Lưu trữ

Astronomers in China have found a lonely cloud of hydrogen. It floats about 12 million light-years away, with no galaxy nearby.

Deep images from the DESI survey show no stars inside. Yet the cloud does not fly apart. Something invisible holds it together.

That something is dark matter. The cloud weighs about 200 times more than its gas. It could be a "dark galaxy": gas that never managed to form stars.

Our model of the universe predicts such starless galaxies. Earlier candidates sat next to big galaxies that could have torn them apart. This one is alone, the cleanest candidate yet. Only Hubble or JWST can rule out very faint stars.

Source: Discovery of an isolated RELHIC Candidate: J1351+0039, https://arxiv.org/abs/2609.29852

Freeze tiny grains of gel inside ice. Warm one side slightly. What happens? The grains start moving through the solid ice, towards the warm side.

The trick: ice sucks water out of the gel. Water melts on the warm side, flows through the grain and refreezes behind it. That new ice pushes the grain forward.

A team at ETH Zurich turned this into a measuring tool. Each grain's speed and swelling reveal how water moves inside the gel.

Gels are everywhere: superabsorbents, drug delivery, soft robots, cartilage. Their key properties are notoriously hard to measure. This method gives thousands of data points from picolitre samples. And it shows the textbook formulas are wrong for real gels.

Source: Extracting hydrogel properties by watching hydrogel particles moving through solid ice, https://arxiv.org/abs/2609.29904

Quantum technologies need light sources that emit photons one by one. Most of that light is lost inside the material. Can we catch it all?

A Max Planck team in Erlangen trapped one organic molecule in a thin crystal layer. The layer is 140 nanometres thick. It sits under a tiny glass dome, above a silver mirror.

The result: 97 % of the light reaches the first lens. That is about 100 million photons per second from one molecule. It is a record for any cold single-photon source.

Quantum communication and optical quantum computers need bright, pure, identical photons. Here, 97.7 % come out alone and 91 % are identical. And molecules are cheap, tiny, and their colour is chosen by chemistry.

Source: 100 million photons per second from a single organic molecule, https://arxiv.org/abs/2609.29736

Iapetus, a moon of Saturn, is strangely rich in ice. And Titan's orbit is oddly stretched. Could one collision explain both?

A researcher simulated impacts on Titan. A body one tenth of Titan's mass hits it at 10 km/s. The blow throws out ice from Titan's outer layer.

About 3.5 Iapetus masses of debris stay in orbit around Saturn. It is 83 % ice, mostly torn from Titan. Titan's orbit becomes stretched too.

A bold idea about how moons are born, put to the test. But the ice ends up on very elongated orbits. Turning it into Iapetus would need gas or debris to round them off. That part is not simulated yet.

Source: A Collisional Origin for Ice-rich Iapetus and Titan's Anomalous Eccentricity, https://arxiv.org/abs/2609.28948

Muons are heavy cousins of the electron. They pass through rock and metal. They already revealed a hidden void inside Khufu's pyramid.

The catch: we only had muons from the sky. About one per square centimetre per second, mostly from above. Images take a very long time.

In Romania, a 10-petawatt laser fired 230 shots. It sped electrons up to 8 GeV, then slammed them into a block of lead. Out came muons, detected up to 42 metres away.

The team imaged a lead block 25 cm thick with these artificial muons. A world first, say the authors. No more waiting for the weak, uncontrolled flux of sky muons.

Source: Imaging with GeV muons produced via laser-wakefield-accelerated electrons, https://arxiv.org/abs/2609.28788

The little clip on your finger measures oxygen in your blood. But it tends to overestimate it in people with dark skin. Low oxygen can then go unnoticed.

One suspected cause: its LEDs emit a broad band of colours. Melanin absorbs some of them more than others. The colour the sensor sees shifts with skin tone.

A Tufts University team tested a physics-based correction for melanin. They used public data from 98 volunteers and 2,991 blood samples.

In their fits, the skin-tone bias fell from +2.8 to +0.1 points. But honesty first: none of these trends is statistically significant yet. About ten times more patients are needed.

Source: Melanin- and linewidth-corrected pulse-oximetry, https://arxiv.org/abs/2609.30122

Water trapped in a fault can make an earthquake worse. Or it can slow it down. Which one wins?

In Paris, a team triggered miniature earthquakes in water-soaked granite. At first, friction heated the water. Its pressure jumped by about 7 megapascals, weakening the fault.

Then it flipped. The fault cracked and made room. Water pressure collapsed, by up to 15 megapascals. And between the two phases, slow earthquakes appeared. The flip showed up in all four experiments.

A first, say the authors: both effects measured directly on one evolving fault. Water pressure even moved up to 50 seconds before rupture. In the lab, the water gave a warning signal.

Source: Competition between thermal pressurization and dilatant strengthening in laboratory ruptures, https://arxiv.org/abs/2609.29617

"Quantum advantage" means a quantum computer beats every classical supercomputer at a task. Until now, it took special lab machines, tuned for the occasion.

A startup, BlueQubit, used an IBM processor rented through the cloud. No special tuning. It ran random circuits on 61 qubits.

One million samples took 19 seconds. The best known classical method would need about 110 years on Frontier, a top supercomputer. The whole experiment used about 11 minutes of machine time.

Almost any user could now reproduce such a test. But careful: this task has no practical use. It is a stopwatch for raw power. And better classical algorithms have cut such claims before.

Source: Quantum computational advantage in random-circuit sampling on IBM superconducting quantum computers, https://arxiv.org/abs/2609.28657

Over 2,000 years ago, the Greek Aristarchus asked a bold question. How much farther is the Sun than the Moon? His trick: at half Moon, Sun, Moon and Earth form a right angle.

He found the Sun 19 times farther than the Moon. The real answer is about 400. Near 90°, a tiny error in the angle changes everything.

A team in Chile rebuilt the experiment. Their tools: a 1.5 m curtain rod, a photo tripod, a tape measure. Aim the rod at the Moon. Measure its shadow cast by the Sun.

Their result: about 97, five times better than Aristarchus, with hardware-store tools. Any class can try it, and learn trigonometry on the sky.

Source: Revisiting Aristarchus's method to estimate the ratio of the distances from Earth to the Sun and the Moon, https://arxiv.org/abs/2609.28513

In the 19th century, scarlet fever killed about 10,000 people a year in England and Wales. Mostly young children, sometimes within a few days.

