PAPER PLAINE

A mathematical model of Age of Discovery voyages reveals how depleting food and water supplies triggered disease in crews, even when most sailors technically survived. The model shows that provisioning levels didn't just determine who lived or died—they determined whether enough healthy crew members remained to actually operate the ship.

Researchers discovered new optimal arrangements for 60, 61, 92, and 99 mutually repelling charged particles confined to a disk using computational simulations. These arrangements are more stable than any previously found, with the new solutions improving on the best-known energies by measurable margins and showing clear, reproducible symmetry patterns.

Researchers created ScienceIDE, a system that converts scientific code repositories into learning environments where AI agents can practice real scientific tasks like fixing code and running experiments. The resulting models (PhAI-IDE) showed improvements not only at scientific programming but also at general reasoning and knowledge tasks, suggesting that learning from real scientific work improves AI across the board.

Large language models appear to improve their decisions in multi-agent games by studying past interactions, but new research shows they're mostly just recognizing statistical patterns rather than reasoning through strategy. When researchers scrambled the patterns in historical data, the AI's performance collapsed back to baseline—suggesting the apparent learning was pattern-matching, not genuine strategic thinking.

Researchers developed a hybrid approach that combines algebra and combinatorial search to solve the Syndrome Decoding Problem, a fundamental challenge in code-breaking and cryptography. By reformulating the mathematical constraints using Grobner bases—a technique for solving systems of polynomial equations—and strategically fixing only part of the information set rather than all of it, the method reduces the computational burden of brute-force search. Tests on parameters from the NIST-standardized Classic McEliece cryptosystem show the approach is computationally feasible and suggests algebraic methods could compete with traditional decoding algorithms.

Two major tokenisation algorithms used in language models differ in two ways: what they optimise for (compression or likelihood) and how they search for solutions (building up or pruning down). By testing all four combinations, researchers found that the search method matters far more than the objective — bottom-up tokenisers consistently performed better at compressing text, regardless of what they were optimising for.

Researchers used a large neural network trained to predict images step-by-step to convert raw camera sensor data into viewable photographs. The method improved image quality by about 0.6 decibels in standard tests and reduced visible artifacts by 21%, even when working with only 3% of its parameters active — but struggles with getting colors exactly right when camera information is incomplete.

When two AI classifiers make the same predictions, you might still need to label new examples to figure out which one actually performs better—because they can rank their confidence differently. This paper shows that you need roughly 56% of your available labels on average to confidently declare a winner, though the exact number depends on what performance metric you care about and how many models you're comparing.

Neural audio codecs used in AI speech systems generate very long sequences that slow down computation. LACE solves this by compressing audio differently at each layer of processing rather than forcing all layers to use the same compression boundaries, achieving better sound quality at faster speeds. When tested on speech synthesis, LACE delivered competitive quality while speeding up inference.

ScienceBuddy is an interactive workspace where AI agents improve themselves through direct collaboration with researchers. The system uses a two-layered learning approach: it refines how it tackles problems while simultaneously training its underlying model, turning researcher feedback into fuel for continuous improvement across tasks like literature review, hypothesis generation, and experimental design.

Researchers created URCHIN, a language model with only 128 neurons organized like a biological brain, that learns language from child-scale datasets rather than the entire internet. The model works identically whether run on a computer's GPU for training or deployed directly on a phone or specialized brain-mimicking chip without any conversion step.

Researchers found a three-step method for compressing large language models that works much better than existing approaches. By optimizing at three different levels—individual matrices, groups of matrices, and the entire model—they reduced errors that usually pile up during compression, cutting WikiText-2 perplexity from 42.1 all the way down to 11.4 on a 60% compressed model.