Systems that hide a model by shuffling and noising layer outputs can be broken: one query more than the layer has inputs reveals it exactly. The attack rebuilt a ResNet-20's layers with zero error using 5,712 queries, and also worked on large image models.

A time-free flow whose map pushes any starting point straight to a finished word sequence in one step, trained from data without a teacher model. It generates in one or a few steps and beats discrete diffusion and continuous flow baselines on language and reasoning tasks.

Hidden text in a product listing can trick a shopping agent into buying something other than asked, yet still pass signature checks. Success rates reached 90%; a fix tying each purchase to the user's request blocked the traceable ones with no false alarms.

Instead of waiting for a whole input, this design splits features into ordered groups and refines its prediction as each group arrives, so data streams in without stalling. It cut latency 4 to 12.5 times versus earlier lookup-table chips, using more table logic.

The model splits in two: the lower half builds the key-value memory attention layers store, and the upper half reuses it instead of caching its own. It matched Transformers while cutting memory and prefill time by orders of magnitude, and handled 1M-token contexts with near-perfect retrieval.

Separately trained monolingual models were compared to see if their representations line up without shared data or training. They do: one rotation maps internal states between models, and swapping in a rotated English state makes a German model answer with the English model's capital.

Rotary encodings spin the vectors a model compares by how far apart words are; tracing a trained Gemma model shows they are not mainly for fading distant words. Fastest spins build position patterns while slowest carry meaning; fixing the slow ones improves performance.

Forty-five scientists rated 2,960 criticisms from human and AI reviews of 82 Nature-family papers on accuracy, importance, and evidence. The strongest AI beat each paper's best human (60% vs 48%), but AI reviews overlapped more and shared 16 blind spots, so they complement humans.

An agent with no outside rewards explores by seeking experience that helps it build models that predict and adapt anywhere in its world. Theory says the best agents visit the most informative, learnable spots first, and in practice this alone grows more complex behaviors.

Training reasoning AIs with trial-and-error feedback eats huge computing power, so this study maps where the time and hardware go and sorts the ways to spread that work across many chips. It turns those patterns into practical rules for building such systems faster and cheaper.