pipette
ESEspañol

Tuesday, September 22

2,844 new papers appeared. We read every one. These 20 are worth your time.

Each dot is one paper. The pink ones made today's edition. Point at any dot to see its title.

Today's edition

Picked for broad interest, a real step forward and honest claims, with no field allowed to take over.

  1. The DNA Legacy of Thomas Jefferson

    Using these DNA data, we find clear evidence of Thomas Jefferson ancestry in descendants of both his wife Martha Wayles Jefferson and Sally Hemings.
    AI summary

    Researchers developed a new technique to recover DNA from rootless hair samples in a museum collection. The DNA from these hairs matches Thomas Jefferson based on their origin, mitochondrial genomes, and Y chromosome markers. The study found clear evidence of Thomas Jefferson ancestry in descendants of both his wife and Sally Hemings. The amount of shared DNA is more likely under direct Thomas Jefferson paternity than other proposed scenarios.

    Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from

    PreprintBroad interest
  2. Exact Realization of Deep ReLU Computation in a Microscopic Statistical-Mechanical System

    Here we establish an exact statistical-mechanical realization of arbitrary ReLU networks.
    AI summary

    The authors built a physical system that exactly copies how deep ReLU neural networks compute. They started from tiny particle arrangements and their possible states, then built a partition function without fixing a neural activation rule. At very low temperature, this system's measurable properties exactly match the hidden states, outputs, and loss of any deep ReLU network. The straight-line segments of ReLU networks emerge from a sudden phase change in the tiny system. This gives a precise physical basis for neural computation and a new physics-based view for designing neural architectures.

    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
  3. Monet: Measuring the Ecosystem of Open-Source Text-to-Image Models Tailored for Harmful Services

    In this study, we present the first systematic, ecosystem-level measurement of Monets.
    AI summary

    Open-source text-to-image models can be modified for harmful purposes, creating what researchers call Monets. The authors built a taxonomy of ten harmful categories and found 23,947 Monets across eight major platforms. About 40.76% of Monets are mirrored across multiple hubs, and 11.99% remain accessible after being banned. Some Monets are part of commercial campaigns, including one with 668 models and 914 paid commissions. The findings suggest platform-specific defenses are insufficient and call for cross-platform cooperation.

    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 use
  4. LLM-Based FORM Code Generation with Verification-Driven Fine-Tuning

    Fine-tuning a compact open-weights model (Qwen3-8B) with quantized low-rank adaptation (QLoRA) yields a specialist that, evaluated on four complementary benchmarks (840 tasks, single attempt each), decisively outperforms frontier models with up to 756B parameters in execution rate and in strict, FORM-verified output matching on the larger benchmarks, and remains statistically indistinguishable from them on the smaller, harder ones.
    AI summary

    FORM is a specialized language used in particle physics to handle extremely large algebraic equations from complex calculations. Current large language models fail completely at writing working FORM code on their first try without help. The authors created 4,633 training examples by using the FORM program itself to check and validate the code. After fine-tuning a small open model, it beat much larger frontier models at producing correct FORM code. The model kept its general reasoning and coding skills, dropping only 2.6 percentage points.

    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
  5. High-resolution Nitrogen Dioxide Maps Reveal Exposure Limit Breaches across Europe

    We reconstruct the daily compliance statistic across those regions and find limit breaches in 91 EU air quality zones deemed compliant by the regulatory monitoring network, covering a population of approximately 135M.
    AI summary

    The authors mapped nitrogen dioxide across Europe using a machine-learning model that combines ground monitors with satellite and other data. The maps show hourly pollution levels at 10-meter resolution for 110 metropolitan areas. They found that 91 EU air quality zones considered compliant by official monitors actually broke the daily nitrogen dioxide limit. About 135 million people live in these zones, and an additional 20 million live in other areas where the daily limit is breached. The maps could help assess whether areas meet the new 2030 EU daily limit.

    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 use
  6. Single-nucleus profiling of social behaviour in a mutualistic reef fish

    Here, we present the first single-nucleus transcriptome atlas of the telencephalon in the cleaner wrasse Labroides dimidiatus, a remarkable model for advanced social cognition.
    AI summary

    The researchers created the first map of gene activity in individual brain cell nuclei from the cleaner wrasse, a reef fish known for complex social behavior. They analyzed about 67,000 nuclei from the telencephalon, the fish equivalent of the forebrain. Fish that were actively interacting socially showed expansion of specific support cells and learning-related gene networks. Fish that were only watching social interactions showed gene patterns linked to motivation and readiness to act. The findings may reveal brain cell types and molecular mechanisms that govern complex social behavior across all vertebrates.

    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
  7. Visualizing micro- nanoplastics in the human brain: Early evidence for roles in microvascular pathology

    In summary our results demonstrate that synthetic polymers exhibit detectable fluorescence, and particles with similar spectral properties are visible in histologic sections.
    AI summary

    The authors examined brain tissue to see if tiny plastic particles were present. They found ten different plastics using chemical analysis. Fluorescent particles appeared in all brain samples. Particles matching polyethylene and polypropylene lined blood vessel walls. These findings may help researchers study links between plastics and disease.

    Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from

    PreprintBroad interest
  8. Ferroelectric switching of odd-parity magnon spin splitting

    Here we report room-temperature transport evidence for ferroelectric switching odd-parity magnon spin splitting in multiferroic BiFeO3.
    AI summary

    The authors found room-temperature evidence that voltage can switch a special spin pattern in magnons, which are spin-carrying waves in insulators. They studied multiferroic BiFeO3, where a spiral magnetic pattern splits magnon branches with opposite spin in an odd way under momentum reversal. An injected in-plane spin current created an out-of-plane magnon spin component that matched theoretical predictions and reversed with ferroelectric switching. They used this voltage-controlled magnon spin to switch a perpendicular ferromagnet without an external magnetic field and demonstrated XNOR logic-in-memory. The results extend odd-parity spin splitting from electrons to collective bosonic modes and achieve nonvolatile electrical control at room temperature.

    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 use
  9. Gliders on Aperiodic Monotilings: Cellular Automata on the Hat and Spectre

    Evolutionary search then discovers gliders on both monotilings; tracked by a sliding window that regenerates the patch along the flight, they travel one million rings at constant speed and heading.
    AI summary

    The hat and spectre are newly discovered shapes that tile a flat surface without repeating patterns. Researchers studied cellular automata on these tilings using computer-generated patches that can be rebuilt from small records. Simple rules produced only short-lived moving patterns, not true gliders. With more complex rules using vertex neighborhoods and visible states, evolutionary search found gliders on both tilings. These gliders travel one million rings at steady speed and direction, following six specific paths aligned with the tiling's structure. Both rule extensions were required to find the gliders.

    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 available
  10. AI-written admissions essays are widespread but penalized

    These improvements, however, came with an apparent AI penalty: Applicants submitting AI-written essays were admitted less often than comparable non-users.
    AI summary

    In 2025, most applicants submitted at least one mainly AI-generated essay, even though this was explicitly banned. After ChatGPT became available in November 2022, essay writing quality improved. However, applicants who submitted AI-written essays were admitted less often than similar non-users. In an experiment, admissions officers could often recognize AI writing and rated suspected AI essays lower than human ones.

    Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from

    PreprintBroad interestEasy to read
  11. ORION-CMR: On-scanner Reporting with Integrated Foundation Model for End-to-End Cardiac MRI Analysis and Interpretation

    We present ORION-CMR (On-scanner Reporting with Integrated fOunda-tioN Model), the first clinically evaluated scanner-native end-to-end CMR foundation model.
    AI summary

    Cardiac MRI scans give detailed heart information but are rarely used because they are complex to perform and read. The authors built ORION-CMR, an AI system that runs directly on the scanner and handles the entire process from images to report. The results show that real-time automated cardiac MRI analysis on the scanner itself is possible.

    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 use
  12. Characterizing Wildlife Response to Biomimetic and Conventional Underwater Vehicles

    We present the first dataset comparing fish disturbance in response to a conventional thruster-driven AUV, a sea turtle inspired flipper-driven AUV, and a diver.
    AI summary

    Scientists compared how fish react to three different approaches: a conventional propeller-driven robot, a turtle-like flipper-driven robot, and a human diver. They tested these at a Caribbean reef biodiversity hotspot and used computer vision to track fish behavior changes. Both robots caused measurable changes in fish behavior. The turtle-like robot generally showed better results, but the difference was not statistically significant after correction for multiple comparisons. Fish showed larger responses to the diver than to the robots, though this comparison is exploratory and based on limited diver data.

    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 use
  13. NAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware

    We present NAVIR, an end-to-end AVSR system targeting the BrainChip Akida neuromorphic processor, which natively supports only sequential two-dimensional convolutional inference.
    AI summary

    Voice control in factories struggles because loud noise ruins audio-only speech recognition. The authors built NAVIR, a system that reads lips and hears speech for the BrainChip Akida neuromorphic chip. The system uses separate modules for video frames, video timing, and audio spectrograms, fused by a small predictor. On noisy test data, NAVIR reached 14.0% word error rate versus 22.5% for audio-only, and 98.6% command accuracy on industrial commands. The system uses roughly 5 times less energy than a Raspberry Pi and over 100 times less than a laptop GPU.

    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 use
  14. ZIL: Zero-shot Image-to-LiDAR Registration

    We propose ZIL, the first foundation model for zero-shot non-synchronized image-to-LiDAR registration.
    AI summary

    Image-to-LiDAR registration finds the camera position of a photo relative to a 3D point cloud made by laser scanning. Current methods struggle with images and point clouds from different times and fail in new situations without specific training. The authors created ZIL, the first model that works without prior training on specific scenarios for this task. ZIL uses Vision and Point Transformers to process inputs and predicts 3D coordinates to improve accuracy without extra labels. Training on 1.4 million frames from 7 datasets, ZIL reduced position and angle errors by up to 87% and 76% compared to prior methods.