Then, in London, deaths fell from about 2,300 to about 80 a year. Before antibiotics even existed. How?

Two mathematicians in Canada studied a century of weekly records, from 1842 to 1939. The epidemics came back every autumn. On top, a longer cycle stretched from about 4 to about 8 years.

A simple epidemic model explains almost all these changes. The disease became less transmissible, while birth rates shifted. The bacterium itself may have evolved. And old archives can still reveal how epidemics work.

Source: Scarlet Fever Dynamics in 19th and 20th Century London, https://arxiv.org/abs/2609.28545

Deep underground, the LZ detector waits for dark matter to bump into a xenon atom. It recently recorded one single, unexplained hit, at 248 kiloelectronvolts. Its significance is intriguing, not a discovery.

The blow was far too strong for ordinary dark matter. Unless it weighs over 74 times a proton — an option already largely ruled out.

One idea: tiny black holes born after the Big Bang, slowly evaporating. They would spray out very fast, light particles. A physicist in India ran the numbers.

Verdict: impossible. Such black holes would also cause many small hits, and LZ saw none. Bonus: at most a millionth of dark matter can be this kind of black hole.

Source: Limits on Primordial Black Hole Evaporation from LUX-ZEPLIN, https://arxiv.org/abs/2609.29462

At the heart of some galaxies, a giant black hole shoots jets of matter across space. Seen by radio telescopes, these are radio galaxies.

An Italian-led team hunted for the most distant ones. Using the VLT in Chile, they confirmed one: TXS 2354+015, at redshift 4.946.

At 500 megahertz, it emits 6.2 × 10²⁹ watts per hertz. That is about twice the previous record. It is now the most powerful radio galaxy known.

It is also the second most distant ever found. Most of these monsters hide behind dust. Many more may be waiting. The team's method, based on how early galaxies vanish in some colours, finds them where the classic radio test fails.

Source: The quest for high-redshift radio galaxies II. Discovery of the most powerful known radio galaxy at z=4.946, https://arxiv.org/abs/2609.28632

The surface of the Sun boils. Cells of hot gas rise, cool and sink again. Simulations predicted tiny whirlpools in between. No telescope could see them. The sharpest earlier views reached about 150 kilometres.

Until DKIST, the 4-metre solar telescope in Hawaii. It filmed a small patch of the Sun for 3.2 minutes, 740 images per second.

Result: 201 whirlpools detected. Their typical width: 38.6 kilometres. More than 99% are under 50. Each turns about once every 37 seconds.

These whirlpools could carry energy upward and help heat the Sun's atmosphere. The team now wants to catch them higher up.

Source: DKIST unveils MHD-predicted sub-50 km photospheric vortices, https://arxiv.org/abs/2609.30051

When a massive star dies, it can leave a black hole. Most of them drift alone through the Galaxy. Almost impossible to see. In our Galaxy, only about two dozen black holes are known, nearly all with a companion star.

A team at Nanjing University built a simulation of the Milky Way's whole history: stars, couples of stars, explosions and orbits.

Result: about 40 million lonely black holes. Around 80,000 lie within 1 kiloparsec of the Sun, roughly 3,300 light-years. So far, only one has been confirmed.

How to find the others? They bend the light of stars behind them. The future Roman space telescope could catch hundreds of such events.

Source: A Census of Stellar-mass Black Holes in the Milky Way with POPKIN. I. Isolated Black Holes, https://arxiv.org/abs/2609.29211

AI now writes police reports from bodycam footage. A human is supposed to catch its mistakes. But what if the mistakes change the human?

Researchers asked ChatGPT and Gemini to summarise a 25-second video of a car hitting a pedestrian. 19 out of 20 summaries left out the collision itself.

Then 328 people watched the video. A day later, they read a summary naming the wrong traffic sign. Only 44.8% still remembered the right one, versus 83.6% with a correct summary.

Believing a human wrote it changed nothing. Neither did distrusting AI. Knowing that AI makes mistakes did not protect memory. The authors question the use of AI summaries in policing.

Source: AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory, https://arxiv.org/abs/2609.28820

To make green hydrogen, we split water. One key step only works with iridium or ruthenium. These metals are rare, with fragile supply chains.

A US company built a robot lab steered by AI. It tested 2,942 materials made from 26 elements, about 240 per week. Humans only move the samples.

Surprise: the winners are rich in palladium, a family nobody expected. The best one lasted 1,000 hours and kept 96% of its palladium.

Palladium is also precious, but produced more than ten times more than iridium. A chatbot alone did worse than random at this search.

Source: AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution, https://arxiv.org/abs/2609.30133

Crease a sheet of paper, and it remembers that fold forever. So one origami usually makes only one shape. Tiny motors at every hinge are one known fix.

A team at Harvard slipped a steel strip, 0.05 mm thick, across each hinge. Squeezed, it bows up or down. Press it, and it flips.

Each flip decides if a fold goes up or down. One sheet folds into a bird or a whale. Another into a box or a boat. A chain spells ORIGAMI.

Some sheets even reprogram themselves when you fold their edges. One unfolds on its own when pulled. Another counts how often it has been squeezed.

Source: Reprogrammable origami through bistable buckled hinges, https://arxiv.org/abs/2609.28790

Quantum sensors use tiny magnets called spins to measure magnetic fields. The more spins you pack in, the more sensitive the sensor. In theory.

In practice, crowded spins disturb each other. Their quantum state collapses too fast. So diamond sensors keep them rare, a few to tens of parts per million.

A Berkeley-led team grew pink crystal needles where one molecule in three is a sensor. That is over 100 times denser than the benchmarks.

At room temperature, the spins stay coherent for about a microsecond. The secret: a neighbour molecule that blocks their energy from hopping around. Next step: turning these needles into real magnetic sensors.

Source: Dense pentacene cocrystal demonstrates room-temperature coherent control, https://arxiv.org/abs/2609.28912

Delivery drones and flying taxis are coming to cities worldwide. Their noise comes from above. Current noise rules were written for roads and airports. Yet drones seem to annoy people more than traffic at the same volume.

Researchers in London simulated identical flights over ten cities, from Tokyo to Nairobi. Same route, 100 metres up, same traffic.

On busy streets, the change was small: traffic already drowns it out. In quiet inner courtyards of London and New York, noise jumped by over 20 decibels.

