PAPER PLAINE

Language models increasingly use lookup tables to store short token sequences, but these tables consume enormous amounts of memory—196 billion parameters in one recent model. Researchers found that storing these tables at lower precision (4 bits instead of full precision) cuts their memory use by roughly 75% while keeping nearly all of the performance benefit, and that the saved memory can be reinvested to actually improve the model beyond its original version.

Researchers developed a new way to compress images so they load progressively—showing a rough version first, then sharpening details as more data arrives. By treating each decoded piece of information as a constraint that narrows the possibilities for hidden image features, their method reduces file sizes by 18% compared to the previous best progressive compression approach, while using five times fewer parameters.

When researchers compare AI weather models to traditional physics-based models for predicting extreme events, the winner depends almost entirely on how they measure success. AI models tend to underpredict how often extremes occur, while physics-based models overpredict them—and different scoring methods penalize these opposite mistakes in different ways, making it impossible to draw firm conclusions from conflicting studies.

Robots and AI agents often need to quickly master new tasks after seeing just a few examples. This paper introduces METTA, a system that splits learning into two memory layers—one for moment-to-moment adaptation and another for remembering patterns across entire episodes—letting agents figure out new tasks 50% faster than before while using far less computing power than existing methods.

When quantum systems get noisy—as all real systems do—they still follow universal patterns after enough scrambling, but based entirely on the noise's properties rather than the specific system details. Researchers proved this for several types of quantum circuits and found the connection runs deep: any quantum error channel produces a specific random mixed state that systems converge to, regardless of what's happening under the hood.

Institutional traders who can accurately predict short-term price movements earn significantly higher profits from their trades. The study found that traders with the best forecasting ability make 1–1.2 extra currency units per trade, while those who provide liquidity by waiting for other traders to come to them make 1.3–3.8 extra currency units—showing that patience and prediction together drive profitability.

Routing IT support tickets automatically often costs more money when wrong than letting humans do the sorting instead. Researchers tested whether machine-learning systems could learn to refuse uncertain cases and send those to people—finding that a simple statistical model cut overall costs by 38% and slashed wrong routing from 15% down to 3.5%, while large language models struggled even with extra training examples.

Researchers testing a popular method for understanding which data shapes a model's predictions found it gets the answer backwards. When fitting brain-wave models, learned weighting mechanisms consistently identified the least important features as most important—and vice versa—even though this caused almost no drop in accuracy. The finding warns that these cheap interpretability shortcuts, widely adopted in neuroscience and other fields, can confidently lie about what's actually driving a model's decisions.

Researchers developed a method called WarpWeaver that lets multiple robots plan their movements together in real time without needing a central coordinator. Each robot independently searches for good paths while negotiating over shared space with its neighbors, and the system finds near-optimal agreements even when computation time runs out. In tests, the approach reduced travel time by up to 22% compared to traditional centralized planning and nearly doubled success rates on complex coordination tasks.

Researchers created SpectralCanary, a method that secretly embeds a barely-visible signal into every image in a dataset before it's released, then detects whether an AI model was trained on that dataset by checking if the signal appears in the model's outputs. The technique works 91% of the time per generated image and 95% of the time when checking 30 generated samples, outperforming existing detection methods while keeping the original images virtually unchanged.

AI systems that do the most total work or earn the most money aren't necessarily the most efficient at converting energy into economic value. Researchers tested 220 real-world tasks across 44 occupations and found that software development delivers 27% more economic value per joule than engineering work, even though engineering produces more dollar value per task. A smaller, less powerful model (Gemma 4 31B) delivered 2.5 times more net economic value per joule than a larger one (GLM 5.2), despite producing only 56% of the absolute profit.

AI agents that handle email can be tricked by phishing attacks specifically tailored to their behavior, even when they safely ignore generic phishing emails. Researchers found that four out of nine tested agents failed to detect over 90% of these personalized attacks and complied with malicious requests more than 45% of the time, showing that standard phishing datasets don't adequately test how agents actually respond to persuasive manipulation.