Fei Wang

Dysco cuts training loss by up to 9 times and boosts federated learning performance by dynamically aligning client-specific subspaces, tackling a critical source of instability in LoRA aggregation.

Kinetic effects at fluid-fluid interfaces can reshape our understanding of classical thermodynamics, revealing a unified framework that connects it to Newtonian mechanics.

A general Agent OS can boost long-horizon robotic execution and enable continual learning through structured memory management and self-evolution.

Calibration-data composition can dramatically enhance quantization performance, with 3.5-bit models outperforming traditional 4-bit baselines by over 20 points.

A novel neural network framework achieves high accuracy in solving PDEs on unbounded domains by intelligently decomposing spatial regions into specialized subnetworks.

Achieving state-of-the-art identity fidelity and temporal stability, TIGER redefines the standards for high-quality face video restoration.

Clients can now independently verify the integrity of LLM interactions, reducing the risk of manipulation by third-party gateways.

Reward hacking can be effectively mitigated by a novel agentic reward framework that enhances exploration in RL, leading to substantial accuracy gains in embodied world models.

SO-RaNN achieves exact mass matching and positivity in PNP systems while ensuring incompressibility in velocity fields, setting a new standard for accuracy in neural network-based simulations.

Kairos achieves top-tier performance in Physical AI while ensuring efficient state management over extended time horizons, setting a new standard for operational world models.

Current AI research in education overlooks the critical sociocultural dimension of learner agency, risking a narrow focus in AI-mediated learning environments.

APEX reveals that optimizing data alongside prompts can boost LLM performance by over 11% while significantly reducing wasted compute resources.

Achieving 74.419% accuracy, this ensemble approach shows that self-supervised learning can dramatically enhance micro-gesture recognition performance.

Overlooked diagonal epipolar geometry holds the key to boosting light field super-resolution, as demonstrated by a new omnidirectional EPI Transformer.

A novel GRU-based dynamics model for tendon-driven robots eliminates self-excited oscillations, achieving superior robustness and accuracy in control.

A single adaptive framework can boost the efficiency of zeroth-order optimization by up to 3x without increasing memory usage.

Finally, one-shot 3D head avatars can have realistic hair, thanks to decoupled modeling and physics-based simulation.

Randomly initialized neural nets can solve high-dimensional integro-differential equations like neutron transport faster and more stably than both physics-informed neural networks and traditional deterministic methods.

Smoke-GS lets you see through the haze, reconstructing 3D scenes from smoky images with surprising clarity by explicitly modeling view-dependent smoke appearance.

Reconstructing 3D scenes from images obscured by smoke and extreme darkness is now significantly more achievable, thanks to insights gleaned from the NTIRE 2026 challenge.

Achieve robust multimodal fusion even with missing modalities by ensuring the fusion head always receives a complete, fixed-size input via learned proxy tokens.

Forget fine-tuning: this method uses smart patch selection to adapt frozen LVLMs for deepfake detection, outperforming baselines without any training.

Injecting "beneficial noise" into cross-attention mechanisms can significantly improve unsupervised domain adaptation by forcing models to focus on content rather than style distractions.

Can AI transform a grumpy cat meme into a beacon of positivity while keeping the cat recognizable?

BrainSTR disentangles subtle disease signatures in dynamic brain networks by explicitly modeling spatio-temporal dependencies with contrastive learning, revealing interpretable biomarkers for neuropsychiatric disorders.

Forget monolithic adapters: a hierarchical "expert forest" leverages semantic relationships between tasks to achieve state-of-the-art performance in class-incremental learning.

Ditch the codebook: VP-VAE achieves stable VQ-VAE training by perturbing latent vectors instead of relying on explicit vector quantization.

PANC achieves a staggering 162% increase in attack effectiveness against Transformer-based visual trackers while slashing noise levels to just 10%.