Upper floors can be louder than lower ones, near the flight height. The team says future routes must be checked in 3D.

Source: Low-altitude aircraft will reshape noise exposure across global cities, https://arxiv.org/abs/2609.29763

Shuffle a 52-card deck. Turn cards over one by one until the first ace. How many cards, on average? Take a guess. It is a puzzle from a Norwegian statistics magazine.

The answer: 10.6. A Norwegian statistician proves it several ways. The general formula is simple: cards plus one, divided by aces plus one.

Even better: the four aces split the deck into five equal parts, on average. Each gap between aces behaves like the first one.

Flip the question and it becomes useful. If the fifth left-hander in a room of 1,000 is the 50th person you meet, there are about 85.

Source: How many cards, until the first ace: variations, extensions, lachrymae, confidence, Dirichlets, https://arxiv.org/abs/2609.29596

Astronomers compared new James Webb images with old Hubble ones. One quasar, GNz7q, had faded by about one magnitude. It sits at redshift 7.19, 700 million years after the Big Bang.

Twenty years of archives show a fast rise, then a smooth fade over about two years in its own time. Only about 0.1% of quasars vary this much.

The best explanation: a star of a few solar masses ripped apart by the black hole. The flare released about 1.5 × 10⁵³ erg. It ranks among the most energetic such events known.

Webb also caught the echo. Dust about half a parsec away, heated to about 1,500 K, glows after the flash. Until now, confirmed star-shredding events stopped near redshift 1.2.

Source: A tidal disruption event in a quasar at redshift 7.19, https://arxiv.org/abs/2609.30441

On 3 July 2026, the James Webb telescope imaged a massive galaxy cluster. A small red dot appeared that was not there in 2024. Its colours said: a galaxy at redshift 16.

A spectrum taken on 17 August told another story. Broad hydrogen lines revealed an exploding star at redshift 3.34. Its gas is racing outward at 7,000 to 9,000 km/s.

The cluster acts as a lens. It magnifies the explosion about 11 times and splits its light into several images. Five models predict the next one between 2028 and 2032.

Timing that comeback could measure how fast the universe expands. SN Helios is the first beyond redshift 2 with a comeback this close. And it warns galaxy hunters: some record candidates may be impostors.

Source: SN Helios: A Multiply Imaged Type II Supernova Opening Time-Delay Cosmography beyond redshift of 3, https://arxiv.org/abs/2609.30440

An attosecond is a billionth of a billionth of a second. Pulses that short can film electrons moving inside atoms and molecules. This field won the 2023 Nobel Prize in Physics.

At the European XFEL in Germany, a team squeezed bunches of electrons five times in a row. Normal operation uses three steps. Simulations put the current at 150,000 amperes.

The electrons then emitted soft X-ray flashes lasting about 100 attoseconds. Some carried up to 3 millijoules. Their peak power reached about 10 terawatts.

Such energy in so short a flash is unprecedented for attosecond X-rays, the authors say. With tens of thousands of pulses per second, researchers can keep the best ones. The length is inferred from the spectrum, not timed directly.

Source: Generation of 10-TW Attosecond X-Ray Pulses in a Free Electron Laser, https://arxiv.org/abs/2609.31534

Drop a ball on a cushion, and it bounces less. Add more cushion, and it should bounce even less, right? A stack of ordinary printer paper says otherwise.

Physicists in Amsterdam dropped a steel ball on stacks of up to 500 A4 sheets. With ten sheets, the rebound fell from 0.71 to 0.50. Then it climbed back, peaking around 58 sheets.

The impact sends a sound wave down the stack. It bounces off the table and comes back up. If it returns while the ball still touches the paper, it pushes the ball upward.

The surprise: the air trapped between sheets does most of the work. Each gap is about 15 micrometres thick. Pump most of the air out, and this paper trampoline stops working.

Source: The Paper-Stack Trampoline: Acoustic Restitution in Layered Media, https://arxiv.org/abs/2609.31213

Take a chessboard with no end. Put a queen in the corner. Then add one queen per column, always on the lowest square no other queen attacks.

The queens split into two lines. One climbs with a slope of 1.618, the other with 0.618. Both numbers come from the golden ratio, φ = (1 + √5)/2.

Mathematicians conjectured this in 2020, and Donald Knuth checked it up to a billion columns. Now Boon Suan Ho, in Singapore, has proved it. No queen ever strays more than about 7 squares from its line.

The key step was checked by computer across 7,014 possible situations. And the author says the proof itself was found with the AI GPT-6 Pro. The checking code is public, with a Lean formalization.

Source: Greedy queens and the golden ratio, https://arxiv.org/abs/2609.31336

Life faces a chicken-and-egg puzzle. Proteins are built by the ribosome, yet the ribosome needs proteins to work well. So which came first?

Three researchers in France and Australia propose an answer: neither. In their hypothesis, one RNA molecule copied itself and also began making peptides. They call it the riboreplisome.

They sketch 12 steps from a simple RNA copier to a ribosome and a genome. Only the first two need a stroke of chemical luck. The other ten could follow through ordinary natural selection.

In their simulations, the single molecule ends almost six times fitter than a "soup" of cooperating RNAs. The main reason: useful products stay next to the genes that made them. It remains a theory, and the authors propose four ways to test it.

Source: Pathways of early evolution from the perspectives of a riboreplisome – the ultimate RNA machine of life, https://arxiv.org/abs/2609.30816

AI models can now guess your age from an eye photo, a chest X-ray, a brain scan or a blood test. If the guess exceeds your real age, you are said to age faster.

A team led from University College London tested five such "ageing clocks". All four organ-based clocks share a hidden bias. They pull estimates toward the average age, more strongly for sick people than for healthy ones.

The result: the link between "older-looking" and "sick" fades with age. For chest X-rays, it even reverses after 70. Recalibrating or retraining did not remove the bias.

Pairing one healthy and one sick person of the same age, the healthy one looked older in 25% to 59% of pairs. The authors warn: no clock can yet judge one person's health alone.

Source: Prediction bias in biological ageing markers, https://arxiv.org/abs/2609.30322

Weather models solve the equations of fluid motion on a grid. To be exact, they would need grid cells about a millimetre wide. Even the best models resolve less than half of the relevant scales.

A PhD thesis at the University of Utah tries something else. It builds the atmosphere from "turbulons", formal eddies of all sizes stacked like building blocks. No fluid equations, no time steps.

