PAPER PLAINE

Language models learn knowledge differently depending on how consistently it appears across years of internet data. Knowledge that shows up sporadically across different time periods improves faster during later training stages and ends up performing 7.3 percentage points better than consistently-appearing knowledge, especially for factual tasks.

As IoT devices move through wireless networks, base stations need to track their changing signal paths in real time. This paper shows how to do that accurately even when using cheap, low-resolution analog-to-digital converters (ADCs) instead of expensive high-precision ones. The researchers developed two methods—one for coordinated networks where devices announce themselves, and another for uncoordinated networks where devices share access randomly—and proved both work well with only 4-bit ADCs while cutting computing power and memory use substantially.

Physicists discovered that quantum simulations designed to measure one specific property of a system actually contain far more information than originally intended. By reinterpreting the internal states these simulations create, researchers showed they can extract data about how information spreads through quantum systems, how entangled particles become, and how well quantum messages could be sent — all from experiments already built to answer different questions.

Researchers discovered that the mathematical structure underlying how AI systems should reason about relationships between entities—like "Alice is Carol's grandmother"—is uniquely determined by a principle from physics called gauge symmetry. When you allow each entity its own independent coordinate frame, only one form of attention mechanism emerges as mathematically valid: one that checks relationships by completing triangles and computing what physicists call Wilson loops. The finding connects three separate AI approaches (knowledge graphs, AlphaFold, and transformer models) that were already using this structure without realizing it was the only option allowed by symmetry.

Financial AI systems now handle everything from answering questions to making investment recommendations, but they rarely track where their uncertainty actually comes from. This survey identifies six distinct sources of uncertainty—from dodgy input data to flawed reasoning to unpredictable market outcomes—and shows that most current systems only measure one piece of the puzzle rather than following uncertainty through an entire financial workflow.

Researchers discovered that a key bottleneck in how language models learn comes from how much their internal weights can shift during training. They developed MuonGS, a new optimizer that allows weights to grow gradually instead of staying fixed, which lets the model explore more effectively and learn 15% faster than existing methods while reaching the same quality.

AI has made it cheap and easy to write impressive-looking research papers, which threatens to break down a system where publications signal genuine expertise. The authors built a mathematical model showing how this creates a dilemma for both job seekers (should you go deep or publish broadly?) and hiring committees (should you trust papers or push for interviews?), and they identified several possible new equilibria that academic communities might settle into.

Learning becomes unstable when feedback signals are sparse and noisy—even if the algorithm theoretically learns the right answer on average. Researchers found that variance in learning can grow without bound in these conditions, showing that unbiased learning signals alone aren't enough; they also need to arrive frequently and clearly enough.

When AI systems must balance conflicting objectives—like being both helpful and honest, or fast and energy-efficient—training becomes inefficient because the best policy changes dramatically depending on how much you weight each goal. This paper measures how much objectives actually conflict and introduces two techniques to learn faster: maintaining a small team of specialist policies that each excel at different goal combinations, and identifying which training examples help or hurt learning on new goals. The methods improved performance by 4–27% across robotic control and other benchmarks.

Researchers created a system that lets humanoid robots learn to control force and motion simultaneously—essential for tasks like opening doors or moving objects that require precise pushing or pulling. The robot learned from human demonstrations captured with force sensors and motion trackers, then used an AI model to predict both movement and hand forces. On real robots performing contact-heavy tasks, this force-aware approach succeeded 80% of the time, compared to just 27% for a system that only focused on motion.

Demand on blockchain trading platforms arrives in sudden, intense bursts that are 2.5 times normal traffic at their peak. By analyzing three months of trading data on Robinhood Chain, researchers found that during these bursts, traders hit about 40% more trades across roughly the same pools, with each surge following a predictable two-part pattern: a slow buildup over hours followed by a sharp drop over minutes.

Researchers created MedVTok, a single system that converts medical images—from X-rays and ultrasounds to 3D CT scans—into a unified format that AI systems can understand. Trained on 28 million images across nine different imaging types, the system outperformed existing methods on 24 different medical tasks, from diagnosis to report writing, all using the same underlying representation.