PAPER PLAINE

AI agents that need to handle many different tasks often run into a problem: learning new skills erases old ones. Researchers created WeAct, a system that lets agents dynamically switch between specialized skill modules during a task rather than using one fixed set of abilities. On eight research benchmarks, this approach improved accuracy by 6.2% over standard methods while using less computing power, and by 27.5% over other switching approaches.

A simple ratio—total wealth divided by average salary—strongly predicts inequality across real economies, just as economic models predicted. The finding holds most clearly in wealthy, developed nations, suggesting this single measure captures something fundamental about how wealth concentrates when investment income matters more than wages.

Researchers built Bombadil, a machine-learning model that learns to recognize electrical patterns common to all plants by analyzing 4 million hours of signals from 100 species across labs, greenhouses, orchards, and forests. Once trained this way, the model can detect seven different stress indicators in any plant—even species it never encountered during training—using just a pair of electrodes.

Speech recognition systems work worse for some speakers than others based on gender and accent, but testing for this bias in African languages has been nearly impossible due to lack of data. Researchers created AfriBED, a dataset of 9 to 18 hours of recorded speech in Bemba and Nyanja from 80 to 127 speakers, specifically designed to measure gender and speaker-type bias in these low-resource languages.

Quantum systems built on tree-like lattices can develop ordered states triggered by weak boundary conditions rather than traditional symmetry-breaking — a mechanism that has no equivalent in standard three-dimensional physics. The researchers confirmed this using precise mathematical models of simple quantum systems, finding an intermediate phase where the unpolarized edges of the system actually determine how the bulk interior arranges itself.

Neuroscientists have created PyGPLA, a tool that untangles the coordination between individual neuron spikes and the electrical hum of brain tissue. Instead of checking every possible neuron-to-signal pair individually—a task that becomes computationally overwhelming with modern brain recordings—the tool extracts the core patterns that reveal how entire populations of neurons synchronize their activity.

A new system called GLOBE lets AI agents improve their own strategies for rewriting product listings in search results by storing what works in a reusable memory bank. Rather than starting from scratch each time, the system learns effective techniques for different product categories and builds them into an evolving handbook—kind of like an agent writing down its own playbook. Testing showed this approach outperforms traditional methods, and the strategies learned on one product type transfer well to entirely new categories.

When traders need to sell large blocks of stock, slowing down the sale usually reduces price damage—but how slow is actually optimal? This paper shows the answer depends on two competing forces: how much the sale itself moves the market (which gets cheaper per share when spread over time), and how fast the trader's original profit opportunity is decaying. The researchers found that for realistic market conditions, the optimal time window can span a range more than ten times wider than typical trading models suggest, and they derived the precise schedule for releasing shares that saves 10-17% on execution costs compared to uniform, constant-rate sales.

When crypto exchanges offer perpetual futures contracts on dividend-paying stocks, they use a funding mechanism based on price differences that fails to account for dividends. Traders who hold these contracts through a dividend payment get shortchanged — they bear the stock's price drop on the ex-dividend date but never receive the cash dividend itself. The paper shows this happens because the standard funding system treats crypto assets and real-world assets identically, missing a crucial distinction: for stocks, fairness requires a discrete payment (an "event credit") at dividend time, not just a continuous funding stream.

Researchers discovered that deep neural networks generalize to new data by progressively expanding coverage of directions in input space during training — a geometric pattern that explains how they classify both correct examples and mistakes. This 'Neural Expansion' framework unifies several disconnected theories of how neural networks work and reveals that generalization emerges not from memorization, but from spreading predictive stability across input space in specific directions.

Researchers developed a machine learning method that quickly computes a key statistical measure called evidence, which tells you how well a scientific model explains observed data. Their approach, called ENFR, outperforms existing methods at this calculation and can be reused across many datasets without retraining, making it practical for real-world model testing.

Brain-computer interfaces that read motor intentions from EEG often fail when tested on data from different labs using different electrode configurations. Researchers developed a method that accounts for these physical differences without needing access to target data, improving accuracy by 8 percentage points across five independent datasets—suggesting the approach captures real physical principles rather than just overfitting.