The resulting 3D cloud fields broadly match high-resolution simulations, with clear flaws. Seen from space, their geometry matches satellite images as well as or better than the reference simulations.

Each independent snapshot costs 100,000 to a million times less computing. But the model needs average profiles from another model. And without time steps, it cannot forecast tomorrow's weather.

Source: Weather and Climate Without Fluid Mechanics, https://arxiv.org/abs/2609.30589

The best robot jumpers cheat a little. They slowly wind up a spring, lock it, then let go. One record holder rests about two minutes between jumps.

Engineers at the University of Illinois took the other road. Their robot weighs 281 grams. It has one carbon leg nearly two metres long, driven straight by a small motor, with no spring at all.

The trick is a ribbon of fabric wrapped around the leg like a candy-cane stripe. It keeps the motor near full power and stops the thin leg from buckling.

Result: a 7.6-metre leap after just 190 milliseconds on the ground. Four small propellers at the top keep the long leg upright. It is the highest jump reported for an electric robot without springs.

Source: A Long-Legged, Direct-Drive Monopedal Robot Achieves Exceptional Jump Height, https://arxiv.org/abs/2609.30530

Two identical roads, 50 AI drivers built on the same model. Each morning, a traffic broadcast names the emptier road. The agents use it well: about 64 minutes per trip, close to the ideal hour.

Researchers in Tokyo then added one sentence. Many drivers will hear this tip and switch, so that road may jam. Result: about 96% of agents stayed on the same crowded road.

Average trip time jumped to 95 minutes. Any agent could have saved 69 minutes by switching alone. After 100 rounds, not one of the ten groups had escaped.

Groups of 20 humans under the same warning stayed near balance. In mixed groups, people increasingly took the road the agents kept avoiding. With 15 agents, humans averaged 44 minutes and agents 80.

Source: Warned alike, AI agents avoid the less-crowded road while people take it, https://arxiv.org/abs/2609.30883

61 Cygni is a pair of orange stars just 3.5 parsecs away. It was the first star ever to have its distance measured. Future telescopes want to search it for Earth twins.

Something invisible tugs on one of the two stars. Gaia and Hipparcos data reveal a small wobble in their motion. No planet has ever been confirmed there.

A team combined 37 years of measurements and a million simulated orbits. Their best fit: a cold giant about eight times Jupiter's mass, roughly eight times farther from its star than Earth is from the Sun.

Close enough, it would wreck the habitable zone, where an Earth twin could live. Farther out and tilted, it could too. So before hunting for Earths there, astronomers must confirm this candidate.

Source: Worlds Next Door. V. A Candidate Solar System Scale Super-Jupiter in the 61 Cygni Binary System, https://arxiv.org/abs/2609.32893

Antihydrogen is the mirror twin of hydrogen: an antiproton with a positron. The GBAR team at CERN wants to drop it and watch it fall. First, they must make it efficiently.

Their recipe: shoot antiprotons into a cloud of positronium, an atom made of an electron and a positron. Sometimes an antiproton steals the positron.

For the first time, they measured how likely this is. About 390 antihydrogen atoms at 6 kiloelectronvolts, about 70 at 4. The odds match two independent quantum calculations very well.

Their uncertainties beat those of the only measurement of the same reaction with ordinary matter, from 1997. These numbers now set how fast GBAR can collect atoms for its free-fall test of antimatter.

Source: First measurement of the antihydrogen production cross section through the charge-exchange reaction of low-energy antiprotons with orthopositronium, https://arxiv.org/abs/2609.33990

In 2024, carbon dioxide in the air grew at a record pace. In 2025, the growth rate fell by 45%. Yet fossil emissions still rose by 0.7%.

How can that be? A team led from France and China combined vegetation models, ocean models and the OCO-2 satellite. They built a carbon budget for 2025 within months, not a year.

The answer is the land. Forests and soils absorbed about 2.4 billion tonnes of carbon in 2025. In 2024 they had absorbed almost nothing, or even released carbon. The ocean barely changed.

The tropics recovered most, especially Africa. Northern lands recovered only partly. CO2 in the air is still rising, just more slowly. And this sink swings with drought and El Niño.

Source: Low latency global carbon budget reveals strong land sink recovery in 2025, https://arxiv.org/abs/2609.34226

Glaucoma is managed by tracking the pressure inside the eye. Today, it is measured only at the clinic, now and then. Daily swings go unnoticed.

A company team in Dubai built a soft contact lens with a tiny laser on it. A dye film, thinner than a micrometre, glows when flashed with green light. Its colour shifts as the lens stretches.

On an artificial eye, raising the pressure turned the laser line slightly redder. The sensor read pressure to within about 1.2 millimetres of mercury. That is about the precision monitoring requires.

No human or animal eye was tested yet. The pump laser and spectrometer are still outside the lens. Temperature also shifts the reading. The authors hold a patent on the design.

Source: Contact lens with stretchable distributed-feedback laser for intraocular pressure monitoring, https://arxiv.org/abs/2609.34476

In a sheet of epithelial cells, cells swirl together in jets and vortices. The lines between clockwise and anticlockwise swirls follow a strict universal geometry. It looks the same zoomed in, zoomed out or rotated.

Physicists in Copenhagen blocked cell division with two different drugs. The cells kept moving and swirling. But the special geometry broke down. Its key number jumped from 6 to about 8.

Blocking cell death instead changed nothing. A computer model confirmed the culprit: without divisions, cells swap neighbours less. The tissue becomes stiff and over-constrained.

So division does more than make new cells. It keeps the cell layer flexible enough to rearrange. The authors suggest anti-cancer drugs that stop division might also change how tissues move. That idea remains untested.

Source: Cell division sets a universal flow geometry in cell layers, https://arxiv.org/abs/2609.34631

TOI-2431 b is a rocky planet one and a half times wider than Earth. It circles its star in just 5.4 hours. Its day side should be partly molten.

Astronomers used Hubble to watch it pass behind its star. The total light dipped by just 69 parts per million. It is the first time Hubble has caught such an eclipse for a rocky planet.

Combined with the TESS satellite, this gives a day side near 2,500 kelvins. That is almost the maximum possible, as if no heat reached the night side. Other lava worlds look cooler than expected.

One catch: some of that light could be reflected starlight. If so, the planet is a bit cooler. The James Webb Space Telescope should settle it in the mid-infrared.

