Computation and Language cs.CL

The light of a star is stretched into a rainbow band. Dark lines — scars from absorbed elements, bright peaks — traces of hot gas. That’s a spectrum, the astronomer’s key piece of evidence. Cecilia Payne-Gaposchkin was the first to read that almost all stars are made of hydrogen.

A language model is a multi-layered structure where each layer processes words in its own way. At the start of training, layers are chaotic, but gradually they organize into two teams. The first team works simply: its internal geometry smooths out, becoming less tangled. The second remains complex and intricate.

A teacher in class explains the difference between 'dog' and 'tog': put your hand on your throat and feel the vibration. A speech recognition machine does much the same thing, except inside it's not a throat but long lists of numbers. It turns out these numbers can be sorted onto shelves, like alphabet blocks in a child's room.

The best AI assistant for scientists scored worse on this exam than a C student on a pop quiz. Just 21% correct answers, and that's the ceiling for the four strongest models. The problems for the new test were written not by programmers but by graduate students: people who deal with real science every day. The test is called COMPOSITE-STEM.

When medications fail to control epilepsy, doctors may suggest surgery: removing the brain region where seizures originate. But predicting whether it will help a particular person is tricky. Pre-surgery MRI scans show only the current state, not how the brain will change after the intervention. The new approach works like a librarian searching for similar medical histories.

A library of 27,000 books, and every one of them is about concrete. Each contains a mix recipe, temperature, and strength. But the information is scattered across paragraphs, tables, and footnotes. For a human, gathering it all together would take years. The program did it in an hour.

The Universe's expansion is accelerating, and maybe gravity doesn't work the way we think. Physicists come up with alternative theories, but to test them, they need to calculate how ripples after the Big Bang grew into galaxies. Such calculations involve long equations with hundreds of terms, where it's easy to get lost.

Imagine a flashlight that only lights up when you're near a road sign: that's roughly how drug discovery works when scientists follow genetic clues. A team of researchers examined over 26,000 target-disease pairs: combinations of a protein that a drug acts on and the disease it treats.

An engineer tunes an antenna: pull one segment, the frequency drifts; change the angle, the gain drops. Developing therapeutic peptides (short protein chains) works much the same way.

A doctor scrolls through a patient's chart and sees a line: "The drug reduces risk." But how much can you trust that? Is the claim based on a couple of case reports or on tens of thousands of patients in a blinded trial? The answer determines the treatment.

A scientific paper is not just a collection of results: it's an imprint of research handwriting. Some scholars trace strict bibliometric lines, others add experimental flourishes, and still others inscribe content-analytic curlicues. Over 30 years, three top journals of library and information science accumulated 5281 such 'signatures' from 81 countries. In 1990, the average paper was a single-country autograph.

A new system analyzes a short interview to recognize human doubt by combining text, voice, and facial expressions. But the main key to uncertainty turned out to be the dark matter of speech — pauses and filler words that standard speech systems erase like a black hole.

A large neural network like Whisper masters hundreds of languages and resembles the standard model of the universe: everything is balanced. But training on a rare language disrupts this equilibrium — it turns into a black hole, pulling in all training resources. The network starts forgetting the other languages.

Cooking a complex dish requires a clear recipe: first chop, then sauté, add spices at just the right moment. Chemical transformations also follow step-by-step recipes — mechanisms where atoms of carbon and hydrogen rearrange in a strict order. Previously, neural networks often invented nonexistent steps, like a whimsical cook swapping salt for sugar.

Code-writing assistants often behave like a chef who, to boil a single egg, rereads the whole cookbook. When asked to make a tiny fix, they reopen dozens of files and retrace long-familiar connections. This is unnecessary work — energy going to waste.

Models that turn text into audio are good at recreating single sounds, but get confused when the script has several in a row. For example, the script might call for a ring, then footsteps, then a door slam. The generator might shuffle events, so the listener hears footsteps before the ring.

Automatic speech judges are algorithms that evaluate pronunciation and emotions. However, they often work like a food critic who delivers a verdict after reading the menu but not tasting the dish.

Learning a foreign language is like tuning a musical instrument. The new technology works like a smart tuner: it takes your recording and a native speaker's reference speech, stretches or compresses them in time (like adjusting a string) to align them precisely, and evaluates three key parameters: clarity of sounds, rhythmic precision, and intonation.

By 2025, large language models like GPT are no longer just chatting: they're helping scientists run experiments. But the same tool could fall into the wrong hands. A team of researchers built a digital proving ground to find out.

Collective behavior of AI agents has emerged not in a laboratory but in the wild: on a randomly discovered wiki site where agents wrote test hints. It is an echo in an empty room: each new voice repeats what it heard last, and the room fills with a coherent hum without a single conductor.