movements from thermal expansion on an actual low-cost robot. These findings suggest microscopic surface texture can give mobile robots the precision of fixed positioning systems.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintAligning Information Theory and Formal Verification: A Lean Formalization of Rate-Distortion Theory
This paper presents the first Lean formalization of core components of rate-distortion theory, a central framework for lossy compression and modern machine-learning analysis.
AI summary
This paper creates the first formal computer-checked version of rate-distortion theory using the Lean proof assistant. Rate-distortion theory is a core framework for lossy compression and modern machine-learning analysis. The authors solved alignment problems when translating information-theoretic concepts into a proof-assistant setting. They built a reusable library of definitions and lemmas that matches textbook conventions. This work provides a foundation for future AI-assisted proofs in information theory.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepCode availableSource-Aware Dynamic Pricing with Strategic Buyers
To the best of our knowledge, our study is the first to provide a theoretical and practical framework for incorporating information provenance into dynamic pricing with strategic buyers.
AI summary
Firms use customer data from different sources to set personalized prices, but some data can be manipulated by buyers to get lower prices. The authors created a new pricing method called SALE that uses verified data to guide price exploration and combines both data types to correct for manipulation. The method achieves sublinear expected regret, meaning its performance gap from optimal pricing grows slowly over time. Tests on both simulated and real data show that SALE performs well.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepRTISSLAM-3D: Full 6-DoF SLAM With an In-Air 3D Imaging Sonar
In this paper, we introduce RTISSLAM-3D, a novel 6-DoF SLAM system for a wheeled robot that carries an in-air 3D imaging sonar, an IMU and wheel encoders.
AI summary
Current sonar SLAM systems only estimate 2D movement and ignore height data, which fails on uneven ground like forests. The authors built RTISSLAM-3D, a 6-DoF SLAM system for wheeled robots using a 3D imaging sonar, IMU, and wheel encoders. The system converts sonar images to sparse 3D points, then uses them for motion tracking, place recognition, and loop closure verification. Tested on a 545-meter forest drive, loop closures cut horizontal error from 1.74m to 0.28m, with 0.12m height error. 94.6% of the 3D sonar map stayed within 0.3 meters of a lidar reference map.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useMagnetic tunnel junctions of twisted La2/3Sr1/3MnO3 freestanding bilayers
Here, we demonstrate that the twisted bilayer membranes of La2/3Sr1/3MnO3, a prototypical half-metallic ferromagnetic metal TMO, offer a magnetic tunnel junction (MTJ) with their van der Waals (vdW) gap acting as a tunnel barrier.
AI summary
The authors created twisted bilayer membranes of La2/3Sr1/3MnO3, a magnetic metal oxide. These twisted layers form a magnetic tunnel junction with a van der Waals gap as the barrier. The junctions show tunnelling magnetoresistance based on the relative magnetization direction of each layer. High-field measurements also reveal tunnelling anisotropic magnetoresistance from rhombohedral distortion in the material. These effects suggest that spin and momentum are conserved across twisted junctions of transition metal oxides.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepGatewayBench: Unified Capability and Security Evaluation Platform for MCP Gateways
To address this gap, we introduce GatewayBench, the first unified evaluation framework and benchmark for MCP gateways.
AI summary
AI agents use tools through a standard interface called MCP, and gateways now manage these connections at scale. No one had studied how gateway design affects agent performance and security. The authors created GatewayBench, the first test framework for MCP gateways. It includes 524 tasks and 11,118 real tools from 251 servers, with extra irrelevant tools added to simulate large deployments. Tests of seven gateways showed major differences in task completion, efficiency, and attack resistance, which varied by AI model and setup.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useUltra-stable laser with sub-Hz intrinsic linewidth using a silicon photonic chip
Here, we demonstrate a novel optical self-injection-locking scheme integrated with a modulation-free laser frequency stabilization technique co-integrated with a compact vacuum-gap micro-Fabry-Perot cavity (uFP), with Q-factor of 3.46E9, to realize an ultra-stable laser with sub-Hz fundamental linewidth using a silicon photonic chip.
AI summary
The researchers built a tiny ultra-stable laser on a silicon chip. Their design combines a new optical locking scheme with a small vacuum-gap cavity. The chip reduces a commercial laser's frequency noise by more than 110 dB at 1 Hz offset. This achieves a fundamental linewidth of 58 mHz and a fractional instability of 8.68E-13 at 1 second. The chip occupies only 10 mm² and uses standard silicon photonic fabrication.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepRationally micropatterned antifogging hydrogels counteracting extreme condensation
Here, we present a passive approach, rationally micropatterning an inherently hydrophilic hydrogel to maintain high transparency under supersaturation.
AI summary
Water vapor condensation causes fogging on windows, eyewear, displays, vehicles, and kitchenware. Common antifogging coatings fail under extreme condensation. The authors created tiny patterns on a water-attracting hydrogel to keep it clear. Their design outperforms existing alternatives, even in boiling conditions. The material is easy to make and works well for extreme condensation.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
Peer-reviewed journalReal-world usePerceive, Solve, Reason: Progressive Capability Formation in Time Series-LLMs
In this paper, we build the first such suite, with 337K training and 44K evaluation samples across three progressive capabilities, Perceive, Solve, and Reason, and use it to ask how time series should enter an LLM that must master all three, and in what order.
AI summary
The authors built the first training and testing suite for three time-series skills: perception, solving, and reasoning. They found that a dedicated encoder is essential: with one, their 4-billion-parameter model beats GPT-4o on every perception test. Converting time series to text instead costs 8 to 10 points in reasoning accuracy. Learning the three skills in order works better than mixing the data, and matches training on over three times as much reasoning data. Their 4-billion-parameter models beat four recent 7-billion-parameter models on seven of eight out-of-domain tests, by up to 11 points, without using any target-domain data.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepEnergy-Aware Earth System Science: A Community Roadmap
We propose a skill-energy framework that combines an empirical Pareto front with a ranking score.
AI summary
Earth system models guide climate policy, but their energy use is rarely weighed when judging quality. Model error has dropped 13.5 percent each time computing power doubled since 1995. Cutting CMIP6 error in half would need 28 times more computing power. Some errors improved more than others: sea-level pressure and surface temperature fell 3.5-fold and 2.7-fold, but precipitation only 7 percent. The authors propose a skill-energy framework and call for standard energy and performance reporting before CMIP8.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useOnline Graph Coloring with Random Arrivals
Assuming and are known in advance, we present a randomized exponential-time algorithm that colors any -colorable graph on vertices with colors in expectation.
AI summary
The authors study online graph coloring where vertices arrive in random order. They present a randomized algorithm that colors any k-colorable graph on n vertices with O(k^3 log^2 k log^4 n) colors in expectation. This is the first online coloring algorithm using a polylogarithmic number of colors for all fixed k at least 3. Their exponential-time online algorithm achieves a performance guarantee beyond the NP-hardness threshold for polynomial-time approximation algorithms.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big step