Source: The First Hubble Detection of a Secondary Eclipse from a Rocky Exoplanet: the 5.4-hour planet TOI-2431 b, https://arxiv.org/abs/2609.31966

Phosphorus is rare in the Universe, yet DNA, RNA and ATP all need it. How did it end up in living matter? Nobody knows for sure.

In space, only three molecules linking phosphorus and carbon have been seen. Chemists suspected a key step: a phosphorus ion meeting acetylene, a simple carbon gas.

A European team fired phosphorus ions at acetylene at the SOLEIL synchrotron near Paris. 99% of collisions made one product, HCCP+. The reaction also gets faster in the cold, unlike what databases assumed.

In a model of an old carbon-rich star, this reaction becomes the main door to phosphorus ions. Its products could lead to the molecules already observed there. Next step: detect them with radio telescopes.

Source: Synthesis of organo-phosphorous species in space: the reaction of P+ with C2H2, https://arxiv.org/abs/2609.33657

Take a network of dots and lines. If each dot has enough links on average, does it contain every possible branching tree of a given size? Erdős and Sós raised this question in the early 1960s.

In September 2026, a report credited a short proof of the full conjecture to a pre-release AI model, GPT-6 Astra. The proof was also checked by computer in the Lean language.

Jay Cummings, a mathematician in California, now redraws it for humans. He reveals the network one dot at a time and counts the neighbours that arrive too early. Swapping dots shows the count can never be too large.

His article is an exposition, not a new proof. It has 32 figures and worked examples. Rival expositions and extensions to other kinds of networks are already appearing.

Source: The Erdős–Sós Conjecture: A Visual Exposition of the Proof Discovered by GPT-6 Astra, https://arxiv.org/abs/2609.32011

Cool atoms to almost absolute zero and they merge into a single quantum wave, a superfluid. Now trap them in a line and add a wall. One side is denser than the other.

Pull the wall away suddenly. Physicists simulated what happens next. Two trains of ripples rush out in opposite directions, like the tidal bores that run up some rivers.

Between the two trains lies a calm zone where no sound travels. Its ray pattern matches a double rainbow, dark band included. The ripples grow longer only as time to the power one third.

With a big enough density difference, the flow outruns the speed of sound. Sound can no longer travel back upstream: an analogue of a black hole's horizon forms. No experiment has seen these ripples yet, but the team suggests tricks to make them visible.

Source: Quantum undular bores, rainbows, and event horizons in superfluid dam breaks, https://arxiv.org/abs/2609.32232

Science publishes more every year. But are papers still saying something new? Two MIT mathematicians checked one measurable part of originality: how much a paper's text echoes recent work.

They used language models to compare 26,000 full papers from eight fields with earlier work. Copy-paste was excluded: they looked for passages that say the same thing in different words.

The share of echoing text was stable until 2020. Then it more than tripled by 2025. A quarter of 2025 papers contain at least one such passage. Fewer than one match in ten is cited.

The rise coincides with AI writing tools, but the study does not prove they caused it. Similar text is not fraud either. The authors warn against turning originality into a score to game.

Source: More of the same? Are scientific papers losing originality?, https://arxiv.org/abs/2609.32963

Where a refugee is first housed shapes their whole future. In Switzerland, officials assign each case to one of 26 cantons, mostly by quota.

From 2020 to 2023, about 2,000 refugee cases were split by lottery. Half got a canton picked by an algorithm to maximise employment. The other half got the usual random pick. Neither officials nor refugees knew which was which.

Over three years, the algorithm group worked about 10% more months. After the Covid years, the gain reached about 17%. Jobs also lasted longer.

Officials kept the final say and followed the advice 97% of the time. The tool costs little: about 50 francs per adult. The authors, who designed it, call their trial a cautious test. It does not replace language courses.

Source: AI-based matching improves refugee employment in a double-blind randomized trial, https://arxiv.org/abs/2609.35448

Neutrinos come in three masses. Nobody knows their exact order. Is the third one heavier than the first, or lighter?

Under 650 m of rock in southern China, JUNO watches 20,000 tonnes of glowing liquid. It catches antineutrinos from eight reactor cores, 52.5 km away. Each one leaves a tiny flash. In 207 days: 8,294 of them.

The result: the world's sharpest values for two neutrino parameters, to about 1 to 2%. And a first JUNO measurement of a third, to about 1%.

Combined with accelerator experiments, the data favour the "normal" ordering. One analysis puts the odds at about 10 to 1. Another gives 2.1 sigma. The team calls it an indication, not a verdict. More years of data will tell.

Source: Measurement of Neutrino Oscillation Parameters at JUNO and Indication of the Neutrino Mass Ordering, https://arxiv.org/abs/2609.37895

Some particles have a spin of one half. Measure it, and quantum rules allow two answers: +½ or −½. Nothing else. So how can a measurement give 7?

In 1988, three physicists found a loophole. Measure very gently. Then keep only the rare cases that end in an unlikely state. The average can land far outside the allowed range.

Their spin experiment was never done. A Toronto team has now done it. They used rubidium atoms cooled into a single quantum wave.

After a tiny magnetic push, the selected atoms moved as if pushed 14 times harder. In spin terms: a result of 7, with an error of ±3. Physics is not broken. The trick is interference — and a lot of thrown-away atoms.

Source: Measuring the spin of a spin-1/2 and getting a result of 7, https://arxiv.org/abs/2609.37822

Far beyond Neptune spins Haumea, a stretched dwarf planet. One day there lasts about 3.9 hours. What is inside it?

Its two moons, Hi'iaka and Namaka, hold the answer. A spinning, flattened world tugs on its moons in a special way. Their orbits slowly twist.

A US team tracked those moons on 20 years of Hubble images. The twist is too small for a uniform body. Haumea must hide a dense, rocky core under an icy mantle.

The core's density, 2,500 to 3,500 kg per cubic metre, fits rock altered by water. So liquid water once filled the inside of this frozen world. Maybe even an ocean. The authors think most large worlds out there followed the same path.

Source: The Dense, Rocky Core of Haumea Revealed by Satellite Dynamics, https://arxiv.org/abs/2609.36037

Each morning on Mars, a thin white cloud grows from the Arsia Mons volcano. In three hours it can stretch 1,800 km. No model had ever reproduced its long tail.

