PAPER PLAINE

When AI systems chain together multiple security checks—tracking who gave permission, enforcing policies, logging actions—critical authorization information gets lost between steps, creating dangerous gaps. Researchers built CONTINUITY, a framework that cryptographically locks security context in place as requests move through an agent's components, ensuring every final action traces back to valid permission. In 2,560 simulated attacks across 128 fault scenarios, the system prevented all harmful actions while completing legitimate tasks without false alarms.

Brain connectivity studies have focused on how two regions talk to each other, but the brain actually works through more complex three-way interactions. Researchers developed a method to map these "triadic" connections—statistical relationships among groups of three brain regions—and found they reveal distinct organizational patterns that simple pairwise analysis misses, providing richer detail about how the brain coordinates activity during rest and active tasks.

Researchers trained smaller AI models to mimic larger ones using just one example, and found the smaller model kept improving for hundreds of training steps and recovered most of the performance gains from using thousands of examples. The breakthrough comes from understanding what matters: a single example exposes the student model to 71.5% of the different situations (called "states") that training on full datasets reaches, but the student absorbs that information slowly—the real bottleneck isn't data scarcity but the model's learning speed.

Mathematicians discovered that two fundamental operations on random systems don't always give the same result when applied in different orders—a property called non-commutativity. By studying what happens when you push these systems to their limits, they found the resulting structures follow a surprising block pattern that's stricter than previously known forms of mathematical order.

When researchers replaced human traders with AI language models in a classic economic experiment, the markets failed to reach equilibrium and produced worse resource allocation than human-run markets. The AI agents showed wildly different trading behaviors depending on which model they used, and analysis of their reasoning revealed they switched from strategic thinking to impulsive "let's just trade now" urgency.

Large language models tasked with fixing bugs in code tend to rewrite far more than necessary, even when they produce correct results. Researchers measured this over-editing problem across top AI models and found that a simple instruction to preserve the original code reduced unnecessary changes by 26% while actually improving fix accuracy by 2.3 points.

When traders act on forecasts over different time periods—some thinking seconds ahead, others days or weeks—their collective buying and selling pushes prices up and down in predictable ways. This paper shows mathematically how an equilibrium emerges: traders anticipate both future signals and the market impact from others' moves, and at balance, the price's predictable component cancels out, leaving only random noise plus the collective footprint of their positions.

When language models encounter typos, garbled text, or shuffled words, the damage spreads differently depending on the type of error and where you measure it. Researchers found that looking only at final output—whether the model gets the answer right—misses crucial information about how the model's internal representations break down, and different errors leave distinct patterns of disruption inside the model that don't always correlate with behavioral mistakes.

Large language models learn facts more effectively when trained on different rewordings of the same knowledge, even when the total number of training examples stays constant. Surprisingly, this benefit holds across different sizes of training batches and types of knowledge, and works regardless of whether the rewordings come from a strong or weak source model.

Researchers showed that certain mathematical properties of a system's connection graph determine how efficiently you can describe that system using simplified quantum representations. They identified three key graph measures—cutwidth, tree-cutwidth, and a new parameter called learning complexity—that control both the overhead needed for these simplified descriptions and how many measurements you need to learn the system from scratch.

A neuron's ability to detect a single synchronized burst versus a specific sequence of inputs depends on one physical property: how spread out the conduction delays are on its receiving branches. The researchers show that this single measurement—which maps onto the anatomical difference between myelinated and unmyelinated connections in the brain—acts as a computational switch, and that this same property predicts the actual diameter of cortical columns, one of the brain's fundamental structural units.

When AI tries to describe everything happening in a long video, it struggles to pinpoint exactly when one event ends and another begins — especially the transition moments in between. This paper introduces a method that uses a visual language model to spot these transitions by analyzing what's actually visible in each frame, then uses that insight to precisely locate and describe event boundaries, outperforming previous approaches on standard video datasets.