Yuhang Wang

Giving vision-language models an explicit action to defer judgment and gather more evidence boosts safety-critical takeover recall from 74.2% to 88.7% over conventional threshold-based alerting.

Standard motion benchmarks fail inside the cabin because drivers mostly sit still—until right before a maneuver, when arm movement spikes 3.4x.

Unlearning in MLLMs can be both effective and efficient, with MCU achieving superior results while preserving critical knowledge across tasks.

Learned representations can identify driver-specific styles with 93.5% accuracy, far surpassing traditional methods that struggle under varied conditions.

Pruning tool outputs directly within the agent leads to a remarkable 39% reduction in token usage without sacrificing performance.

SA-Homo achieves unprecedented precision in homography estimation, maintaining accuracy even with 8× scale discrepancies that typically cripple existing methods.

E^3RL not only overcomes the autoregressive curse but also enhances LLMs' reasoning capabilities, achieving up to 6.514% better performance than previous state-of-the-art models.

FastContext cuts coding agent token usage by 60% while boosting resolution rates by 5.5% by decoupling code exploration from task-solving.

Coding agents can achieve superior repository exploration, outperforming classical methods by effectively leveraging line-level context for bug diagnosis and code retrieval.

Unlike naive approaches that cause flickering and visual artifacts, 4D-GSW embeds robust watermarks into dynamic 3D scenes by respecting the physics of motion.

Predicting when drivers will engage or disengage driving automation requires more than just video—CAN bus data and route context are key, and handover/takeover events exhibit distinct temporal dependencies.

Autonomous vehicles can now better identify the unexpected, thanks to a new method that boosts out-of-distribution detection by up to 20% without retraining.