Cloud ice usually forms on dust grains. Forming straight from vapour needs extreme humidity. On Earth, that is considered impossible in real air.

A Paris team added this "dust-free" ice to a Martian weather model. In a cold pocket over the volcano, humidity reaches 100,000 times saturation. Hundreds of crystals per cubic centimetre appear. The cloud's tail then comes out right.

The authors warn: this is their draft before peer review, with known errors fixed later. The reviewed version says the clouds "suggest" this process. If it holds, a textbook idea has finally been seen at work.

Source: Homogeneous Nucleation of Water Vapor Evidenced by Elongated Clouds on Mars, https://arxiv.org/abs/2609.37259

Earth's plates glide on a soft layer of the mantle. A little molten rock may lubricate it. But how easily does that magma flow, deep down?

Squeeze most things and they stiffen. Experiments said basalt magma does the opposite, down to a point. Then critics called it a lab artefact.

A Chinese-German team simulated basalt melt atom by atom. A neural network sped up the quantum-level forces. Up to 600 runs per condition. Verdict: at 2,073 K, viscosity drops about 19% by 3 GPa, then rises again. Aluminium atoms swapping partners make the flow easier.

The twist: this dip does not explain why magma pools under the plates. Heat does. Hotter depths make melt runny. Cooler heights slow it until it stalls.

Source: Anomalous pressure-dependent viscosity of basaltic melts and its role in asthenosphere melt accumulation, https://arxiv.org/abs/2609.36623

Paint every point of the plane. Rule: two points exactly 1 apart never share a colour. How many colours do you need? Asked in 1950, it is still open: between 5 and 7.

Now change the ruler. Mathematicians can measure distance in countless other ways, called norms. In d dimensions, the usual ruler needs a number of colours that explodes.

Noga Alon, Matija Bucić and James Davies tackle "typical" norms. There, 2d colours always suffice. And some norms need exactly 2d. Ten dimensions, twenty colours.

Along the way they solve a 1978 conjecture by Schoenberg. They also credit ChatGPT 6 Pro with the proof of their last missing lemma. The authors helped steer it with their own ideas.

Source: Hadwiger–Nelson Problem for Typical Norms, https://arxiv.org/abs/2609.36220

In May 2025, a 75,000-tonne container ship ran aground in the Red Sea. Its GPS had reportedly been fed a false position. It stayed stuck for over five weeks. How common is that?

A Georgia Tech team checked position reports from over 367,000 ships. The survey spanned about ten weeks. They flagged impossible motion, then looked for many ships fooled at once.

Result: 17,936 suspicious episodes and 31 hotspots. At least 22 show strong signs of GPS spoofing. Ships loop in circles off the Black Sea coast. Others jump onto a Jordanian airport, far inland.

The same fake track that preceded the Red Sea grounding appeared there five months earlier. The authors propose spoofing risk maps for sailors, like weather charts.

Source: She Spoofed Sea Ships by the Sea Shore: Measuring Large-Scale GPS Spoofing in Global Maritime Traffic, https://arxiv.org/abs/2609.37676

Touch a ripe seed pod of Impatiens balsamina and it bursts, flinging its seeds. Engineers in Hangzhou copied the trick to deliver tiny robots inside the stomach.

Their capsule is about 8 mm wide. A soft silicone dome is turned inside out, like a spring. A gel with iron nanoparticles holds it in place.

Magnets guide the closed capsule to its target. Then an alternating magnetic field warms the gel to about 41 °C. The gel gives way. In under 50 ms the dome snaps back and shoots out its cargo at 1 m/s.

It worked on a printed track and on a pig stomach outside the body. No living animal yet, and the gel must still face stomach acid. Four of the authors have filed a patent.

Source: A Bioinspired Magnetothermally Triggered Capsule for Rapid Microrobot Release, https://arxiv.org/abs/2609.37598

Sulfur is a common element in the universe. But in the cold clouds where stars are born, less than 5% of it shows up. Where is the rest?

A Japanese-Spanish team recreated those conditions in a vacuum chamber at −263 °C. They made single sulfur atoms on a film of space-like dust, sometimes buried under ice.

They warmed the dust to 180 K, then to room temperature. About half of the atoms stayed put, chemically bonded to the silicate. Even atoms dropped on ice dug their way down to it.

In a model of a forming cloud, 10 to 20% of all sulfur gets stuck. Invisible to telescopes, it should only escape when shock waves heat the dust. JWST could check.

Source: Thermally irreversible sulfur chemisorption on silicate grains as a major sulfur reservoir in molecular clouds, https://arxiv.org/abs/2609.37663

Seeds carry their own microbes. Some help the plant, others make it sick. Farmers fight the bad ones with chemicals.

A Paris team tried another weapon: cold plasma. It is electrified air, full of reactive molecules, yet close to room temperature. They zapped seeds of the lab plant Arabidopsis for 5 or 15 minutes.

Untreated, up to 100% of the seeds grew mould, depending on the batch. After 5 minutes of plasma, almost none. After 15, none at all. Every batch still sprouted fully, at the same speed.

The surprise came later. DNA tests on the seedlings showed a reshuffled fungal community. The dominant Penicillium faded in some batches, and other fungi took its place. Cleaning a seed also decides who moves in next.

Source: Cold atmospheric plasma decontaminates Arabidopsis thaliana seeds and remodels seedling fungal microbiota, https://arxiv.org/abs/2609.37222

Khi hai sao neutron hợp nhất, chúng rèn ra các nguyên tố nặng trong một chớp sáng gọi là kilonova. Thiên Hà của chúng ta chỉ nên có vài tàn tích của những vụ nổ như vậy. Chưa từng có cái nào được tìm thấy.

Một nhóm ở Bắc Kinh và Notre Dame đã nghiên cứu HESS J1507−622. Nguồn tia gamma này nằm xa mặt phẳng Ngân Hà một cách khác thường. Không có sao xung, không có vỏ, không có vầng sáng tia X nào giải thích được nó.

Mô hình của họ khớp nó như một tàn tích kilonova, 200 đến 3.000 năm tuổi. Mô hình nghiêng về tuổi dưới 1.000 năm, ở khoảng cách 3,8 đến 14,3 kiloparsec. Vụ nổ có thể đã sáng ngang những hành tinh sáng nhất.

Vậy tại sao không có ghi chép lịch sử nào? Nó mờ đi chỉ trong vài ngày, ở rất xa về phía nam. Một kính thiên văn MeV tương lai hoặc một tiếng vọng ánh sáng mờ nhạt có thể ngã ngũ vấn đề.

