Yufeng Wang

Model selection for scientific forecasting is fundamentally protocol-dependent, with optimal choices reversing based on dataset size, system complexity, and evaluation metrics like state refresh versus closed-loop performance.

Multimodal LLMs can cleanly disentangle camera ego-motion from real-world scene dynamics without any drone telemetry or pose sensors, using just four optical-flow-derived residual tokens per slice to unlock robust aerial video reasoning.

CE3D++ enables flexible, dialogue-driven 3D and 4D scene editing by leveraging a large language model to interpret user intentions and coordinate visual tools seamlessly.

Achieving a Relational Delivery Score of 66.6% reveals that even advanced models struggle with the complexities of merging interacting pull requests safely.

Instruction leakage can lead to misleadingly high accuracy in spatial relation tasks, revealing a critical flaw in goal-conditioned models that could misguide future research.

A novel zero-parameter gating mechanism boosts UAV video segmentation performance by over 4% while maintaining temporal stability.

LLM-generated test suites are shockingly bad at catching even simple code mutations, with even the best models failing to detect over 60% of them.

By strategically dropping "anchor" Gaussians and their neighbors, DropAnSH-GS overcomes the limitations of naive dropout in 3D Gaussian Splatting, leading to more robust and globally informed scene representations under sparse-view conditions.