STEM Synapse

Paper of the Day

Monday 16 November

AI & Machine Learning

Preprint, not yet peer reviewed

Goal Encoding in LLMs

When a language model works through a task with several steps, does it hold a goal internally before it acts?

Why it matters A language model's next intended step can be located in a few internal components, read out in plain words, and transplanted to trigger the same action.

The researchers gave models coding tasks made of several sub-tasks and looked inside. They found the next sub-goal recorded in the output of a small set of attention heads, the components a model uses to pull in information from earlier text, and could read it out in plain words. Moving that signal into a different setting, such as another programming language, triggered the same operation there. Because such a model cannot carry an internal state forward for long, the authors propose that it keeps re-reading the goal from its instructions and checking progress, an idea they test with targeted experiments.

Sharma et al., arXiv cs.AI, 12 November 2026.

Read the paper: Goal Encoding in LLMs

Preprint, not yet peer reviewed

Negative-Only Policy Optimization for One-Sided Verifiable Rewards

One way to train AI models is to let them attempt problems and reward the answers an automatic checker confirms as right.

Why it matters AI models can be trained to improve even when automatic checks can only catch wrong answers and never confirm right ones, by learning from the failures alone.

In many real cases no such checker exists: cheap tests, like a handful of unit tests for code, can prove an answer wrong but cannot prove it right. Rewarding everything that passes then teaches the model shortcuts that slip through the tests. The authors propose learning only from answers known to be wrong: the model is steered away from confirmed failures, and answers that pass are treated as unknown. On coding, logic and maths benchmarks the method solved more problems than existing approaches, and the gain held when it was carried over to other benchmarks.

Xu et al., arXiv cs.LG, 12 November 2026.

Read the paper: Negative-Only Policy Optimization for One-Sided Verifiable Rewards

Tech & Engineering

Preprint, not yet peer reviewed

Implementation of a 1D Gaussian-modulated CV-QKD under finite-size parameter estimation

Quantum key distribution lets two parties share a secret encryption key using light, with security that rests on physics rather than on hard mathematics.

Why it matters A simpler way to share encryption keys using quantum light delivered 0.364 megabits per second over 12 km of ordinary optical fibre.

A leading version encodes the key in two properties of a light wave, which places heavy demands on the sending equipment. Encoding in only one property is simpler, but a complete working system had been hard to build. The researchers built one, running over 12 km of ordinary optical fibre with separate lasers at each end and the full chain of processing needed to distil a secure key. It produced 0.364 megabits of secret key per second under realistic assumptions about limited data. That figure shows what the simpler design costs in performance.

Nascimento et al., arXiv quant-ph, 12 November 2026.

Read the paper: Implementation of a 1D Gaussian-modulated CV-QKD under finite-size parameter estimation

Space, Astronomy & Aerospace

Preprint, not yet peer reviewed

Effects of Lensing by Sgr A* on Semicoherent Searches for Continuous Gravitational Waves. I. Signal Model and Mismatch Studies

Spinning neutron stars are expected to give off faint, steady gravitational waves, and astronomers search for them toward the centre of the Milky Way.

Why it matters The Milky Way's central black hole would distort gravitational waves from neutron stars behind it in ways current searches leave out of their models.

For a star behind the giant black hole there, gravity would bend the waves along two paths, and the two copies would overlap and interfere. As the star moves, the interference shifts in ways the search templates leave out. The authors added simulated signals, with and without this effect, to noise and ran a standard search on them. Magnification helped detection, fast sideways motion took some of the gain away, and faint signals were affected more than strong ones. Limits from such searches therefore cannot simply be rescaled; the full distorted signal has to be simulated.

Sharma et al., arXiv astro-ph, 12 November 2026.

Read the paper: Effects of Lensing by Sgr A* on Semicoherent Searches for Continuous Gravitational Waves. I. Signal Model and Mismatch Studies

Physics

Preprint, not yet peer reviewed

Circular phonons generate negligible magnetization in insulators

Laser light can make the atoms in a crystal move in tiny circles, and some experiments have suggested that this motion turns a non-magnetic material into a strong magnet.

Why it matters Atoms set circling by laser light do not turn silicon carbide into a strong magnet, according to a direct measurement of the magnetic field outside the crystal.

The evidence was indirect: it came from how the crystal twisted light passing through it, which can be read in more than one way. Here the researchers also measured the magnetic field just outside a silicon carbide crystal while driving it with intense infrared light. The twisting of light was as large as in earlier reports, but the field outside showed that any magnetism was below 0.001 tesla, far too small to explain it. Whatever twists the light, it is not a large magnetization.

Wu et al., arXiv cond-mat, 12 November 2026.