Nguồn: HESS J1507-622: A Plausible Young Galactic Kilonova Remnant, https://arxiv.org/abs/2609.38341

Có thể có những thiên hà chứa rất nhiều khí nhưng hầu như không có sao. Chúng gần như vô hình với các kính thiên văn quang học thông thường. Chỉ có sóng vô tuyến từ hydro của chúng mới có thể làm lộ diện chúng.

Một nhóm nghiên cứu Úc đã đối chiếu hai cuộc khảo sát vô tuyến khổng lồ, từ Arecibo và từ kính FAST của Trung Quốc. Một nguồn, AGC 322753, xuất hiện trong cả hai. Cùng vận tốc, cùng hình dạng vạch: không phải là nhiễu giả.

Tín hiệu của nó có hai đỉnh, dấu hiệu của một đĩa đang quay. Nó chứa khoảng hai tỷ khối lượng Mặt Trời hydro. Thế nhưng các ảnh quang học sâu không cho thấy gì ở đó.

Các ngôi sao của nó nặng nhiều nhất là 50 triệu Mặt Trời, nên khí nặng hơn sao ít nhất 40 lần. Lượng khí đó quá nhiều đối với một thiên hà hoàn toàn không có sao. Những ảnh sâu hơn sẽ cho biết nó thực sự tối đến mức nào.

Nguồn: AGC 322753: a Dark Galaxy Candidate detected by ALFALFA and FASHI, https://arxiv.org/abs/2609.39152

Một chất siêu rắn có trật tự như tinh thể nhưng lại chảy không ma sát, như chất siêu lỏng. Các phiên bản làm từ nguyên tử chỉ xuất hiện ở vài phần tỷ độ trên độ không tuyệt đối.

Một nhóm tại Đại học Hồng Kông đã dùng polariton, nửa ánh sáng nửa vật chất. Họ đặt một màng perovskit lên một con chip silic nitrua khắc hoa văn nano. Rồi họ chiếu vào đó các xung laser.

Ở nhiệt độ phòng, ngưỡng thứ nhất tạo ra một chất ngưng tụ. Ngưỡng thứ hai tách nó thành một hoa văn đều đặn, với chu kỳ 3,33 micromet. Hoa văn nền bên dưới chỉ rộng 290 nanomet.

Ánh sáng từ hai phía vẫn đồng điệu: hệ số tương quan 1,0121, rất gần 1. Trật tự tinh thể cộng với tính kết hợp lượng tử: đó là siêu rắn, các tác giả nói. Nó cũng có thể tạo ra ánh sáng nén ngay trên chip.

Nguồn: Room-temperature polariton supersolids, https://arxiv.org/abs/2609.40009

Các họa sĩ đã dùng những dụng cụ giống cây cọ từ thời tiền sử. Cọ ngày nay có đủ mọi kích cỡ, cho nghệ thuật, cho móng tay hay cho tường. Liệu hình dạng của chúng chỉ là thói quen?

Một nhóm ở Orsay, Rennes, Paris, Bordeaux và Palaiseau đã đo khoảng 200 chiếc cọ. Chiều dài tuân theo một định luật lũy thừa của chiều rộng, chỉ phụ thuộc vào hình dạng đầu cọ. Hình nón, tròn hay dẹt: ba số mũ.

Khi ấn lên giấy, những chiếc cọ có kích thước rất khác nhau chống lại với độ cứng gần như bằng nhau. Thế nhưng cắt ngắn một chiếc, độ cứng của nó thay đổi rất nhiều. Tỷ lệ thương mại giữ cho độ cứng không đổi.

Với cọ tròn, kích thước vết màu tăng theo độ sâu ấn cọ. Vậy nên một họa sĩ, hay một robot vẽ tranh, kiểm soát nét vẽ chỉ bằng lực. Chiếc cọ tròn nào trong tay cũng được.

Nguồn: The hidden side of paintbrushes: a geometry shaped by power laws, https://arxiv.org/abs/2609.39410

Hãy tô màu mọi điểm và mọi đường của một mạng lưới, sao cho không có xung đột giữa các phần tử kề nhau. Các nhà toán học từng phỏng đoán rằng danh sách màu riêng không bao giờ đòi thêm một màu nào. Ý tưởng này có từ cuối thập niên 1990.

Ngày 24 tháng 9 năm 2026, nhà toán học Jonathan Noel đã yêu cầu ChatGPT 6 Astra Ultra bác bỏ nó. AI đã đưa ra một phản ví dụ. Noel kiểm tra chứng minh và viết lại.

Mạng lưới này có 20 điểm, mỗi điểm có ba đường. Bốn màu là đủ để tô nó. Nhưng với những danh sách bốn màu được chọn khéo léo cho từng phần tử, việc tô màu trở nên bất khả.

Toàn bộ bài báo chỉ gói gọn trong sáu trang, và chứng minh có thể kiểm tra bằng tay. Giả thuyết đã sai. Một câu hỏi mới mở ra: liệu danh sách riêng có bao giờ đòi thêm nhiều hơn một màu?

Nguồn: The List Total Colouring Conjecture is False, https://arxiv.org/abs/2609.38417

Liệu bạn có thể đi qua mọi điểm của một mạng lưới đúng một lần, theo các mũi tên một chiều, rồi trở về nhà? Đây là một bài toán khó kinh điển. Từ năm 1962, các phương pháp tốt nhất cần khoảng 2ⁿ bước cho n điểm.

Hai nhà nghiên cứu ở Tokyo, tại Đại học Tokyo và công ty CyberAgent, đã vượt qua giới hạn đó. Thuật toán ngẫu nhiên của họ chạy trong khoảng 1,9133ⁿ bước. Đây là lần đầu tiên có cải tiến như vậy cho mạng một chiều tổng quát.

Mẹo là: xóa và nhân đôi các mũi tên một cách ngẫu nhiên, rồi xếp các bản sao vào các nhóm. Sau đó đếm số lời giải theo modulo hai, ở mỗi tổng trọng số. Một số đếm lẻ chứng minh rằng hành trình tồn tại.

Các tác giả cho biết ChatGPT 6 Astra đã tạo ra chứng minh chính. Họ đã kiểm tra và chỉnh sửa nó. Với mạng hai chiều, kỷ lục vẫn đứng ở mức thấp hơn nhiều, 1,657ⁿ.

