University of Florida's DWM Slashes AI Safety Verification Errors by 127x
A deterministic state-to-image decoder replaces noisy GAN surrogates, shrinking reachable sets and letting safety proofs transfer to real vision-based controllers.
20 stories
A deterministic state-to-image decoder replaces noisy GAN surrogates, shrinking reachable sets and letting safety proofs transfer to real vision-based controllers.
A new arXiv preprint extends recent softmax findings to KL and JSD divergences, showing standard code silently corrupts gradients when distributions are close.
A new neural field encoder replaces multiresolution hash grids with a single adaptive tetrahedral mesh, enabling exact mesh extraction and editing.
A differentiable physics model written as convolutions in PyTorch recovers soil and flux parameters from urban temperature data, no training required.
A new benchmark and framework called FF-TRUST tackles sleep staging when labels are noisy and training data comes from mismatched clinical sources.
A new graph attention based reconstruction model sets accuracy and latency records on three industrial defect benchmarks using only one to eight normal samples.
Google researchers give harness optimizers a structured log of past edits, and that history beats starting from the evolved harness alone when the model or benchmark changes.
TUM's DeepBA turns bundle adjustment into a 2M-parameter graph network. It recovers cameras, focal lengths and depth from 2D matches alone, starting from scratch, and gets close to billion-parameter models.
AnisoFlow makes the prediction target of flow models fully learnable, cutting ImageNet FID from 1.90 to 1.55 and killing patch seams in 8-bit inference.
miHoYo's Gamer model turns a Qwen3-VL backbone into an agent that plays Honkai: Star Rail with keyboard and mouse at human speed.
PixVid skips the VAE and renders video directly in pixel space, enabling 960p generation at 480p cost and 34ms control-to-screen latency.
A controlled study of 200+ multimodal models from scratch finds that vision needs its own parameters, but neither separate pretraining nor a dedicated encoder.
A learned energy function inspired by Hopfield and Tank cracks Sudoku-Extreme in 42 milliseconds, exposing weak reasoning benchmarks along the way.
A new benchmark shows LLMs that refuse biased prompts when identity is stated directly still produce stereotypes when identity is hinted at through cultural cues.
Kosmo-GFM packages depth, pose, and dynamic 3D point prediction into a reconstruction pipeline that matches top accuracy at a fraction of the runtime.
A new gradient inversion attack dissects GANs layer by layer to reconstruct private client images from federated learning gradients at pixel-level fidelity.
A training-free decoder exploits spatial and temporal stability in diffusion language models to deliver up to 14x throughput gains without hurting accuracy.
Researchers carve out an 'eager' fragment of timeline-based planning that skips the exponential determinization step and directly synthesizes strategies.
A Google DeepMind team proves that two natural ways to train language models on their own decoded outputs are broken by construction, and concludes that teacher forcing remains the right tool for supervised training.
A NeurIPS 2026 paper reformulates Bayesian experimental design as a game against an adversarial nature, yielding Sibson's alpha mutual information as a robust objective with PAC-Bayes guarantees.