Edition of November 19, 2026

  1. At The Knowledge Frontier, we pick out the research papers of the day and explain them simply.

    The web has changed in nature. We no longer search, we scroll. Pages look alike, texts are churned out, and we end up reading a great deal without keeping any of it. Educating yourself there has become a strangely passive affair.

    Here, you do more than read. Every article ends with a quiz: a few questions on what you have just read. When the subject lends itself to it, it is an experiment — you set the values yourself and recover, in turn, the researchers' result.

    @VirScience
  2. Physics

    ATLAS HUNTS FOR JUMPS BETWEEN VACUUMS

    Detector-style display of a particle collision: thin curved glowing tracks burst outward from one bright central point through concentric rings of a cylindrical detector, in blue light.

    The strong force glues quarks together. Its theory predicts something strange. The vacuum comes in several distinct versions, and nature can tunnel between them. These jumps are called instantons. Decades of study have never confirmed them.

    An instanton should spray particles almost evenly in every direction. The ATLAS collaboration looked for that signature at the LHC. It used data from 2015, recorded while collisions were kept deliberately rare.

    A neural network sorted the collisions using five track measurements. In the final selection, 32,789 events remain. The fit attributes about 3,900 of them to a signal. Against background alone, the excess reaches 3.5 sigma.

    So, instantons at last? Not yet. No background model alone describes the data, the team writes. The quarks these jumps should produce, and their mass spectrum, still have to be checked.

    Source: Search for the low-mass isotropic signature of QCD instantons in pp collisions at √s = 13 TeV using the ATLAS detector, https://arxiv.org/abs/2611.22092

    PreprintExperiment❓ Quiz
  3. Earth & climate

    THE WORLD'S CO2 STALLS IN 2026, BY CANCELLING ITSELF OUT

    Three large concrete cooling towers of a coal-fired power station releasing white steam at dusk, beside a highway streaked with long trails of car headlights.

    How much CO2 will fossil fuels and industry emit worldwide in 2026? A team projects 38.1 billion tonnes. That is only 0.10% more than in 2025. Almost a plateau.

    Yet nothing is standing still. China drops by 194 million tonnes, mainly in industry and ground transport. From January to June 2026, new-energy vehicles made up 49.4% of new car registrations.

    India climbs by 187 million tonnes, led by power plants. Coal and lignite supplied 69.5% of its electricity from April to June. Rises and falls almost cancel: 94% of the movement vanishes in the global sum.

    The last three months are still a forecast. With another forecasting method, 2026 even ends slightly down. The sign of such a tiny change depends on how the year is completed.

    Source: Projection of global daily CO2 emissions in 2026, https://arxiv.org/abs/2611.22406

    PreprintData analysis
  4. Space & astronomy

    A RUNAWAY BLACK HOLE NEEDS A WIND TO MAKE WAVES

    A bright, fuzzy elliptical galaxy at lower left, with a long, thin straight streak of light running from its edge across the dark sky to a glowing, almost closed ring at upper right.

    Some giant black holes get kicked out of their galaxies. One candidate, RBH-1, was found with Hubble and confirmed with JWST. It races away at about 1,000 km/s.

    Ahead of it sits an arc-shaped shock wave, one to two kiloparsecs wide. Behind it, a straight trail stretches about 60 kiloparsecs back to its galaxy.

    Two astrophysicists at Northwestern University simulated the scene. Gravity alone barely stirs the surrounding gas. A wind blowing from the black hole is needed. It must still be blowing, or have stopped less than 3 million years ago.

    What drives that wind? The black hole could, while feeding at just 1% of its theoretical maximum. And the straight trail reveals calm gas around it. Its turbulence stays below about 30 km/s.

    Source: Shocks and Trails in the Circumgalactic Medium Driven by Runaway Supermassive Black Holes, https://arxiv.org/abs/2611.21738

    PreprintSimulation❓ Quiz
  5. Life sciences

    WHY WOULD A NETWORK OF SPRINGS NEED TO SLEEP?

    Close-up of a tabletop web of small steel balls linked by thin coiled springs and a few straight rods, resting on a dark surface in deep green light.

