Hao Li

Achieving trillion-parameter performance with just 35 billion parameters by scaling agent horizons reveals a new frontier in model efficiency.

Flux-GS slashes the storage overhead for mobile 3D rendering while preserving essential visual fidelity, enabling high-quality graphics on resource-constrained devices.

Spatial reasoning can be transformed from isolated frame predictions to dynamic scene understanding, significantly boosting performance in multi-view and video tasks.

Achieving a 45.2% boost in prediction accuracy, the Multi-Adapter PPO framework revolutionizes wavelength selection in LIBS by leveraging reinforcement learning and cross-attention techniques.

Models with similar success rates can exhibit vastly different strengths and weaknesses, revealing the hidden complexities of mobile manipulation capabilities.

Factorized Neural Operators achieve superior physical modeling by separating transient and persistent responses, leading to enhanced accuracy and interpretability.

Incorporating egocentric human video data into robotic training can lead to substantial performance gains, with ACE-EGO-0 setting new benchmarks in VLA tasks.

AutoMine outperforms existing methods in scenario mining, achieving a HOTA-Temporal score of 36.38 in a competitive setting.

Reasoning VLMs falter under semantic distractions, often mistaking irrelevant cues for evidence, which can lead to incorrect answers.

Mixed-authorship documents can be harder to detect than purely human or AI-generated texts, challenging existing assumptions about AI-text detection.

Integrating raster and vector data could revolutionize geospatial AI, unlocking richer insights from Earth observation data.

Achieving superior safety alignment with LLMs using only 100 harmful samples, SafeSteer drastically cuts alignment costs while maintaining model performance.

Forget data selection—reordering your existing dataset using these four simple guidelines can significantly boost LLM training performance and stability.

Spatial foundation models aren't as "all-round" as we thought: SpatialBench reveals surprising generalization gaps and the critical importance of domain alignment over naive data scaling.

Achieve stable and accurate robot control with significantly less data by learning the system's energy function directly.

Over-eagerly adapting VLMs for robot control can actually hurt performance, suggesting the original VLM representation already encodes surprisingly useful action priors.

Injecting region-specific awareness into the diffusion model's initial noise unlocks significantly better compositional image generation, outperforming existing methods in cross-region coherence.

LLMs that ace general web browsing still fail miserably at autonomous scientific literature discovery, revealing a critical gap in research-oriented AI agent capabilities.

AI-guided simulations reveal the precise mechanisms of siloxane poisoning in gas sensors, paving the way for designing sensors that resist degradation.

SPD-SheafNets learn richer geometric representations than standard GNNs by operating directly on matrix-valued features, achieving SOTA on molecular property prediction by capturing relationships between directions.

Unified benchmarks reveal the state-of-the-art in simultaneously addressing multiple real-world image degradations like blur, low-light, and rain.

Density-dependent shifts in the potential of zero charge reveal a more accurate predictor for oxygen reduction activity than traditional magnetic descriptors in M-N-C electrocatalysts.

AI coding agents are surprisingly bad at logging, requiring humans to silently fix 72.5% of their logging mistakes.

Inverting the standard multimodal learning paradigm, Uni-ViGU shows a video generator can serve as the foundation for both video generation *and* understanding, achieving competitive performance on both.

Robots can now better assemble boxes in the real world thanks to a video-generative value model that anticipates future states, moving beyond static snapshots for more reliable task progress assessment.

Structured composition unlocks significantly better agent performance compared to flat skill invocation, even with the same skill set.