Fan Zhang

VLM agents exhibit vastly different skill evolution patterns, revealing that initial performance scores can be misleading without considering improvement dynamics.

Predicting fine-grained traffic from coarse data could revolutionize traffic management systems by drastically improving prediction accuracy without the burden of extensive data collection.

A training-free intervention can boost VLA model success rates by over 4% without any additional training or modifications.

Real-time safety filtering for VLA models can be achieved without additional training by harnessing attention heads to localize targets and avoid collisions with moving obstacles.

Struct-Searcher achieves a remarkable 17.2% accuracy boost in multimodal information seeking by effectively managing conflicting evidence through a dynamic structural graph.

Robot video world models can be significantly improved by distilling a multimodal reward function and stabilizing long-horizon inference, leading to better instruction following and manipulation accuracy.

Forget A100s for long-context LLMs – Salca achieves up to 74x better energy efficiency with a sparsity-aware hardware accelerator.

Model-based RL can leverage non-differentiable domain metrics to drastically improve spatiotemporal forecasting, especially for capturing extreme events.