    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 useCode available
  15. BATOseq-PE: A public batoid DNA library with the first molecular insights into the cryptic diversity, genetic health, and management of Peruvian marine rays

    To bridge this gap, we developed BATOseq-PE, the first public genetic reference library for Peruvian marine batoids, covering 14 species sequenced in vitro from local specimens, complemented by a novel in silico assembled mitogenome for Notoraja martinezi.
    AI summary

    The authors created BATOseq-PE, the first public genetic library for Peruvian marine rays, covering 14 species. The library adds 73 new DNA barcodes and includes first public records for several high-priority ray species. Using genetic diversity data combined with fishing and ecological information, they identified areas where heavy fishing overlaps with extinction risk and low genetic diversity. The authors suggest tailored management strategies based on these data to help conserve ray populations.

    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 use
  16. ClimTip-GML: A global bias-corrected and downscaled dataset for assessing impacts of climate tipping events

    Here, we present ClimTip-GML, the first globally bias-corrected and downscaled climate dataset for impact assessment of large-scale tipping scenarios, comprising eight key variables at 0.25{\deg} spatial resolution from three general circulation models (GCMs): CESM1-CAM5, HadGEM3-GC31-MM, and MPI-ESM1-2-HR.
    AI summary

    The authors created ClimTip-GML, a new global climate dataset for studying tipping events like Amazon rainforest dieback or Atlantic Ocean circulation collapse. It contains eight climate variables at 0.25 degree resolution from three models, covering 100-year simulations with and without tipping scenarios at 2°C warming. They used generative machine learning trained on reanalysis data to correct biases and sharpen resolution while keeping physical consistency across space, time, and variables. Validation showed much lower biases, better small-scale patterns, improved correlations between variables, and consistent long-term responses to forcing and tipping events. These results allow better assessments of Amazon and Atlantic tipping impacts, which can directly guide mitigation and adaptation policies.

    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 use
  17. The Choreographic Genome: Amplifying the Silent Structure of Text into Dance

    We present an embodied visualization instrument that amplifies not what a text means, but how it is built.
    AI summary

    Current text-to-motion systems only use the meaning of words, discarding everything else. The authors built a tool that turns the hidden structure of text into dance. They first organize dance movements into 256 style categories, then map raw text bytes onto these categories. Physics routines smooth the result into fluid full-body movement. Tests with poems, code, and non-Latin scripts show each text creates a unique dance.

    Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from

    PreprintBroad interest
  18. InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality

    We present a reinforcement learning framework that trains insertion policies entirely in simulation for direct deployment without real-world demonstrations or policy fine-tuning.
    AI summary

    The authors built a reinforcement learning system that trains robots to do precision insertion tasks entirely in simulation. The robot uses fingertip force feedback and target poses to find the right alignment and correct its movements. The trained policies work on real robots with gaps as small as 0.02 mm and handle different hole shapes without retraining. One policy trained only on a simulated hexagonal task succeeded 95.0% of the time on eight new real-world tasks. The system scored 20/20 on ManipulationNet's peg-in-hole benchmark using fully autonomous motions.

    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 use
  19. Direct imaging of sub-parsec binary SMBHs in the central galaxy of the NGC 5044 group

    The 4.9, 8.4, 15.2, and 23.6 GHz VLBA images reveal two compact, flat-spectrum, and variable radio cores separated by just 0.38 parsec at the heart of the galaxy, forming the tightest directly imaged binary AGN known by an order of magnitude.
    AI summary

    Binary active galactic nuclei are two supermassive black holes in one galaxy that are both pulling in gas at the same time. Scientists used the Very Long Baseline Array to observe NGC 5044, the central galaxy of a galaxy group. They found two compact radio cores separated by just 0.38 parsec, making this the tightest directly imaged binary AGN known. This discovery suggests the older outbursts may have been powered by jets from the secondary core, without any reorientation of the main jet. Massive elliptical galaxies at the centers of clusters appear to be good places to search for binary supermassive black holes.

    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
  20. Lifespan EEG Reference Charts Reveal Aperiodic Confounds in Beta-Band Biomarkers Across Neurological and Neuropsychiatric Disorders

    In this paper, we establish the first comprehensive lifespan reference charts for beta oscillations (N = 22,094, ages: 1-100 years) using complementary analytic approaches that isolate periodic oscillations from aperiodic background activity.
    AI summary

    The authors created the first comprehensive reference charts for beta brain waves using data from 22,094 people aged 1 to 100. Beta wave strength follows a three-phase pattern: it rises through childhood and adolescence, peaks near age 50, then falls in older age. When properly separated from background brain activity, beta frequency shows a previously unknown dip during adolescence, bottoming out around ages 12-13. Most patient groups showed reduced beta power, but schizophrenia showed a unique pattern: lower raw beta yet higher adjusted beta when background activity was removed. The findings suggest that unmeasured background brain activity systematically distorts beta measurements and can reverse how they appear to change with age.

    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