Nguồn: Breaking the 2^n barrier for directed hamiltonicity, https://arxiv.org/abs/2609.39062

Giờ đây chúng ta có bản đồ đấu nối hoàn chỉnh của dây thần kinh bụng ruồi giấm, đến từng khớp thần kinh. Các mô phỏng dựng trên đó tạo ra những nhịp điệu giống như của chân. Nhưng liệu cách đấu nối chính xác có quan trọng không?

Một nhóm do Hồng Kông dẫn dắt đã mô phỏng mạch điều khiển chân từ hai bản đồ ruồi độc lập. Họ so sánh mỗi bản đồ với 30 phiên bản xáo trộn giữ nguyên các thống kê tổng quát. Tiêu chí của họ được ấn định từ trước.

Nhịp điệu hóa ra rất phổ biến: nhiều mạng xáo trộn còn dao động tốt hơn. Thế nhưng chỉ mạch thật mới khiến các nhóm cơ đối kháng thay phiên nhau rõ rệt. Nó đạt 0,31, so với tối đa 0,14.

Bí mật nằm ở một quy tắc được mô tả từ hơn một thế kỷ trước: các nơ-ron kích thích một cơ thì ức chế cơ đối kháng của nó. Cả hai bản đồ đều mang quy tắc này; không mạng xáo trộn nào có. Chuyển các tín hiệu đầu vào giữa các cơ đối kháng đã xóa sạch sự phối hợp.

Nguồn: How much of fly walking is written in the wiring?, https://arxiv.org/abs/2609.38665

Sâu bên dưới châu Phi và Thái Bình Dương có hai vùng khổng lồ nơi sóng địa chấn chậm lại. Chúng vươn cao hơn 1.000 km từ lõi. Nguồn gốc của chúng còn gây tranh cãi, và một khoáng vật tên davemaoite có thể góp phần.

Nhưng lý thuyết dự đoán davemaoite cứng hơn khoảng 30% so với các phép đo trong phòng thí nghiệm. Khoảng chênh này đã kéo dài hàng chục năm. Một nhóm ở Vũ Hán, Bayreuth và Michigan đã mô phỏng hàng triệu nguyên tử để tìm hiểu nguyên nhân.

Các mẫu trong phòng thí nghiệm đầy những hạt tí hon. Rìa hỗn loạn của chúng vô hình trước tia X tiêu chuẩn. Chỉ 5,8% mất trật tự đã làm độ cứng trượt của khoáng vật giảm 37%, đủ để lấp khoảng chênh.

Áp dụng cho lòng sâu Trái Đất, điều này đặt ra một mức sàn: hạt davemaoite phải rộng ít nhất khoảng 100 nanomet. Điều đó phù hợp với các dấu hiệu cho thấy những vùng rộng lớn này có hạt thô và tồn tại lâu dài.

Nguồn: Elasticity of polycrystalline davemaoite constrains its grain size in the lower mantle, https://arxiv.org/abs/2609.39423

Ốc tai điện tử gửi các xung điện thẳng tới dây thần kinh thính giác. Kết quả khác nhau rất nhiều giữa người này với người khác. Có người hiểu lời nói tốt, có người hầu như không.

Một nhóm tại Vanderbilt và các trung tâm khác đã dựng một bản sao số của tai được cấy ghép của từng bệnh nhân. Nó chỉ dùng ảnh chụp và các phép đo do chính thiết bị cấy ghép thực hiện. Nó ước tính còn bao nhiêu sợi thần kinh khỏe mạnh.

Trên các bệnh nhân của nghiên cứu, điểm sức khỏe thần kinh này đi đôi với khả năng hiểu lời nói, cả trong yên tĩnh lẫn trong tiếng ồn. Hệ số tương quan đạt khoảng 0,75 đến 0,78. Nó giải thích được 50 đến 60% khác biệt giữa các bệnh nhân.

Sau đó, sáu bệnh nhân được tắt vài điện cực ở nơi dây thần kinh có vẻ yếu nhất. Điểm nhận biết từ tăng trung bình 10 điểm, và tất cả đều giữ thiết lập mới. Một thử nghiệm lâm sàng bài bản đang được lên kế hoạch.

Nguồn: Estimates of neural health using cochlear implant digital twins correlate with speech recognition, https://arxiv.org/abs/2609.38634

Phân tử amoniac là một kim tự tháp tí hon. Nguyên tử nitơ của nó có thể đi xuyên qua mặt phẳng của ba nguyên tử hydro và lật ngược lại, như chiếc ô bị gió thổi bật. Trong chất lỏng, điều đó quan trọng đến mức nào?

Một nhóm nghiên cứu ở Seoul và Suwon đã mô phỏng 512 phân tử amoniac bằng một trường lực được học máy. Họ so sánh nguyên tử cổ điển với hạt nhân lượng tử, từ 170 đến 250 kelvin.

Hạt nhân lượng tử hạ thấp rào năng lượng của cú lật, từ 0,272 xuống 0,213 electronvolt. Mỗi lần lật làm đứt tạm thời các liên kết hydro và nới lỏng lớp vỏ phân tử lân cận bao quanh nó.

Trong chất lỏng ấm hơn, điều này hầu như không thay đổi sự khuếch tán. Dưới khoảng 210 kelvin, khi mỗi phân tử bị các phân tử lân cận giam giữ, các cú lật lượng tử giúp nó thoát ra. Khuếch tán tăng 18% ở 170 kelvin.

Nguồn: Nuclear quantum effects enhance diffusion in supercooled ammonia through accelerated pyramidal inversion, https://arxiv.org/abs/2609.38688

Language models learn mostly from web text. But a growing share of that text is now written by other AIs. In June 2026, 27.5% of filtered web tokens were AI-written; by August, 31.1%.

A team from the University of Maryland and the company Pangram Labs trained 800 small language models. They added different amounts of AI text to fixed human text.

AI text helped only models short of data. Beyond the usual budget of about 20 tokens per parameter, adding it raised errors on human text. Fresh human text kept helping.

Skipping the filter at August 2026 levels would cost 1.6 times more computing power. Mixed test sets hid the damage in 95.5% of harmful runs. Pangram sells the detector used here, a conflict to keep in mind.

Source: How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text, https://arxiv.org/abs/2609.40295