Read the paper: Circular phonons generate negligible magnetization in insulators

Chemistry & Materials Science

Ultra‐Stable Phosphide Barrier Layer for Thermoelectric Bismuth Telluride

Bismuth telluride is a widely used material for devices that turn a temperature difference directly into electricity, but generators built from it have been held back by the lack of a reliable barrier layer, the thin coating that keeps the material from reacting where it is joined to other parts.

Why it matters Generators that turn heat into electricity with bismuth telluride gain a nickel phosphide barrier layer that stayed stable for 90 days at about 250 °C.

The researchers propose a family of compounds called metal phosphides and pick out nickel phosphide. It reacts very little with bismuth telluride, nickel moves through it slowly, and it bonds strongly at the atomic scale. The joint stayed stable for 90 days at about 250 °C, the longest reported for any such barrier. A test generator built with it turned up to 6.9% of the heat into electricity.

Chen et al., Advanced Materials, 15 November 2026. doi:10.1002/adma.75713

Read the paper: Ultra‐Stable Phosphide Barrier Layer for Thermoelectric Bismuth Telluride

Biology & Ecology

Preprint, not yet peer reviewed

Parental effects can be transmitted through the rhizosphere microbiome

Plants can pass on the effects of what they lived through, such as crowding by neighbours, to their offspring.

Why it matters In a common reed, parents that faced competition had offspring with different root microbes, which may account for part of how the offspring grew.

Known routes include chemical marks on DNA; whether the microbes around the roots are another route was not known. The researchers grew eight genetic lines of common reed with or without a competing species, then raised their offspring in identical conditions. Offspring of parents that had faced competition grew in a more cautious, resource-saving way. They also had different communities of bacteria and fungi around their roots, and statistical modelling suggested these microbes may explain part of the change in the offspring. Parental experience may therefore shape the next generation through soil life as well as directly.

Chen et al., bioRxiv ecology, 14 November 2026. doi:10.64898/2026.11.12.766069

Read the paper: Parental effects can be transmitted through the rhizosphere microbiome

Climate, Energy & Environment

Temperature-Dependent Decay of Infectivity of Infectious Salmon Anemia Virus (ISAV) in Raw Seawater

Infectious salmon anemia virus, a distant relative of the flu virus, threatens salmon farms worldwide and also wild fish.

Why it matters A virus that threatens farmed and wild salmon loses its ability to infect sooner in warmer seawater, which shortens how far it can spread between fish.

Controlling it depends on knowing how long the virus stays dangerous once it is loose in the sea. The researchers kept the virus in natural seawater at five temperatures from 4 to 17 °C and, after set times, tested whether it could still infect young salmon. The warmer the water, the faster the virus lost its ability to infect. In a model of how the virus drifts, the average distance over which fish could be exposed fell by about 40% as the sea warmed from 4 to 17 °C. The findings bear on how salmon farms guard against the disease.

Kannimuthu et al., Environmental Science & Technology, 15 November 2026. doi:10.1021/acs.est.5c16253

Read the paper: Temperature-Dependent Decay of Infectivity of Infectious Salmon Anemia Virus (ISAV) in Raw Seawater

Psychology & Cognitive Science

Preprint, not yet peer reviewed

Environmental Preference and Context-Supported Episodic Recall

Do people remember things better when they come across them in places they like?

Why it matters People remembered more of what they saw in virtual places they personally liked; a general taste for nature made no difference, in a study of 60 people.

Little research had looked. Here, 60 volunteers explored four virtual environments, some urban and some natural, some regular in layout and some irregular. They were then tested on what they remembered, straight away and after a delay, and asked which places they liked. A general liking for nature or for orderly spaces made no difference to memory. But in tests that used the setting as a cue, people recalled more objects from the particular environments they preferred. The authors suggest that how a place affects the mind may depend on how each person feels about it, not only on its design.

Huynh et al., PsyArXiv, 15 November 2026. doi:10.31234/osf.io/dxa2h_v9

Read the paper: Environmental Preference and Context-Supported Episodic Recall

Math & Computer Science Foundations

Preprint, not yet peer reviewed

Combinatorial Bernoulli Factories with Polynomially Many Coins

Suppose you have coins whose odds are unknown and want to use them to make a random choice among complicated options, such as ways of pairing people up, that reflects those odds exactly.

Why it matters Coins with unknown odds can drive a random choice among arrangements, such as pairings, that matches those odds exactly, now with far fewer flips.

A procedure that does this is called a Bernoulli factory. An earlier one could need an astronomical number of coin flips. The authors give one algorithm, covering every problem where such a procedure can exist, whose expected number of flips grows at a manageable rate with the size of the problem and with how lopsided the coins are. For pairings of n items it needs roughly n to the eighth power flips, where the earlier method needed at least n to the power 2n minus 1.

Leme et al., arXiv cs.LG, 12 November 2026.

Read the paper: Combinatorial Bernoulli Factories with Polynomially Many Coins