    Even jellyfish and hydra sleep, without a centralized nervous system. Could the need for sleep come from learning itself, even without biology? A Syracuse University team put the question to a network of springs.

    The network learns tasks one after another by changing the stiffness of its springs. Pull on one spring, and another must answer with a set tension. With each new task, fewer networks manage to learn the next one.

    The reason is exact. Some springs end up carrying almost all the load, others almost none. Learning capacity plus this imbalance form a fixed budget. What one gains, the other loses.

    Random internal jiggles let each spring sense its role and rebalance. But during a task, the readings are biased. Repair works best in a separate phase, like sleep.

    Source: Why do continual physical learners need sleep?, https://arxiv.org/abs/2611.22070

    PreprintSimulation❓ Quiz
  6. Mathematics

    RANDOM NETWORKS STOP BEING COLOURABLE AT A FIXED DENSITY

    Hand-drawn network of about twenty round dots joined by thin straight lines on dark paper, the dots in teal, ochre and rose, with a few unconnected grey dots around it.

    Take a network of dots and a box of q colours, at least three. Linked dots must get different colours. Now add random links, one at a time. Sooner or later, no colouring works at all.

    In 1999, Achlioptas and Friedgut showed the switch is abrupt. But its position could still wobble as networks grow. The conjecture named after them says it settles at one fixed density.

    Gaia Carenini, at the University of Cambridge, announces a proof. It holds for any number of colours from three up. The key idea: pairs of dots that every valid colouring paints alike.

    One question stays open: where exactly is the threshold? The proof shows it exists, not its value. The preprint has not yet been peer-reviewed.

    Source: A Proof of the Achlioptas-Friedgut Convergence Conjecture, https://arxiv.org/abs/2611.21665

    PreprintTheory❓ Quiz
  7. Physics

    A NEW RULE FOR HOW DENSE A FUSION PLASMA CAN GET

    Inside a vast round reactor chamber lined with dark panels, a thick bright ring of glowing plasma circles a tall central column, with fine wisps of light along its edges, in cold blue light.

    In a fusion reactor, more fuel means more power. Output grows with the square of plasma density. But push the density too far, and the temperature at the plasma's edge collapses.

    To stay safe, operators use a rule of thumb, the Greenwald limit. Yet real limits can stray from it by 50% or more.

    A team led from MIT pooled data from five tokamaks. They hold 325 density-limit events and 4,359 stable discharges. A machine-learning search found a better gauge. It combines edge collisions, edge pressure and the twist of the magnetic field.

    This gauge raises far fewer false alarms. And for ITER and ARC, it points to a wide margin. Both might run denser, and so produce more fusion power, than expected.

    Source: Collisionality scaling of the tokamak density limit, https://arxiv.org/abs/2611.21899

    PreprintData analysis❓ Quiz
  8. Computing & AI

    A LANGUAGE MODEL REWRITES ITS OWN WEIGHTS

    Open desktop computer case showing its circuit boards and fan, beside a monitor whose glowing screen is filled with columns of small numbers, a few brighter than the rest, in amber light.

    Language models are trained once, then frozen. Improving one usually takes another costly training run. Could a model fix itself instead, just by reasoning about its own insides?

    A team from five universities let an open model see its own weights. It could read them, edit them and test itself, but never train. The largest version, with 27 billion parameters, set to work.

    Its best edit made it 6.6 points more accurate on 490 unseen problems. Handed a damaged copy of itself, it also found the broken matrix and repaired it.

    Smaller versions of the same model failed. The 2-billion model made no edit, and the 4- and 9-billion models broke themselves. Even the largest withdrew its own edit in five sessions out of eight.

    Source: Language Models Can Rewrite Their Own Weights, https://arxiv.org/abs/2611.21885

    PreprintExperiment❓ Quiz
  9. Space & astronomy

    JUPITER'S POLAR CYCLONES ARE NOT ALL ALIKE

    Oblique view of a giant planet's pole with a ring of large white swirling cyclones around a central one, set in cream and brown cloud bands.

