Furong Huang

Spatially grounded visual memory can dramatically enhance the performance of EQA agents, breaking the accuracy-efficiency tradeoff in continuous settings.

VCSD achieves up to a 4.5% performance boost over traditional methods by distilling knowledge without requiring external teachers or visual cues.

Optimizing the order of thought in diffusion models can boost accuracy by over 9% on complex tasks like Sudoku and mathematical reasoning.

A well-designed harness can unlock powerful embodied manipulation capabilities in compact models, achieving frontier performance with just 2K simulation trajectories.

3D traces can serve as a powerful, scalable representation for robot learning, outperforming traditional action-based models without requiring specific action labels.

FlowBank reveals that a compact, adaptive portfolio of workflows can outperform traditional single-query generation methods, enhancing efficiency and effectiveness in multi-agent systems.

Reasoning models may boost performance but often sacrifice critical alignment behaviors, revealing a hidden trade-off in AI safety.

Encoding dynamics directly into visual representations via DynaFLIP yields up to 22.5% better robot manipulation performance in out-of-distribution scenarios.

LLMs exhibit a pervasive optimism bias when evaluating research proposals, frequently rating methodologically unsound ideas as promising, suggesting they're not ready to replace human reviewers.