PAPER PLAINE

Machine-learning models that mimic chaotic systems like weather often look good on standard tests but fail at something crucial: responding correctly when you nudge the system. Researchers developed a new test based on linear response theory that checks whether AI models respond the same way the real system does, finding that different AI architectures succeed and fail in completely different ways—some match long-term statistics perfectly while botching response properties, and others do the opposite.

Researchers discovered that diffusion-based samplers can be made faster by starting their computation partway through, rather than from the beginning. By pulling the problem back to a simpler form, then using a mathematical bridge to map samples forward, they cut the work by roughly half while improving accuracy—especially when trying to find rare outcomes that simple methods would miss.

Programmable metasurfaces—electronically tunable antenna arrays—can estimate where a radio signal is coming from by processing waves as they pass through them. By training these surfaces directly for the task rather than fitting them to a generic mathematical operation first, researchers showed that simpler, cheaper designs perform just as well, even when the control system uses only 1–3 bits of precision.

World models need to unlearn information that becomes false as environments change—unlike traditional AI that treats all forgetting as failure. This paper shows that current methods can't tell the difference between a model that correctly updated to new conditions and one that catastrophically forgot important rules, and proposes a new way to measure which knowledge should stick around forever and which should change.

When two different AI models encounter each other and disagree, the smaller or less confident one often persuades the larger one to change its answer—contrary to what you'd expect. Researchers tested seven open-weight language models on reasoning tasks and found that a model's size and certainty don't predict who wins an argument; instead, the result depends entirely on which specific pair of models is talking to each other.

When neural networks train on simple tasks, they break a fundamental symmetry by preferring large weights in some layers over others — a biased drift that shouldn't matter mathematically but consistently happens anyway. This directed pattern emerges from random noise in training, follows predictable laws, and eventually hits a hard boundary where the network can no longer drift further.

Machine learning systems often learn from past experience, but when the world changes, old information can hurt rather than help. Researchers created a method that uses statistical evidence to decide when a system should start discounting its history and adapt to new conditions—and found that this evidence-based approach prevents premature switching about 38% of the time compared to systems that always adapt.

A new method called SNAP learns to understand 3D scenes by generating novel views of them — but it works better by using a simpler decoder and focusing on reconstructing features rather than raw pixels. The technique matches or beats specialized 3D-supervised methods while learning from unlabeled data, and its features surprisingly become viewpoint-invariant, meaning they work even when the camera moves.

Researchers created a new method for solving optimization problems across networks of computers by strategically breaking the network into smaller subgraphs, rather than treating it as one whole system. The approach, called GATE, assigns each computer a specific piece of the problem to solve and passes lightweight messages between neighbors, reducing both the computation and communication needed at each step while still guaranteeing the solution converges reliably.

Researchers found that 3D avatars can be animated using simple math instead of heavy neural networks, cutting animation processing by up to 1,000 times. Their method, called GALA, replaces complex per-frame computations with a lightweight predictor and linear blending, achieving 60 frames per second on mobile phones while keeping visual quality nearly intact.

Most digital systems demand the same level of identity verification from everyone before they can do anything—like requiring a passport check just to enter a store. This paper proposes a smarter approach: ask for more personal information only when the specific action warrants it, letting people participate with minimal disclosure at first and proving their identity gradually as needed. The model works across borders and different legal systems, matching how much verification is required to how risky the transaction actually is.

When artificial intelligence language models are given identical news but presented in different formats, their trading decisions shift dramatically—sometimes swapping which model dominates the market by 40 to 48 percentage points. These shifts happened without any change in the actual number of buy or sell orders, meaning AI traders were simply switching strategies based on presentation alone. The finding reveals that even when markets contain diverse AI participants, how information is packaged can eliminate that diversity from actual trading decisions.