    NASA's Juno probe flew low over Jupiter's north pole eleven times. Its passes ran from May 2023 to May 2024. There, eight cyclones circle a central one, forming an octagon.

    Juno's microwave radiometer sees beneath the clouds, down to where water condenses. The eight outer cyclones shine brighter than their surroundings. Either they are a few degrees warmer, or poorer in ammonia gas.

    Either way, it looks like a warm core, as in Earth's tropical cyclones. Five of the eight show about twice the effect of the other three. Those three also look more irregular in Juno's images.

    The surprise sits in the middle. Big and cloudy in visible light, the central cyclone almost vanishes in microwaves. The team suggests it may have a different origin.

    Source: Three Dimensional Structure of Jovian Polar Cyclones Revealed by Juno Microwave Radiometer, https://arxiv.org/abs/2611.21841

    PreprintObservation❓ Quiz
  10. Mathematics

    THREE DISCS IN A CIRCLE, AND A LOPSIDED MINIMUM

    Top view of a round brass tray on dark paper holding three flat discs of pale wood of about the same size, two side by side and a third below touching both.

    Draw a circle of radius 1. Inside, place two discs that touch each other and the edge. Then slip in the largest disc that still fits beside them, touching both.

    How small can the three areas be, added up? Anatoly Eydelzon, at the University of Texas at Dallas, proves the answer. It is exactly 11π/18, about 1.92.

    You might bet on three equal discs. Wrong: the best radii are 1/2, 1/2 and 1/3. The same minimum also appears in another order. Start with discs of 1/3 and 1/2, and the third one measures 1/2.

    The hard part is proving the third disc really is the largest that fits. The paper offers three separate proofs. They use Descartes' circle theorem, an inversion, and contact-point coordinates.

    Source: Three Circles in a Unit Disk: Three Solutions to an Extremal Problem, https://arxiv.org/abs/2611.22121

    PreprintTheory
  11. Earth & climate

    WARM NIGHTS COST BRAZIL'S COWS THEIR MILK

    Black-and-white and reddish dairy cows lying on straw in an open-sided barn at night, large fans hanging from the roof and a few dim lamps lighting a hazy aisle.

    Heat stress cuts milk yields. Dairy farms use rules of thumb to say when it starts. A Brazilian team tested them on a huge dataset. It covers 788,528 milk records from 87,809 crossbred cows, over 24 years.

    The key factor is the night. Some nights stay hot after sunset. Cows then cannot shed the heat built up during the day. On average, cows then give 2.37 kg less milk a day.

    High yielders start losing milk around the classic thresholds. The lowest yielders show no loss at all. After seven nights without relief, yield is 8.6% lower.

    Their genes matter too. At the same high milk output, half- and 5/8-Holstein crossbreds hold steady longer. Cows that are 7/8 Holstein start losing milk much sooner.

    Source: Milk yield responses to heat stress in tropical dairy cattle, https://arxiv.org/abs/2611.22980

    PreprintData analysis❓ Quiz
  12. Computing & AI

    AI HAS CHANGED SOME WORDS OF SCIENCE, NOT SCIENCE ITSELF — YET

    Large open bound volume on a worn wooden table with a few lines highlighted in yellow, in front of dim shelves packed with old bound books.

    Since late 2022, large language models have spread through research labs. Has science itself changed? A team at the New Jersey Institute of Technology checked. It read 2.5 million full papers, from 2015 to 2025, across five disciplines.

    Some words did change. "Delve" was nearly absent before 2023, peaked in 2024, then fell sharply in 2025. In all, 796 words surged and 199 declined.

    But about a thousand words weigh little against tens of millions of distinct words. The richness of their vocabulary shows no break in 2022. Scientific terminology keeps rising in every field.

    Most long-term trends simply carried on. There are a few exceptions: computer science teams from Asia grew faster after 2022. The title's last word is chosen with care: "yet".

    Source: Large Language Models Have Not Yet Transformed Science, https://arxiv.org/abs/2611.22975

    PreprintData analysis❓ Quiz
Legal notice